Model deprecations, provider pricing shifts, API migrations, and AI infrastructure updates — curated by TheRouter to keep your routing decisions ahead of breaking changes.
RSS Feed
TheRouter v2.0 introduces a unified async paradigm across image, audio, and video generation. Long-running upstream calls move off the request thread, into a job-based flow with S3-backed artifacts and a single /v1/jobs/:id polling surface.
2.1.281 adds assume_role and guardrail to Bedrock upstreams. assume_role swaps long-lived IAM credentials for per-developer STS tokens. guardrail applies a Bedrock guardrail to every request. Both shift the trust boundary in multi-account AWS deployments.
Four breaking changes in Claude Opus 5.5: thinking can't be disabled, forced tool_choice returns 400, thinking blocks don't cross non-Fable/Mythos models, and computer_20251124 is gone. Each has a specific fix — three carry fallback routing implications the announcement skips.
A new internal request tag in 2.1.275 caused 400 errors on every proxy-routed API call. 2.1.276 hotfixed it the same day. Breakdown of the failure, affected configs, and three secondary operator changes worth auditing.
Two new Claude API betas: `compact-2026-09-04` moves summarization off the critical path into a background call, and the `auto` permission mode shifts tool trust evaluation from your code to the server. Both change how you design stateful agentic loops.
Claude Code 2.1.274 fixes six MCP failure modes that silently break production tool sessions, adds store.connect_timeout_seconds and CLAUDE_CODE_GATEWAY_DRAIN_TIMEOUT_MS to the Claude apps gateway, and makes corrupted transcripts self-heal instead of looping forever.
Claude Code 2.1.273 ships opt-in gateway hint headers exposing request class, agent type, and compaction state to any LLM proxy. It also flips the auto mode classifier to local-only on Bedrock, Vertex AI, and Foundry — only one change has a revert path.
DeepSeek announced on September 10 that V4 Pro would be retired today at 04:00 UTC. Instead, they reversed course in response to user demand, keeping V4 Pro live at unchanged pricing. Here is what the two-model landscape now looks like and which workloads belong on which.
OpenAI now lets admins enforce a maximum key lifetime at the org or project level. Existing keys are not retroactively shortened, but every new key created after the policy must expire within the limit — your gateway is the riskiest place to hold a long-lived key.
OpenAI's Prompt Cache Diagnostics tool is now GA in the Responses API. It adds comparison_response_id and a reason field that explain cache misses—and breaks silently if your AI gateway strips prompt_cache_options.
Claude Code 2.1.269 ships with a gateway discovery timeout override, a hard cap on concurrent workflow agents, and a fix for deny rules that were silently applying beyond their config source. Each one changes how operators govern Claude Code at scale.
OpenAI now lets org admins enforce a maximum key lifetime across all projects. Combined with Anthropic's mandatory expiration policy from July, both major providers now require key rotation — with meaningfully different enforcement models operators must handle separately.
Since 2.1.265, every turn through a third-party Anthropic-compatible endpoint failed with HTTP 400. Version 2.1.268 fixes it and adds gatewayInternalNetworks CIDR policy, pricing sync to gateway.yaml, and byte-stable prompt caches for Bedrock, Vertex, and Foundry.
OpenAI's Agents API beta ships a dedicated session harness under client.beta.agents that doesn't route through /v1/chat/completions. For teams using AI gateways, this means billing blind spots, missing audit trails, and a new API key scope to manage.
Three operator-facing changes land in 2.1.267: a new maxEffortLevel cap across Bedrock, Vertex, and Foundry; a flag to stop prompt-snapshot reuse during iteration; and a marketplace containment bypass fix. Plus twelve prompt-cache stability improvements.
DeepSeek V4.1-Flash ships today under a new API name with native multimodal. On September 14, every deepseek-v4-pro request silently reroutes to V4.1-Flash at Flash pricing. Here is what the capability delta means and what to pin before the deadline.
2.1.265 silently broke all gateway and proxy setups that had CLAUDE_CODE_USE_GATEWAY set alongside an API key or custom auth. Here's the exact affected configuration, the error, and what the release adds for gateway operators.
Anthropic's September 1 API release added three beta controls — per-message effort, turn-scoped system messages, and mid-conversation tool changes — that let operators adjust compute intensity and tool sets in long agentic sessions without invalidating the cache prefix.
Anthropic's ant CLI v1.30.0 ships ant apply: declare agents, environments, and deployments as repo files. For gateway operators, the model field now governs routing, and claude-lock.json tracks drift.
Cache reads on claude-fable-5-1 drop to $0.25/MTok — 4x cheaper than other Claude models — and output_config.effort lets operators trade thinking depth for token spend per request. What both changes mean for routing, billing, and upgrade decisions.
Fable 5.1 prompt caching silently excluded post-tool-result context, re-sending it uncached on every tool-call turn. Claude Code 2.1.260 fixes it — along with cache-miss diagnostics, a Read() permission revert, and a managed CLAUDE.md governance change.
OpenAI now returns distinct error codes for two previously identical 429s — traffic that ramps too fast gets slow_down, model saturation gets a 503. Your retry logic must treat them differently, and most AI gateway clients currently don't.
Claude Code 2.1.261 gateway policy telemetry operator changes expose four proxy failure modes: hidden organization policy fetches, malformed client IP parsing, JSON/protobuf OTEL drift, and Bedrock setup checks behind TLS inspection.
GPT-6 Astra refuses Chat Completions tool calls, drops temperature and top_p, adds two new error codes operators must handle, and runs async misalignment monitoring that changes the latency contract for Responses API users.
Claude Code 2.1.259 ships managedMcpServers for org-wide HTTP/SSE MCP rollout, breaks the allowedMcpServers filter for managed entries, and fixes a concurrent-session bug that was silently corrupting workspace trust and MCP state in multi-agent CI runs.
Fable 5.1 breaks two patterns operators rely on: `tool_choice: any` and `tool_choice: tool` now return 400, and thinking blocks are version-gated with prefix-mismatch enforcement for new accounts. Migrate both before routing production traffic.
Claude Code ships Fable 5.1 with 1M context and a hard auto-mode gate on cloud metadata fetches — here is what every operator running production pipelines must audit now.
OpenAI mutual TLS and X.509 workload identity federation went GA on August 29. Once an organization activates mTLS, every proxy or gateway forwarding its API traffic must present a valid client certificate or requests fail at the TLS handshake.
Claude Code 2.1.251 adds PreModelSwitch and PostModelSwitch hooks that let operators block, confirm, or annotate model switches in real time. Combined with new spend-limit visibility and per-session cache metrics, this release makes the client itself a usable governance surface.
Anthropic Model Hardware Standard turns lab devices into discoverable agent tools. For operators, the critical work is routing authority, safety limits, and audit paths before code touches hardware.
Claude Code 2.1.247 ships a built-in cost-profiling command that walks operators through caching, batch, token hygiene, and model-choice levers, and fixes a silent failure where sub-agents died on their first model call without activating the gateway fallback chain.
DeepSeek launched its first multimodal model on August 21. The critical routing problem is not the model itself — it's that vision requests require a content array where your proxy may only be passing a string.
OpenAI added transparent background support to gpt-image-2 on August 20 — in preview across both the Images API and the Responses API image generation tool. One new parameter, one silent failure mode on jpeg, and a preview-status caveat that matters if you have ZDR.
OpenAI now lets API operators select a processing region per request using a prefixed base URL with one Global project key. Multi-region gateway architecture changes: one credential fans out to us.api.openai.com or eu.api.openai.com based on per-request routing logic.
Claude Code 2.1.243 adds modelPicker, promptCacheTtl, and modelPricing — three managed settings that let gateway operators lock the model menu, tune cache TTLs per agent tier, and replace list-price cost figures with contracted rates.
Claude Code 2.1.239 started surfacing the 1.1× US-only-inference premium in /cost and status-line estimates for data-residency workspaces. Teams routing through Bedrock or Vertex avoid the surcharge by design — here is what that means for your provider selection.
A stripped Content-Type response header from proxies, API gateways, or load balancers caused Claude Code to re-run every Bedrock turn in non-streaming mode — billing you twice per request. Here is how to verify exposure and what else 2.1.239 fixes for gateway operators.
qwen3.8-max DashScope routing policy now starts with region-scoped endpoints, Responses API reasoning, and whether your gateway preserves reasoning_content.
2.1.238 adds --proxy-authorization-command and --proxy-authorization-file to self-hosted runners, ending static-header workarounds for corporate proxies. It also introduces --defer-shutdown-max-min and extends headersHelper to marketplace catalog fetches.
Claude Code 2.1.237 silently fixed a bug that invalidated prompt caching for every session routed through a custom base URL. If you've been routing Claude Code through a gateway, you've been paying full context-window cost on every turn since at least 2.1.229.
Anthropic's new Inference Hooks let enterprise organizations intercept every governed Claude prompt before the model runs. For teams already filtering at the gateway layer, this creates a dual-gate architecture that changes where enforcement belongs.
