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.
OpenAI's new org-level key creation controls let admins enforce service-account-only, project-keys-only, or disable all key creation. The first structural control that prevents shadow routing via personal API keys.
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.
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.
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 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.
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.
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.
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.
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.
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.
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.
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.
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.
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 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.
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.
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 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.
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.
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.
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.
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.
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?
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.
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.
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).
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 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.