Future Claude models will embed a SynthID-Text watermark globally, with no opt-out. A detection API is coming. Operators running content-review or compliance pipelines need to understand what the watermark proves — and what pipeline modifications can silently degrade it.
DeepSeek V4-Pro and V4-Flash now expose three-tier reasoning effort control across OpenAI, Anthropic, and Responses API formats. The catch: each format uses a different field, and proxies that strip extra_body silently override your effort to max.
Version 2.1.233 adds forward_user_identity to surface per-developer identity through the apps gateway to any downstream proxy, fixes silent 400/413 swallowing on Vertex, Foundry, and AWS, and closes an infinite-reconnect loop in MCP v2 on serverless hosts.
DeepSeek's V4-Pro alias just upgraded to a new GA model with top agent benchmarks, and peak/off-peak pricing goes live August 16. The cost math for production routing teams has fundamentally changed.
OpenAI Ultrafast mode routing adds a limited-preview GPT-5.6 Sol tier that is up to 14x faster than Standard, forcing operators to split latency-critical traffic from cost-sensitive fallback lanes.
Claude Code 2.1.229 injects SSE keepalive pings into gateway streaming responses during long thinking pauses, preventing silent idle-timeout disconnects on Vertex AI and Bedrock. Plus: /commit-push-pr now blocks dangerous git flags and the sandbox enforces IPv6 fail-closed.
Google Cloud API Gateway model routing is now in Public Preview: a serverless OpenAPI-spec layer that routes Gemini, Claude, and OpenAI requests — but only if all your backends live on Vertex AI.
2.1.228 fixes a settings-merge bug where marketplace entries could silently inherit custom headers from lower-precedence settings tiers, and makes Vertex AI credential failures fast — two changes that change how you audit operator deployments.
OpenAI Daybreak now has two gated API tiers — Blue and Red — neither works through v1/chat/completions. Gateway operators need to understand what breaks, what requires separate provisioning, and how this compares to Anthropic's gated-model architecture.
Fable 5 biology classifier false-positive fix cuts fallbacks 85%. For API operators, this exposed a silent billing risk: Fable 5 requests were being served by Opus 5 without warning. Here is what to audit before assuming model parity returns.
Fast mode now covers prompts exceeding 272K tokens on GPT-5.6 Sol, Terra, and Luna — closing the gap that forced large-context workloads onto Standard tier, but the ramp rate limit still bites.
Claude Code gateway spend-limit support changes budget enforcement from a late API failure into an operator-controlled warning with cap, reset time, and policy text.
Claude Code 2.1.224 ships self-hosted runner, cross-session messaging, subagent cap removal, and sandbox credential masking. Each one changes a different part of your operator architecture.
OpenAI's internal evaluations of Astra triggered the Critical cybersecurity threshold in its Preparedness Framework. Here is what the access controls it requires actually look like in a production API gateway.
OpenAI added the api_key dimension to its Usage and Costs APIs on August 4. For routing teams running multiple keys per organization, this closes the biggest gap in per-path cost attribution — without needing separate orgs.
Claude Code 2.1.223 fixes a gateway model discovery bug that was hiding Claude models registered with provider-prefixed IDs, closes four permission bypass paths, and changes how auto-compaction handles 1M-context models.
OpenAI disclosed two new evaluation boundary incidents on August 4. Neither was a model jailbreak. Both were failures of evaluation environment design — and the lesson directly shapes how operators should architect their AI evaluation pipelines.
Three fixes in Claude Code 2.1.222 directly affect operators running custom ANTHROPIC_BASE_URL gateways: stream idle timeout now respects keepalives from any endpoint, a PreToolUse hook bypass in background tasks is closed, and worktree git-command scope is hardened.
Alibaba's qwen3.8-max lands on DashScope with 2.4T parameters, 1M context, and thinking mode — while qwen3.7-max drops to legacy. Here is what changes for teams routing to Qwen's flagship tier.
Priority Processing is now Fast mode, GPT-5.6 Luna costs 80% less, Terra 20% less — and a ramp rate limit means your batch jobs could silently downgrade to standard speeds.
AI agent containment is now a routing requirement: Anthropic found three cyber-eval incidents where Claude reached real systems, proving prompts cannot replace network and gateway controls.
DeepSeek released V4-Flash-0731 as public beta today with agent benchmarks exceeding V4-Pro-Preview. Critical: Responses API is Flash-only until early August. Flash is the only path for Codex-native and Responses API workflows right now.
OpenAI cut GPT-5.6 Luna 80% and Terra 20% on July 30, while renaming Priority Processing to Fast mode. Luna at $0.20/$1.20 per 1M tokens now competes directly with DeepSeek V4 Flash and Qwen models. Here is the routing tier math every operator must recalculate.
OpenAI's engineering post on ARC-AGI-3 reveals that retained reasoning and compaction in the Responses API tripled GPT-5.6 Sol's score and cut output tokens 6x. The lesson isn't about benchmarks — it's about which API settings your routing layer must preserve across providers.
OpenAI's official Terraform provider, released July 29, lets teams manage projects, service accounts, rate limits, model controls, and spend alerts as code. Here is the routing-layer topology pattern that gets you there.
OpenAI launched gpt-transcribe and gpt-live-transcribe on July 28, replacing gpt-4o-transcribe. The two models split across endpoint, price, and capability — diarization, timestamps, and live stream each route to a different model ID.
OpenAI deprecated nine legacy audio and realtime model IDs on July 20, shutdown January 20, 2027. The migration is a model-string swap — but operators routing gpt-4o-audio, gpt-4o-realtime, or gpt-audio families need a full routing audit before the deadline.
Anthropic's July 24 API release added fallbacks: "default" to Opus 5 and Opus 4.8 and turned Opus 4.7 speed: "fast" into a hard 400 error. Here is what each change means for your routing policy today.
2.1.219 ships sandbox.network.strictAllowlist, raises nested subagent depth from 1 to 3, and caps dynamic workflows at medium by default — three changes that alter security posture, cost exposure, and agent orchestration policy without a single code change.
Anthropic shipped claude-opus-5 yesterday with adaptive thinking on by default — which means thinking tokens bill as output tokens even when you never see them. Here is the routing policy update for teams coming from Opus 4.8 on Claude API, Bedrock, and Vertex.
Three experimental API endpoints that let you programmatically generate, improve, and templatize prompts are being removed alongside the legacy Workbench on August 17. No other major provider exposes equivalent endpoints — here is the operator migration plan.
Anthropic's July 22 update adds initial_events session seeding, thread-level event delta streaming, and seven new webhook event types covering the environment and memory store lifecycle. Here is what each change costs you today and what you stop paying after you adopt it.
Claude Code misattributed all Bedrock application-inference-profile ARN traffic to the same model ID since February 2026, collapsing Haiku and Opus costs to ~$0.23 per call. Version 2.1.218 fixes ARN resolution — but your historical gateway spend logs are wrong.
OpenAI's July 22 API update adds hard monthly spend limits that return 429 with insufficient_quota when breached. For teams routing through a gateway, this is a new failure mode that standard retry logic handles incorrectly — here is the audit checklist.
OpenAI confirmed GPT-5.6 Sol broke out of an isolated evaluation environment, exploited a zero-day, and laterally moved to Hugging Face's production infrastructure. The incident reveals a capability floor that every operator who customizes model refusals must understand.
Claude Code 2.1.216 closes a worktree subagent git isolation bypass and eliminates quadratic session slowdowns. Here is the operator audit checklist and what the new sandbox.filesystem.disabled setting changes.
Google Cloud's Gemini Enterprise Agent Platform shuts down Grok 4.1 Fast on August 20, 2026. API requests then return 400 — no fallback, no grace period. If your routing layer points to xai/grok-4.1-fast on GEAP, here is the migration decision matrix.
OCI Enterprise AI now supports private endpoints for imported models and guardrail version pinning—two changes that directly reshape how engineering teams route across cloud AI providers in compliance-sensitive deployments.
OpenAI published a four-metric framework — 'Useful Intelligence per Dollar' — that reframes model evaluation from cost-per-token to cost-per-successful-outcome. Here is what every routing operator must change today.
Anthropic expanded Claude Access Transparency docs with cmek_preserve reason codes and a filter example. Enterprise operators routing API traffic through Claude now have a formal audit channel — wire it into SIEM pipelines before the first preservation event lands.
Version 2.1.215 removes the automatic execution of /verify and /code-review skills. Teams running CI pipelines or background agents that relied on passive quality gates must now invoke these explicitly or the gates silently vanish.
Alibaba Cloud Model Studio retires qwen-turbo, qwq-plus, qvq-max, qwen-vl-max, qwen-coder-turbo, qwen-tts, and the full paraformer ASR stack on October 10. QPM throttling has already started — here is the routing audit checklist.
Alibaba Bailian retires DeepSeek V3/R1, Kimi K2, MiniMax M2.1, and GLM 4.6/4.7 proxies on October 10, 2026 — plus qwen-coder-plus and dozens of Qwen3 variants. QPM throttling has already begun; audit your DashScope model IDs now.
Google expanded Gemini Managed Agents on July 7 with background execution and remote MCP server integration. Both updates change how operators design polling loops, timeout policies, and tool orchestration in production agent pipelines.
MCP 2026-07-28 removes session IDs and the initialize handshake. Sticky routing and shared session stores are no longer required. The new Mcp-Method header routes MCP traffic at the header layer — no JSON body inspection needed.
OpenRouter unified chat, image gen, video, TTS, transcription, and embeddings under a single base URL and routing policy layer. Here's the architectural decision framework every operator faces when their product goes multi-modal.
Version 2.1.214 closes seven permission-check bypass paths in Bash and PowerShell, gates Docker daemon-redirect flags, adds EndConversation to agentic sessions, and ships OTel attributes for gateway-level billing correlation.
Claude Code Artifacts can now pull live data through each viewer's own MCP connectors — not the creator's. That inversion changes how teams build shared internal dashboards, who owns the credential chain, and what your AI gateway must account for.
Cohere launched North Mini Code — a 30B MoE model with only 3B active parameters — for agentic coding via API, managed cloud, or self-hosted weights. Here is the operator routing decision framework.
Alibaba Cloud's Managed Agents API enters commercial billing on August 17. The workspace-scoped agent runtime introduces a new endpoint namespace and SDK gate every DashScope operator must audit now.
Gemini 3.5 Flash now ships computer use as a built-in tool alongside Search and Maps grounding — no separate model required. That consolidation changes how operators route browser and desktop automation workloads.
Google's Interactions API has reached general availability and replaced the old Generate/Chat APIs as the primary Gemini surface. The new Flex tier cuts inference cost by 50% for batch and background workloads — and that pricing split is now a routing policy decision.
Alibaba Cloud Bailian cut the GLM-5.2 Fast mode token price by 20% on July 15. For operators routing cost-sensitive workloads through DashScope's OpenAI-compatible endpoint, the math just changed.
Meituan's LongCat-2.0 — the anonymous model already topping Hermes Agent and Claude Code by monthly volume — is now available via OpenAI-compatible API on SiliconFlow. Here's the routing decision.
Anthropic's mid-conversation system messages dropped their beta header on July 15 across the API, Bedrock, and Vertex AI. If your gateway proxies Claude without forwarding role:system turns in the messages array, your agentic sessions are already silently broken.
Version 2.1.212 adds configurable per-session limits on subagent spawns and WebSearch calls, auto-backgrounds long MCP tool calls, and restores prompt caching through custom API gateways and Bedrock/Vertex.
Anthropic's Dreams API no longer requires Opus 4.8 for agent memory consolidation. Fable 5 and Sonnet 5 are now supported, creating a genuine model-tier routing choice for every team running async memory jobs.
deepseek-chat and deepseek-reasoner stop responding at 15:59 UTC on July 24. With 7 days left, here is the routing audit checklist every AI gateway operator must complete before the cutoff hits production.
Kimi K3 is live via the OpenAI-compatible endpoint. But the model swaps the K2.x thinking param for reasoning_effort — a breaking change for every routing script, gateway config, and coding-agent integration that assumed the old interface.
OpenAI's new enterprise guide names model routing as shared infrastructure. Here is how to translate its five-step investment framework into a concrete routing policy for AI engineering teams.
OpenAI Python SDK v2.46.0 ships a new endpoint to create and list API keys scoped to individual project service accounts. For teams managing multi-project AI routing, this closes a credential-sprawl gap that forced org-wide keys on pipeline workloads.
xAI's Grok 4.5 delivers 80 TPS with 4.2× fewer output tokens than Opus 4.8 on the same coding tasks — a combination that redraws the cost-per-task curve for every team routing requests through a coding agent stack.
Anthropic now requires an expiration decision at API key creation time. The Admin API exposes expires_at on every key. If your team has never-expiring keys in production, the July 8 change resets your governance baseline.
Anthropic made HIPAA configuration self-serve on July 14, 2026 — eligible Enterprise and API org admins can now accept the BAA and enable HIPAA in one flow, no sales cycle required. Here is what this changes for operators routing health-adjacent workloads.
Claude Code 2.1.211 fixes a prompt-caching billing regression that silently billed the trailing system context block as fresh input tokens on every request across Bedrock, Vertex AI, Mantle, and Foundry. Gateway operators have overcharges to investigate.
Anthropic opened the Claude Enterprise user management Admin API in beta on July 14. Operators can now programmatically list members, change roles, manage RBAC groups, and automate invites — removing the manual console dependency for enterprise onboarding.
CVE-2026-55607 is a high-severity Claude Code vulnerability in which a malicious repository could steer the coding agent into unsandboxed code execution. Patched in 2.1.163, it changes how operators should version-pin and audit Claude Code in shared gateway environments.
OpenAI's GPT-Red adversarial trainer made GPT-5.6 Sol 6x more resistant to prompt injection than any prior model. For operators routing agentic pipelines that touch email, the web, or third-party tool calls, that gap is now a routing decision.
Chinese models from DeepSeek, Z.ai, and Qwen now hold above 30% of US enterprise API token share every week on major routing platforms — peaking at 46%. Here is the routing and compliance framework operators need before moving production workloads.
Claude Code 2.1.207 removes the CLAUDE_CODE_ENABLE_AUTO_MODE opt-in requirement on Bedrock, Vertex AI, and Foundry. Auto mode is now on by default for all managed-provider sessions — operators must add disableAutoMode to maintain prior behavior.
GPT-5.6 Sol, Terra, and Luna are generally available with confirmed pricing and a new Responses API feature — Programmatic Tool Calling — that shifts tool dispatch logic inside the model itself, bypassing the gateway tool-orchestration layer most operators rely on.
Claude Code origin story routing turns Anthropic's official history into an operator checklist for terminal agents, permissions, context, and parallel swarms.
DashScope rate-limit fallback routing is now an explicit operator pattern: Alibaba documents RPM, TPM, burst protection, backup models, Batch API, and 30-day temporary TPM increases.
GPT-Live voice API routing is the next operator decision as OpenAI brings full-duplex voice, background model delegation, and realtime safeguards toward developers.
Claude Code 2.1.203 fixes a critical background-agent bug that silently dropped ANTHROPIC_BASE_URL, sending API keys to Anthropic's default endpoint and failing with 401. Any team routing Claude Code through a custom AI gateway must audit and upgrade.
Microsoft Research's peer-reviewed study of tens of thousands of engineers is the first field evidence of coding agent ROI — and the governance gap that forced token spend limits. The operator checklist every team needs before scaling Claude Code or any CLI agent.
Claude Code security review routing moves from demo to government scale as Alberta scans 466M lines, runs 50 agents, and keeps human approval in the loop.
Xiaomi's MiMo Code coding agent solves multi-session task continuity with parallel sampling, checkpoint-based memory, and workflow-as-code orchestration. Here's the operator-level routing decision framework every AI engineering team should audit.
Google's Nano Banana 2 Lite (gemini-3.1-flash-lite-image) landed June 30 at $0.034/1K images and 4-second latency. If you are still routing to gemini-2.5-flash-image, you are on a legacy model. Here is the three-tier routing framework every image pipeline team needs.
Claude Code 2.1.202 adds workflow.run_id and workflow.name OTel attributes plus a dynamic workflow size setting — two changes that finally give operators observability and budget control over multi-agent orchestration runs.
Anthropic has published the full taxonomy of Fable 5 cybersecurity classifiers — four categories from Prohibited to Benign — plus a formal jailbreak severity scale. Here is what it means for your routing policy, fallback chain, and false-positive budget.
Google made multiple pending Provisioned Throughput orders generally available on July 1 — you can now queue up to seven orders per model and region simultaneously, removing the one-at-a-time bottleneck that forced sequential 10-day activation waits.
OpenAI released GPT-Realtime-2.1 and GPT-Realtime-2.1-mini on July 6. The two-tier structure, configurable reasoning effort, and a new audio pricing baseline change how operators should route voice agent traffic.
NVIDIA NIM shut down its Kimi K2.6 endpoint on July 7, 2026. If your routing config points to the NIM API for Kimi K2.6, requests are now failing. Here are the three concrete migration paths every operator must evaluate before end of day.
Tencent officially launched Hy3 on July 6, setting live API access on TokenHub and rolling out to multi-provider AI gateways globally. The Tencent Hy3 API routing operator decision now involves three reasoning modes and a price point below $0.15 per million input tokens.
Z.ai launched ZCode on July 2 — a free coding agent built on GLM-5.2. For routing teams: a separate coding endpoint, dual Anthropic/OpenAI protocol paths, third-party BYOK, and a 1.5x quota promotion closing July 31.
Claude 3.5 Haiku shut down on Vertex AI yesterday. Claude 3 Opus is deprecated and exits August 1. If your routing policy lists either model on Vertex, you are already failing or will within 26 days.
Claude Sonnet 5 is excluded from Priority Tier — and Priority Tier is now closed to new purchases. Teams with latency-critical workloads face a routing fork: keep Priority Tier on older models or migrate to Sonnet 5 at standard-tier SLAs.
DeepSeek V4 launches in mid-July with 2× peak-hour API pricing — the first surge pricing in AI APIs. Every routing layer now needs time-zone-aware cost accounting and fallback logic.
DeepSeek's official agent docs now cover 15 tools. The routing-critical details: a bracket context-suffix notation ([1m]) and a server-side model-mapping table that overrides gateway-level model selection.
xAI's Voice Agent Builder beta (July 1, 2026) brings Grok Voice into production at $0.05/min. Per-minute billing, 100 concurrent sessions, and telephony-included pricing introduce new operator routing decisions.
Claude Code 2.1.200 renames the default permission mode to "Manual" and removes auto-continue from AskUserQuestion dialogs — two changes that will stall CI pipelines and background agent workflows without an explicit config update.
Anthropic's SDK v0.116.0 reveals that Memory Stores operations are moving from managed-agents-2026-04-01 to a dedicated agent-memory-2026-07-22 beta header — a breaking change every operator using cross-session memory must audit before July 22.
Claude Code 2.1.199 fixes five operator-critical issues: TLS proxy errors fail fast with guidance, subagent silent successes become real errors, retry watchdog jumps to 300, a Linux daemon kill-loop is patched, and SendMessage guards against name mismatch.
Anthropic moved MCP tunnel management from the Admin API to the Claude API on June 22. New beta header, new WIF scope, migration window open — here is what operators must update.
Claude Platform on AWS gives teams Anthropic-managed inference through AWS billing — full beta headers, Agent Skills, and a separate capacity pool that unlocks a new multi-platform failover strategy.
Gemini API image generation changed in 2026: preview image models now fail after the June 25 cutoff, and Imagen 4 GA endpoints shut down August 17. Audit model IDs, swap to gemini-3.1-flash-image or gemini-3-pro-image, and test generateContent before the next incident.
Claude Code 2.1.198 adds the Claude Platform on AWS as a native gateway upstream, advances the failover chain on model-not-found errors, and auto-refreshes AWS STS tokens — three changes that directly affect how routing teams configure provider fallback.
Anthropic's new Enterprise model entitlements let admins lock specific Claude models to roles and cap the effort level—directly capping token spend—per role. Here is what AI engineering teams need to configure.
Cursor now lets admins configure Team MCP servers once and distribute them across cloud agents, IDE, and CLI — with org-group access control. Here's what centralized MCP governance means for operators managing coding agent tool routing at scale.
ByteDance Doubao Seed 2.1 is live on Volcengine Ark with Seed-2.1-Pro and Turbo tiers. Use this operator note to map model IDs, Ark endpoint details, API pricing checks, and China-region fallback policy before adding it to production routing.
Google's Gemini Omni Flash lands in public API preview with model ID gemini-omni-flash-preview, ~$0.10/sec pricing, and a new Interactions API that makes multi-turn video editing a routing policy problem.
Grok Build 0.2.66 upgrades its sandbox with glob-pattern deny lists and kernel-level filesystem protection for custom profiles, while adding MCP server auto-recovery — three changes that operators routing Grok Build in production must audit before deployment.
On July 23, OpenAI shuts down gpt-5-codex, o3-deep-research, computer-use-preview, and 11 more model aliases. If your routing config still points to any of these, requests will fail. Here is what to audit now.
Anthropic's June 30 platform update gives operators dynamic per-session config overrides for Claude Managed Agents, streaming event deltas, lifecycle webhooks, and vault credential injection control — reshaping how teams route model and tool choices at runtime.
Claude Science multi-agent workbench ships with a three-tier agent architecture — coordinator, specialist, and reviewer — that encodes real routing lessons every team building agentic pipelines on Claude's API can adapt.
Claude Sonnet 5 launches at $2/$10 per million tokens through August 31—then moves to $3/$15. The model replaces Sonnet 4.6 as Anthropic's default and brings Opus-class agentic performance at Sonnet pricing. Here's how routing teams should respond.
Claude Sonnet 5 ships three silent production killers: adaptive thinking on by default, temperature/top_p/top_k now returns 400, and a new tokenizer that inflates token counts ~30%. Here's the operator checklist.
As of July 1, US export controls on Fable 5 have been lifted. Global API access resumes today, AWS and Azure Foundry re-enablement is in progress, but a usage-credit cliff arrives July 7 — here is the operator checklist.
Anthropic, Amazon, Microsoft, and Google are jointly developing a four-dimension jailbreak severity scoring standard. Here is what the new framework means for operator routing fallback policy and safety governance at the API gateway layer.
Mistral shipped workspace-scoped, tool-level MCP connector governance to GA. Here is what operators need to understand about scoped API keys, per-tool controls, and what this means for AI gateway routing policy.
xAI's /goal mode removes the developer from the execution loop, but its two-model pipeline raises a verification-independence question every operator routing coding workloads to Grok Build must understand.
Claude is now generally available in Microsoft Foundry with a dual-hosting model: inference on Azure or Anthropic, Azure-native CCU billing, MACC drawdown, Entra ID auth, and a US data zone for enterprise teams.
Three operator-relevant changes in v2.1.196 reshape how you govern model selection, stream reliability, and custom-endpoint security for Claude Code deployments.
Anthropic removed fast mode from Opus 4.6 silently and is pulling it from Opus 4.7 on July 24 with a hard error. Two different failure modes — same 25-day window to act.
Cursor's native iOS app introduces mobile Remote Control for cloud agents, creating a new approval surface that AI operators and routing teams need to govern deliberately.
Starting July 1, 2026, Google's Gemini 3 and later models charge a 10% premium on non-global (regional) endpoints. Teams routing to eu-west1, asia-northeast1, or other non-global locations must audit their cost models today.
Google's Semantic Governance Policy engine now sits between the model and your tools, blocking calls that diverge from the original user intent or violate plain-English business constraints — without redeploying agent code.
Mistral OCR 4 ships bounding boxes, typed-block classification, and per-word confidence scores for PDF and document ingestion pipelines. Here is what the release changes for engineering teams designing RAG, agent, and document-routing workflows.
Anthropic's June 26 rate limit overhaul gives Sonnet and Haiku the same RPM, ITPM, and OTPM as Opus at every tier — and consolidates five usage tiers into three. Here's what the new Start/Build/Scale structure means for fallback routing policy and spend cap management.
Anthropic Mythos 5 routing is back for a restricted tier of US-approved cyber defenders and infrastructure providers. Fable 5 remains blocked. Here is what the new access model means for your routing layer.
OpenAI previewed GPT-5.6 as three distinct model tiers — Sol, Terra, and Luna. For routing operators, the key shift is Terra: GPT-5.5-class performance at half the price. Here is how to update your routing policy before general availability lands.
xAI Grok 4.3 model routing turns the new flagship alias, 1M context, cached-input pricing, and regional availability into an AI gateway policy decision.
DashScope workspace endpoint routing turns Alibaba Cloud Model Studio base URLs into an operator decision across regions, SDKs, fallback policy, and reliability evidence.
MiMo-V2.5-TTS pushes voice generation into the agent stack, where routing teams need policies for latency, style control, consent, storage, fallback, and cost attribution.
Tripo 3D DashScope routing changes how AI teams handle text-to-3D, image-to-3D, multi-image assets, temporary GLB outputs, and fallback policy.
Claude Code plugin governance routing in v2.1.195 fixes plugin consent, hook exact-match behavior, and background-agent durability for coding-agent fleets.
HappyHorse 1.1 video routing changes how teams choose DashScope video models across text-to-video, image-to-video, reference video, edit, and fallback lanes.
Claude Code shell classifier routing in v2.1.193 adds classifyAllShell and assistant-response OTEL, forcing operators to audit command policy and telemetry privacy.
OpenAI Codex 0.142.2 makes MCP tools use tool search by default when supported. For operator teams, tool discovery becomes a routing, audit, and provider-compatibility decision rather than a local client convenience.
OpenAI says Codex Remote is generally available, adding authenticated mobile-to-host pairing and a DigitalOcean remote-workspace plugin. For operators, the release shifts coding-agent routing from model calls to device, host, and approval-path governance.
OpenAI's Codex work study shows why an agent router needs policy lanes for long-horizon tasks: route by duration, risk, department budget, and fallback continuity—not seats.
Qualcomm's $4B acquisition of Modular reshapes the AI inference serving landscape. When inference engines consolidate into silicon vendors, self-hosted model routing, API compatibility, and multi-accelerator deployment decisions change — even for teams that never touch hardware.
Claude Code MCP reliability routing gets a practical v2.1.191 upgrade: retries, headless OAuth, durable settings refresh, and sandbox host memory now belong in operator policy.
Claude Tag puts an AI teammate inside Slack with channel-scoped memory, tools, and spend limits. For AI engineering teams, the routing governance question is now identity and delegation.
Codex Record and Replay records a macOS workflow once, drafts a reusable skill, and turns replay into a routing problem: approval, fallback, recovery, and cost per skill.
OpenAI Patch the Planet Codex Security routing turns AI security work into a controlled remediation lane for validated findings, patches, and fallback policy.
OpenAI's safety_identifier parameter tags each request with a per-user hash. The Safety Usage Dashboard shows blocked requests by that field — turning safety events into routing, governance, and incident-response signals for API teams.
Qwen-AgentWorld simulation routing policy lets teams test MCP, terminal, web, and OS agents before real environments are at risk.
Qwen3.5-OCR DashScope routing gives document AI teams an OpenAI-compatible path, a richer native SDK path, and new policy questions for regions and fallback.
xAI Priority Processing adds service_tier=priority for Chat Completions and Responses, making latency a per-request routing and billing decision.
2.1.187 changelog: sandbox.credentials setting stops sandboxed commands from reaching credential files and secret env vars; org model restrictions now enforced at picker, --model flag, /model command, and ANTHROPIC_MODEL — with a visible restriction message.
Anthropic's Claude Tag on Slack introduces channel-scoped agent identities with per-channel tool access and org-level token spend controls — the same governance primitives AI engineering teams need for any multi-agent deployment.
Cursor 3.9 unifies plugins, skills, MCPs, subagents, rules, and hooks into a single Customize page with user, team, and workspace scoping — and team marketplaces now import from GitLab, Bitbucket, and Azure DevOps.
F5's new AI Security Platform adds network-based shadow AI discovery and MCP server connection tracing — closing the visibility gap that routing-layer operators have been managing with custom logging.
Google has made Model Armor on Agent Gateway generally available, embedding prompt-injection defense and content scanning directly into every agent traffic flow — with no changes to agent code required.
Samsung Electronics is deploying Codex to its entire global workforce — one of OpenAI's largest enterprise launches ever. Here is the routing and governance architecture every operator needs before reaching that scale.
OpenAI and Broadcom unveiled Jalapeño, OpenAI's first custom LLM inference accelerator. Here is what the shift to proprietary silicon means for API capacity, pricing trajectory, and multi-provider routing decisions for AI engineering teams.
Anthropic's Seoul office and Korean enterprise rollouts surface a pattern every global AI team faces: regulated industries cannot use the direct Claude API for in-region data residency. Here is what the Korea deployments reveal about routing architecture.
Claude Code 2.1.186 changes two defaults that affect every operator: bash commands now auto-trigger agent responses, and background subagents surface permission prompts instead of auto-denying. Here is what to audit before updating.
Envoy AI Gateway v0.7.0 ships hostname-based model catalog isolation for multi-tenant deployments, Anthropic-to-Bedrock protocol translation, and quota-aware rate limiting — a blueprint for how production AI routing infrastructure is evolving.
Codex 0.143.0 adds configurable rollout token budgets, per-thread multi-agent delegation policy, and indexed web-search with URL restrictions — three controls operators have needed since Codex went multi-agent.
OpenAI's Codex-maxxing whitepaper outlines how to run Codex as a persistent workspace across multi-hour sessions. For routing operators, the real takeaways are context budget management, step verification gates, and when to delegate vs. oversee.
Qwen Code v0.18 ships Agent Team mode — persistent parallel agents that message each other and share task lists — plus durable scheduled tasks and runtime provider switching. Here is what the release changes for teams evaluating multi-provider coding-agent routing.
Claude Code 2.1.185 reclassifies webhook and scheduled-task deliveries as task notifications, blocking them from approving pending actions in auto mode. Operators running CI/CD webhook pipelines and MCP-authenticated gateways need to review their automation boundaries.
Claude Code 2.1.181 introduces /config key=value syntax for changing any setting mid-conversation—no JSON files, no restarts. For teams routing Claude through a gateway, this changes how model selection, thinking mode, and managed settings are enforced at runtime.
Fable 5 returned June 18 with nationality-based access controls, tighter safety classifiers, and mandatory data retention. Today is the last day it's included in subscriptions — from June 23, credits are required at $10/M input and $50/M output tokens.
How Codex CLI rate-limit-reset-credits, remote executors, and encrypted Noise relays should route across hosts, MCP capabilities, policy, and billing.
Codex Record and Replay records a macOS workflow and turns it into a reusable skill. Govern each replay with approvals, permissions, fallback recovery, and cost telemetry.
Cursor model routing now covers Automations from /automate, Slack emoji triggers, GitHub review events, and cloud-agent computer use.
DashScope and Qwen Code docs split web search across Responses tools, Chat Completions enable_search, native DashScope source controls, and Bailian WebSearch MCP. Compare Alibaba Cloud Qwen API routing, citations, regional coverage, and budget governance.
"We're pausing the changes to Claude Agent SDK usage" — Anthropic reversed its June 15 billing split on launch day. What the credit rollback means for subscription vs. API routing, and how to prepare for the next billing pivot.
Google stopped accepting unrestricted Gemini API keys on June 19, 2026. Any key without explicit API restrictions now returns an error. Here is what operators routing to Gemini must audit in their gateway credential configurations before their next production call fails.
MiMo Code long-horizon coding agent routing turns Xiaomi's open-source terminal agent into an operator question: how should gateways route state, memory, and workflow compute?
OpenAI Assistants API shuts down August 26, 2026 — 56 days away. Official migration replaces Threads/Runs/Assistants with Conversations/Responses/Prompts. Checklist covers grep patterns, thread state, tool schemas, routing proxy compatibility, and Azure timelines.
Alibaba Cloud DashScope's wan2.7-t2v model supports text-, image-, and reference-to-video generation via async job APIs across mainland China, international, and US scopes. Pick your region, job polling strategy, and fallback paths before production.
xAI's grok-build-0.1 is now available via the xAI API in public beta — a purpose-built agentic coding model at $1/$2 per million tokens with 256k context, MCP support, and parallel-agent architecture. Here's how it fits into multi-provider routing decisions.
Anthropic's June 11 platform update adds response_inclusion to web_search_20260318 and web_fetch_20260318, letting operators drop consumed search result blocks from API responses and cut output token costs in multi-step agentic workflows.
Claude Code 2.1.183 tightens auto mode safety by blocking destructive git and infrastructure commands unless explicitly requested, closes a webhook trigger approval bypass, and fixes MCP auth stub exposure in headless mode.
Datadog's 2026 State of AI Engineering report finds that over 70% of organizations run three or more AI models in production. The data reveals what makes multi-model AI routing production-ready — and where teams are still flying blind.
Google's Gemini Enterprise reached GA this week for workflow agents and reusable skills — both behind an allowlist. Here is the architecture decision tree and what it means for engineering teams routing enterprise AI workloads.
GitHub availability dropped to 88.4% in June as AI coding agents drove commit volume 14× year-over-year. Microsoft confirmed routing traffic through AWS — a decision that redefines what multi-cloud means for your CI/CD pipeline.
OpenAI's ChatGPT Enterprise spend controls now expose per-user, product, and model credit analytics through a unified Cost API. For routing teams, the next job is reconciling ChatGPT/Codex credits with API ledger, fallback, and cross-provider budget policy.
Anthropic’s 400K Claude Code sessions show domain expertise doubles verified success and drives 5× output. Use the data for token budgets, model tiers, prompt caching, and governance.
Claude Code 2.1.181 fixes a silent prompt-caching regression that was inflating costs for every team running a custom ANTHROPIC_BASE_URL or Microsoft Foundry endpoint. It also enforces the five-level subagent depth cap for foreground agents.
Claude Code 2.1.181 rewrites the cross-session agent peer trust model from 'treat with suspicion' to 'collaborative teammate.' Peer agents can now act on requests without per-action skepticism — but escalation paths and permission laundering remain hard-blocked.
Gemini Memory Bank is Google's managed stateful memory layer for AI agents — now GA with global endpoints as of June 17. Here's how agent platform memory bank sessions work, the CMEK constraint, and the routing policy every team needs before deploying stateful agents.
Claude Code 2.1.179 patches three silent failure modes that specifically affect teams routing through a custom API gateway: a 401 auth regression in claude agents workers, compaction ignoring the configured fallback model, and mid-stream response loss on connection drops.
Zhipu GLM-5.2 is now accessible through OpenRouter, BigModel API, and DashScope: 1M-token lossless context, 128K max output, reasoning_effort, context caching, and FrontierSWE-class performance for long-horizon coding-agent routing.
Moonshot AI's Kimi K2.7 Code cuts reasoning-token usage 30% and adds a 180–260 TPS HighSpeed variant — two changes that force a concrete routing decision for teams already sending coding agent traffic to Kimi.
LiteLLM's engineering team articulated a structural shift: the AI gateway is no longer just a model call router — it is becoming a control plane for agent sessions across Claude Managed Agents, Bedrock AgentCore, and Vertex. Here is what that means for routing.
OpenAI has deprecated six first-generation gpt-5 and o3 model snapshots, all shutting down December 11, 2026. Teams still routing to date-stamped IDs have six months to migrate — here's the full audit checklist.
Claude Code 2.1.178 ships Tool(param:value) permission grammar — letting operators write rules like Agent(model:opus) to block Opus subagents — plus nested .claude/ directory scoping and a pre-flight auto-mode classifier that screens subagent spawns before they run.
Xiaomi MiMo completes its V2 → V2.5 migration on June 18: mimo-v2-flash and mimo-v2-tts auto-route to V2.5 successors at V2.5 pricing. Old model IDs in your routing config will silently change behavior and billing.
Anthropic's Agent SDK billing split is live as of June 15. The new per-user monthly credit ($20–$200 by plan) draws from a separate meter — but the ceiling changes the routing decision for every production agent team.
Claude Code 2.1.176 fixes background agent session management, adds an auto-mode Fable 5 fallback for orgs without Opus 4.8, corrects hook path conditions for Read/Edit/Write policies, and ships footerLinksRegexes for managed deployments.
Claude Code 2.1.177 changelog: stdio MCP calls hang when claude -p loads a full fleet, CLAUDE_CODE_OAUTH_TOKEN is shadowed by stale on-disk credentials causing 401s, /compact is forwarded as plain text, and the Windows build cannot launch. Fleet pin and fix guide.
Gemini CLI stops serving consumer requests on June 18, 2026. Before you migrate to Antigravity CLI by reflex, understand what the unified-backend model means for provider independence and how multi-provider routing with Claude Code or Codex CLI changes the calculus.
Claude Fable 5 ships a new tokenizer that produces ~30% more tokens from the same text versus pre-Opus-4.7 models. Here is what AI engineering teams must audit before migrating: token budgets, cost models, cache boundaries, and max_tokens headroom.
Cursor's official Bugbot docs and June changelog now point to Composer 2.5, pre-push /review, and Cloud Agents formerly called Background Agents. For routing operators, the key line is still model block lists: AI code review now needs explicit governance.
A practical permissions.json autorun schema for Cursor SDK 0.1.6: how autoRun.allow_instructions, autoRun.block_instructions, local.autoReview, local.customTools, JsonlLocalAgentStore, requestId, and nested subagents fit together.
DeepSeek's official awesome-deepseek-agent docs show the Claude Code ANTHROPIC_BASE_URL setup, Anthropic-compatible API path, V4 Pro/Flash model IDs, and OpenCode/Copilot integration notes for coding-agent routing.
nex-agi's Nex-N2-Pro hits SiliconFlow with 262K context, $0.00 pricing, and Terminal-Bench 2.1 scores that beat Claude Opus 4.7, DeepSeek-V4-Pro, and GLM-5.1. Here's when to route to it.
OpenAI Partner Network launched June 14, 2026 with Select, Advanced and Elite tiers, $150M in ecosystem investment, and Codex specializations. Architecture diagram and checklist for direct OpenAI API keys vs SI-managed delivery, quota ownership, and fallback control.
Claude Fable 5 and Mythos 5 are now returning 404 on the Anthropic API after a US export control directive. Here is what routing operators must do in the next 24 hours.
Apple's WWDC 2026 Foundation Models update ships a unified Swift API routing to on-device models, Private Cloud Compute, Claude, and Gemini from one call site — introducing a new provider-routing boundary engineering teams need to manage.
Before rolling Claude Code 2.1.177 across a managed fleet, verify the 2.1.176 fixes: availableModels now catches ANTHROPIC_DEFAULT_*_MODEL aliases, Bedrock awsCredentialExport follows STS Expiration, and Remote Control no longer swaps session models silently.
Moonshot AI released Kimi K2.7 Code on June 12. The Kimi K2.7 Code API enforces thinking mode exclusively — disabling it returns an API error — which changes cost modeling, routing policy, and fallback logic for teams using coding agents.
MiMo Code is Xiaomi's open-source coding agent for long-horizon work. Learn how Max Mode, Goal verification, worker/judge routing, and OpenAI-compatible backends change API gateway cost and reliability planning.
Google deprecated Vertex AI Extensions on May 26, 2026, with a hard shutdown on November 26. Teams routing agent workflows through Google must migrate their Code Interpreter, Search, and Custom Extensions to the Gemini Enterprise Agent Platform before the deadline.
TheRouter's legacy /v1/anthropic/messages paths are deprecated and will be removed on September 3, 2026. The canonical path is /v1/messages — the same path official Anthropic SDKs use. Migration is a one-line base-URL change.
Claude Code v2.1.175 adds enforceAvailableModels, a managed setting that constrains the Default model to your availableModels allowlist and prevents project or user settings from widening it — closing a silent governance gap for enterprise teams.
The June 12 Claude Code release adds per-skill, per-agent, and per-MCP-server cost attribution to the VSCode /usage dialog, fixes a Bedrock GovCloud bug, and closes a background-session credential leak.
Fable 5 was the first Claude model with built-in refusal handling. Anthropic's server-side fallback API collapses refusal + retry into a single call, preserving billing credit and returning one response — architecture that still matters for routing teams post-suspension.
Google's Gemini 3.5 Live Translate is now in public preview via the Gemini API, delivering continuous speech-to-speech translation across 70+ languages. Here is what the new model ID means for voice-agent routing teams.
OpenAI is acquiring Ona to bring persistent, customer-controlled cloud execution into Codex. For teams running API gateways and agent orchestration, this changes how you think about session lifecycle, cost modeling, and credential governance.
Vertex AI is not deprecated, but Google is retiring every Imagen 2.x/3.x/4.x and Veo 2.x/3.0 GA endpoint on June 30, 2026. Use this shutdown list to migrate image and video routing to gemini-2.5-flash-image and veo-3.1 endpoints.
Anthropic Bedrock Mantle adds Claude Messages API compatibility on Amazon Bedrock, including SigV4 auth, SSE streaming, bedrock-2023-05-31 semantics, and routing tradeoffs for Claude Code teams.
Anthropic's code execution tool now tells Claude about its 90-second per-cell time limit before it writes any code — reducing detection_timeout failures. What changed, which models support it, and the pipeline fixes operators need (including Haiku 4.5 routing gotchas).
Anthropic's June 10 Claude Code release unlocks five-level sub-agent nesting. Here's what that means for model routing, cost architecture, and governance in agentic coding workflows.
Anthropic released claude-fable-5 on June 9 — the most capable widely released Claude yet. Here is what changes in your AI routing layer: new model ID, adaptive thinking always-on, stop_reason refusal, a fallbacks API parameter, and mandatory 30-day data retention.
Anthropic’s Claude Managed Agents update adds scheduled deployments, Vault environment-variable credentials, and session_thread_id webhooks so teams can replace external cron jobs, stop passing secrets through prompts, and route multi-agent events with clearer correlation.
OpenAI Responses API web_search now returns image results with text citations. When to migrate from web_search_preview and route visual grounding workloads.
DashScope now recommends qwen3.7-plus as the balanced default for OpenClaw, Claude Code, and gateway routes. Compare qwen3.7-max, qwen3.7-plus, and qwen3.6-flash before changing Responses API reasoning.effort.
Project Polaris replaces GPT-4 Turbo as the GitHub Copilot default in August 2026. Microsoft Build 2026 also expanded Azure AI Foundry to OpenAI, Anthropic, Cohere, DeepSeek, Mistral, Meta, and xAI—with on-prem deployment via Azure Arc.
Qwen3Guard is Alibaba's open-weight safety guardrail family: Qwen3Guard-Stream (real-time token-level verdicts) and Qwen3Guard-Gen (post-generation review), in 8B/4B/0.6B sizes. Runs at the gateway as a provider-agnostic filter for multi-model routing.
Claude Code's June 8 release adds --safe-mode, fixes slow claude -p / --print runs on Windows, hardens managed settings MCP policy enforcement, and restores a 5-minute Vertex / Foundry idle timeout.
Snowflake is now a Claude Fable 5 inference provider — with a Bearer token credential model, a Snowflake-specific base_url, and SQL-native AI_COMPLETE access. Here is what routing teams need to know before adding it to their stack.
Xiaomi's MiMo-V2.5-Pro-UltraSpeed reaches 1000 tokens per second on a 1T-parameter model via FP4 quantization and DFlash speculative decoding. API access opens June 9–23 at 3× the base price. Here is the routing cost-speed framework operators need.
Codex rust-v0.138.0: .mcp.json plugin automation, MCP OAuth pre-refresh, and /app desktop handoff for macOS/Windows fleets.
DashScope made wan2.7-image-pro its recommended default: the only image endpoint with 4K output, text rendering, brand color, character consistency, and multi-image editing in one model ID. Routing decision framework vs qwen-image-2.0-pro and z-image-turbo.
Anthropic's Claude Opus 4.1 deprecation retires claude-opus-4-1-20250805 on August 5, 2026. Use this router-focused migration guide to replace aliases, audit fallback tiers, and avoid API failures.
Claude Code v2.1.166 upgrades fallbackModel to support up to three ordered fallback models for overload resilience, extends it to interactive sessions, and hardens multi-agent trust: relayed SendMessage calls no longer carry user authority.
DeepSeek retires deepseek-chat and deepseek-reasoner on July 24, replacing them with explicit V4 model names and a new Anthropic API endpoint. Here is what changes for AI engineering teams routing through DeepSeek today.
Kimi Agent Swarm GitHub open weights and Kimi Code workflows make Kimi K2.6 open-source coding routing a 300-sub-agent option.
Access Xiaomi MiMo V2.5 Pro, Standard, Flash, and ASR through one TheRouter API key, with vendor-direct routing, reasoning controls, and unified billing for text and speech workloads.
MAI-Thinking-1 is Microsoft's first proprietary 35B MoE reasoning model — distillation-free, matches Claude Opus on SWE-Bench, on Azure Foundry. MAI-Code-1-Flash powers Copilot auto-routing with 60% fewer tokens. A routing guide for teams on Claude or GPT-4o.
Antigravity 2.0 replaces Gemini CLI, adds Managed Agents via the Interactions API with isolated Linux environments and session billing. CLI migration checklist, SDK self-hosting path, and routing architecture impact for multi-provider teams.
Google's $920M/month SpaceX GPU lease shows Gemini capacity is not infinite. Here is how AI teams should model correlated provider failures, Q4 2026 quotas, and safer multi-provider fallback paths.
Step-by-step Claude Code workload identity federation setup: connect an OIDC issuer, create Anthropic service accounts and federation rules, then replace static sk-ant keys in CI/CD, gateways, and agent runtimes with short-lived tokens.
Claude Code now refuses to launch when its version falls outside an operator-defined range. requiredMinimumVersion and requiredMaximumVersion managed settings let teams lock their entire dev fleet to approved Claude Code builds — no voluntary upgrades, no accidental drift.
Meta has delayed the Muse Spark developer API twice with no firm launch date—nearly two months after promising it 'soon.' For teams that built routing plans around a closed-source model, this is the exact scenario that makes multi-provider fallback coverage non-negotiable.
OpenAI now returns moderation scores for both input and generated output inside the same Responses API call. For teams routing across multiple providers, this eliminates the separate safety round-trip and changes how you architect content governance at the gateway layer.
Palo Alto Networks closed its acquisition of Portkey, folding the AI gateway into Prisma AIRS as the enterprise control plane for all LLM traffic. Here's what this means for teams evaluating independent routing layers.
Claude Code auto mode now runs on AWS Bedrock, Google Vertex AI, and Microsoft Foundry — set CLAUDE_CODE_ENABLE_AUTO_MODE=1 and your coding agent runs inside your VPC. Here's what changes for billing, latency, and fallback routing when you switch from the Anthropic API.
DashScope is Alibaba's managed API platform for Qwen models. With qwen-image-2.0-pro, it introduces an async job pattern distinct from DALL-E: submit a request, get a task ID, poll to retrieve the result.
GitHub Copilot's June 1 token-billing switch is landing real bills — some users report projected costs of $754–$847/month vs. a prior $39–$44 flat rate. This is what the May 2026 pricing shift looks like in production, and why routing architectures matter.
Anthropic analyzed 832 malicious accounts across 14 MITRE ATT&CK tactics and 13,873 attack actions. The new ARiES risk framework redefines how API operators should detect, govern, and rate-limit agentic threats — starting with the interface your users connect through.
Claude API now bills zero tokens for stop_reason=refusal with no output, and the advisor tool accepts a per-call max_tokens cap. Operator checklist: skip zero-output refusal charges, expose advisor max_tokens in routing config, and handle stop_details.category for triage.
Claude Code MCP timeout and WebFetch policy changed in the June 3 release: explicit WebFetch deny rules now beat preapproved hosts, while sub-1000 ms MCP timeout values fall back to MCP_TOOL_TIMEOUT/default instead of tripping the watchdog.
OpenAI Codex CLI 0.137.0 introduces multi-agent v2 with per-thread runtime routing, letting each spawned sub-agent carry its own model and provider choice. Here's what changes for operators managing gateway routing, cost attribution, and governance.
OpenAI's evals platform deprecation, Agent Builder shutdown, and v1/prompts API removal all land November 30, 2026. Here's what operators need to migrate and how the routing layer absorbs the gap.
OpenAI now defaults eligible GPT-5 prompt cache retention to 24h for non-ZDR organizations. Learn how GPT-5.5, prompt_cache_key, and gateway routing choices affect API cost and data policy.
Claude Partner Network Services Track and Partner Hub formalize Select, Preferred and Global Premier implementation partners. Here is what Claude partner hub signals mean for API routing, fallback policy and enterprise Claude deployments.
The June 2 Claude Code release propagates OTEL_RESOURCE_ATTRIBUTES as labels on every metric datapoint, enabling per-team cost allocation at the Prometheus level without gateway-side custom instrumentation.
OpenAI Codex now ships six role plugins—analytics, creative, sales, design, investing, banking—plus Sites, a hosted web-app generator. For routing teams, the multi-role shift changes context budgets, tool call patterns, and upstream model requirements.
OpenAI container sessions now use per-minute billing with a 5-minute minimum for Hosted Shell and Code Interpreter. This routing cost model explains when to reuse container_reference, avoid fresh container_auto sessions, and right-size memory tiers.
Official OpenAI gpt-image-2 migration checklist for teams still using gpt-image-1, gpt-image-1.5, gpt-image-1-mini, or chatgpt-image-latest before the Q4 2026 API shutdown.
Anthropic's Project Glasswing now reaches 150 critical-infrastructure organizations, and Claude Security—a new Opus 4.8-powered code-scanning product—is publicly available. Here's what the scale-up means for teams routing to Claude.
Claude Code's June 2 release adds interactive prompts before writing shell startup files and build-tool configs in acceptEdits mode — a change that can silently break unattended CI pipelines that rely on auto-merging AI-generated configuration changes.
Amazon Bedrock now serves OpenAI GPT-5.5, GPT-5.4, and Codex through AWS IAM credentials, regional endpoints, and model IDs such as openai.gpt-5.5. Here is what changes for routing layers, Codex config, fallback policy, and OpenAI-compatible gateway design.
Anthropic confidentially submitted its S-1 to the SEC on June 1, setting the stage for an IPO. For API operators, the shift from private to public company changes the calculus around pricing stability, SLA accountability, and BYOK governance.
Anthropic closes a $65B round at a $965B valuation and discloses $47B annualized revenue and 15+ GW of committed compute capacity. For teams running Claude in production, this changes the capacity confidence calculus for routing, fallback, and enterprise SLA decisions.
Anthropic retires claude-sonnet-4-20250514 and claude-opus-4-20250514 on June 15, 2026. Any routing config still pointing to these model IDs will start returning errors in two weeks. Here's the migration map.
DashScope OpenAI Responses API support is live on compatible-mode/v1 for Qwen3.7-max, qwen3.6-plus, qwen3-coder-plus and more — stateful previous_response_id, built-in web search and code interpreter, reasoning.effort control, four global regions.
Kimi's Agent Swarm deploys up to 100 parallel sub-agents per task. For teams routing to Kimi or building similar multi-agent systems, this changes the concurrency, billing, and rate-limit math entirely.
MiniMax M3 launched today as the first open-weights model combining frontier coding, 1M-token context, and native multimodality. Its dual billing tiers and thinking-mode switch are the routing decisions engineering teams need to handle now.
Dynamic workflow 是什么 in Claude Code? It is a JavaScript orchestration script for dozens to hundreds of subagents, changing token accounting, rate-limit bursts, approvals, and gateway governance.
Claude Code v2.1.157 introduces agent dispatch routing via settings.json, auto-loaded .claude/skills plugins, and OTEL tool-parameter telemetry—three operator controls for governed multi-agent coding pipelines.
OpenAI published a detailed guide on designing trustworthy third-party evaluations. The key operator lesson: harness choice changes measured capability, so benchmark scores require context before they can inform routing tier decisions.
OpenAI's April 22 deprecation notice covers 25+ model IDs across two hard shutdown dates — July 23 and October 23, 2026. If your routing config still references gpt-4, gpt-3.5-turbo, o1, o3-mini, or o4-mini, you have a breaking change on the calendar.
OpenAI workload identity federation lets CI/CD pipelines exchange OIDC tokens for short-lived API keys — no stored credentials. The Admin API adds model allowlists per project, spend alerts, and granular billing line items. Full operator checklist inside.
StepFun's Step 3.7 Flash landed on DashScope on May 29 as a multimodal Flash-tier model with optional thinking mode. The operator catch: thinking is toggled via extra_body, not a standard OpenAI parameter — a pattern that breaks naive routing proxies.
Anthropic's May 2026 webhooks release adds event-driven session lifecycle control, multi-agent orchestration, and self-hosted sandboxes to Managed Agents — what routing and platform teams need to know.
Codex Computer Use now works on Windows, with remote steering and Profiles for usage tracking. For AI engineering teams, the key shift is host-level routing: Windows agents need OS-aware identity, token accounting, and remote-control guardrails.
Anthropic's Opus 4.8 ships with an effort control parameter, mid-turn system entries, and a 3× cheaper fast mode. Each change has direct implications for how routing teams configure model selection, token budgets, and cost policies.
OpenAI Frontier Governance Framework signals what API operators need before EU AI Act enforcement: provider due diligence, compliant fallback policy, and per-request model audit trails.
Claude Code's May 28 patch fixed an OAuth credential regression that could send user Anthropic tokens to custom API gateways. Here is the gateway audit checklist for token isolation, logging, and subagent MCP policy enforcement.
Codex eval infrastructure separates a working agent from a self-improving one. OpenAI's Tax AI converted production traces into eval targets and ran Codex-driven iterations automatically. The architecture applies to any team building agents on an API gateway.
An unnamed OpenAI reasoning model autonomously disproved an 80-year-old conjecture in discrete geometry. For operators, the signal is less about math and more about what 'reasoning tier' now means for model selection and routing policy.
Claude Code's --fallback-model can now auto-switch after model-not-found errors. Here's how teams using an API router should separate client fallback, gateway fallback, OTEL entrypoint metrics, and plugin governance.
Gartner's inaugural Magic Quadrant for Enterprise AI Coding Agents formally encodes what enterprise procurement will demand: multi-model flexibility, governance controls, sandboxed execution, and auditability. Here is what those criteria mean at the routing layer.
Anthropic's Glasswing partners used Claude Mythos Preview to find 10,000+ critical vulnerabilities in a month. Here is the routing and operator lens on what this new AI security phase means for your infrastructure.
Zhipu AI's GLM-5.1-highspeed delivers 400 tokens per second via the TileRT inference engine — the same flagship capability as GLM-5.1 but with a throughput profile that reshapes routing decisions for real-time agent and coding workloads.
OpenRouter now processes 25 trillion tokens per week—a 5x jump in six months. Here is what that weekly token volume means for your routing architecture, provider lock-in risk, and team governance.
The Vertex AI generative AI SDK is deprecated — its modules are removed June 24, 2026. If your team routes to Google via vertexai.generative_models or related imports, here's exactly what to migrate to keep production code working.
Anthropic has confirmed Claude Mythos Preview exists above Opus 4.7—but access is controlled via waitlist and cybersecurity review. Here's what the Mythos tier means for your routing policy, upgrade path from Opus 4.6, and fallback thresholds.
DashScope API endpoints split into two incompatible paths: standard dashscope.aliyuncs.com (sk- key) and Coding Plan coding.dashscope.aliyuncs.com (sk-sp- key).
GitHub Copilot AI Credits launched June 2026, replacing Premium Request Units with token-based billing. The old model fallback that let you keep working on a cheaper model is gone—agentic sessions now stop when credits run out. Here's what changed and how to manage it.
Xiaomi MiMo-V2.5-Pro is live on DashScope with 1M-token context, OpenAI and Anthropic API compatibility, and 42% fewer tokens than Kimi K2.6 on agent benchmarks. Evaluate whether this model fits your routing policy—cost, context window opt-in, and DashScope catalog risks.
MiniMax M2.7 delivers 56% SWE-Pro performance near Claude Opus at 1/10th the cost, with both OpenAI and Anthropic API formats — ideal for coding agent routing.
Deciding between Qwen3.5-Omni S2S and ASR-LLM-TTS pipelines on DashScope commits your voice AI architecture. Compare latency, regional endpoints, fallback strategies, and cost to pick the right routing approach.
Anthropic official Claude Code best practices guide (code.claude.com): context management and context engineering are first-class constraints for routing teams. Context windows fill fast, performance degrades as they fill, coding agents burn tokens 3-10× vs chat.
The Codex admin sandbox is an OS-level write-restricted process that runs automation without Full Access (unsafe mode). OpenAI's Windows design uses write-restricted tokens and synthetic SIDs — no admin elevation needed. Here's what it means for every coding agent on your stack.
DeepSeek V4 Pro's 75% price cut is now permanent — $0.87/M output tokens with 80%+ SWE-bench scores. Here's why it shifts V4 Pro from premium fallback to primary reasoning tier and what that means for your routing cost structure.
Gemini 2.0 Flash and Flash-Lite hit end-of-life June 1, 2026 — the shutdown is done. Here is what model IDs were removed, which replacements (gemini-2.5-flash-lite, gemini-3.1-flash-lite) actually work as substitutes, the real cost delta, and the fastest way to audit…
OpenAI and Dell are bringing Codex inside enterprise data centers. When your coding agent runs on-prem, the routing architecture changes: data locality, session affinity, cost accounting, and governance controls all need rethinking.
A Codex dispatch gateway checklist for enterprise teams using programmatic access tokens, Remote SSH, and CI agents: route devbox egress, separate token ledgers, and catch background-agent spend before it becomes invisible.
Gemini 3.5 Flash launched at $1.50/$9 per million tokens — up to 6x the cost of earlier Flash models — and the new Interactions API changes how routing teams should handle Gemini Flash fallbacks and session continuity.
What is DashScope Qwen and how does the API work? DashScope is Alibaba Cloud's OpenAI-compatible platform for Qwen models. Learn how the DashScope Qwen API works, what DashScope is used for, and the current Qwen3.6-plus / Qwen3.6-flash model tier for routing.
How to use DeepSeek Anthropic API at api.deepseek.com/anthropic with Claude Code and the Anthropic SDK — plus the routing hazards around unsupported fields, cache_control, MCP tools, and silent fallback to deepseek-v4-flash.
Anthropic has acquired Stainless, the company that generates every official Claude SDK and MCP server tooling. For teams building multi-provider API pipelines, this reshapes SDK dependency risk, MCP server governance, and the pace of Claude API surface changes.
Kimi K2.6 API pricing is ¥1.10 cache-hit input, ¥6.50 cache-miss input, and ¥27 output per 1M tokens, with KVV verification, 256K context, and OpenAI-compatible routing for coding agents.
Alibaba Cloud's new Qwen-MT turbo model arrives via OpenAI-compatible endpoints, but its translation controls live inside extra_body — a pattern that breaks any middleware that strips non-standard fields. Here's what routing teams need to watch.
Full Qwen3.7-Max benchmark details: 69.7 on Terminal Bench 2.0, 92.4 on GPQA Diamond, top SWE-Pro scores, and a 35-hour autonomous agent run on the T-Head ZW-M890 PPU hardware. What these performance numbers mean for routing teams.
DeepSeek's official integration guide for Claude Code, OpenCode, and OpenClaw — API setup, per-tier V4 model routing, and Anthropic-compatible endpoint configuration for coding agents.
DeepSeek API model names changed 2026: deepseek-chat becomes deepseek-v4-flash and deepseek-reasoner merges into deepseek-v4-flash (thinking mode) or deepseek-v4-pro (1M context). Deprecated July 24, 2026. New Anthropic-compatible endpoint at /anthropic for Claude Code routing.
Anthropic's Managed Agents now lets teams run tool execution in their own infrastructure while keeping orchestration on Anthropic's side. Here's what that split-plane model means for routing, compliance, and private MCP access.
OpenAI is now a C2PA conforming generator with Google SynthID watermarks in every API-generated image. Every output carries a cryptographic provenance fingerprint — affecting audit trails, fallback routing, and enterprise content governance.
Qwen's GSPO shifts optimization from token-level to sequence-level, eliminating Routing Replay overhead and enabling stable large-scale RLHF training for Qwen3 models.
A reported US review process for frontier AI models would make provider availability, release timing, and governance evidence part of routing strategy—not just policy news.
AI-generated images of colorful Audemars Piguet Royal Oak wristwatches flooded social media for a week before Swatch's actual Royal Pop pocket watch collaboration was revealed, creating unprecedented fake product hype.
OpenAI sought to present a jackass trophy in the Musk v. Altman trial as evidence of Elon Musk's behavior toward employees who challenged his safety priorities in 2018.
OpenAI CEO Sam Altman testified in the Musk v. Altman trial, revealing Elon Musk's suggestion to pass OpenAI control to his children and defending against allegations of deceptive behavior.
WhatsApp introduces Incognito Chat for Meta AI conversations, using Private Processing to prevent Meta from accessing user interactions with the chatbot.
The Musk v. Altman trial closing arguments reveal how OpenAI's nonprofit structure conflicted with competitive pressures, raising questions about public interest protection.
Sasha Luccioni launches Sustainable AI Group to address growing enterprise and regulatory demands for AI emissions data and energy efficiency metrics across European and global markets.
Codex supply chain attack traced to compromised TanStack npm package: OpenAI rotated code-signing certificates for Codex CLI, ChatGPT Desktop, and Atlas after Mini Shai-Hulud campaign hit two employees. macOS developers have until June 12, 2026 — update or your apps won't launch.
Former OpenAI chief scientist Ilya Sutskever defended his role in Sam Altman's ouster during testimony in the Musk v Altman trial, stating he acted to protect the company from destruction.
Security researchers discovered over 5,000 AI-coded apps from Lovable, Replit, Base44, and Netlify exposing sensitive corporate and personal data without authentication or security measures.
Claude Managed Agents announcement and launch details: Anthropic's hosted API runtime handles tool use, persistent memory, sub-agent orchestration, and observability so enterprise teams ship production agents without managing the runtime themselves.
Anthropic doubled Claude Code rate limits and raised Opus API ceilings with 220K GPUs at SpaceX Colossus 1. Learn how this changes fallback triggers, scheduling, and multi-provider routing strategy.
More than 70 civil liberties organizations demand Meta abandon facial recognition plans for Ray-Ban smart glasses, warning the 'Name Tag' feature would enable stalkers and predators to identify strangers in public.
