DeepSeek Deprecates deepseek-chat Model Name: V4 Flash Migration and the New Anthropic API Routing Playbook

DeepSeek officially retires the deepseek-chat and deepseek-reasoner model names on July 24, 2026, at 15:59 UTC. This is not a model sunset — the underlying V4 Flash engine stays live — but it is a namespace change that breaks every routing configuration, SDK integration, and cost model still referencing the legacy names. At the same time, DeepSeek has launched a dedicated Anthropic API format endpoint at https://api.deepseek.com/anthropic, creating a new routing path for Claude Code and other Anthropic-native clients. For AI engineering teams managing multi-provider routing, this dual change demands migration work before the July 24 deadline.

What changed in the DeepSeek model catalog

The DeepSeek API now exposes two explicitly named model tiers:

The legacy aliases deepseek-chat and deepseek-reasoner map as follows until July 24:

• deepseek-chat → non-thinking mode of deepseek-v4-flash
• deepseek-reasoner → thinking mode of deepseek-v4-flash

After the deprecation date, requests using the old model IDs will be rejected by the API. The pricing page is explicit: deepseek-chat and deepseek-reasoner are aliases, not independent models. They exist for backward compatibility only.

The Anthropic API endpoint changes how coding agents route

The bigger structural change is DeepSeek's new Anthropic-format endpoint. The base URL https://api.deepseek.com/anthropic accepts Anthropic SDK calls natively, with automatic model-name mapping:

• Claude model names starting with claude-opus → routed to deepseek-v4-pro
• Claude model names starting with claude-haiku or claude-sonnet → routed to deepseek-v4-flash

This means a Claude Code user can switch the ANTHROPIC_BASE_URL to DeepSeek and keep using the Anthropic Python SDK unchanged. The API translates Anthropic message formats, tool calls, and streaming into DeepSeek-native execution.

What works and what does not

The compatibility matrix is documented in the DeepSeek Anthropic API guide. Key points:

• Fully supported: max_tokens, system, messages (text and tool_use), stream, temperature, thinking (budget_tokens ignored), stop_sequences, tool_choice, top_k, top_p, json output
• Partially supported: metadata.user_id (used for rate-limit isolation; other fields ignored), effort in output_config
• Not supported: type: "image" messages, type: "document" messages, type: "redacted_thinking", MCP tools, container uploads, code execution tools
• Ignored: anthropic-beta, anthropic-version, container, mcp_servers, service_tier, cache_control on tools/messages

The practical implication: DeepSeek's Anthropic endpoint covers the core coding-agent workflow (text, tools, streaming, thinking) but does not support Claude-specific features like image input through messages, MCP tool orchestration, or redacted thinking blocks. Teams running Claude Code with heavy document or image context may see unexpected behavior if they route blindly through DeepSeek's Anthropic endpoint.

Why this matters for AI engineering teams

1. Routing configuration audit required

Every AI gateway, proxy, SDK config, or preset that maps "model": "deepseek-chat" or "model": "deepseek-reasoner" needs updating before July 24. This includes:

• Gateway routing presets and fallback chains
• Coding-agent configuration files (Claude Code, Cursor, GitHub Copilot with DeepSeek plugin)
• Cost dashboards and billing reconciliation systems that aggregate by model ID
• Automated test suites that assert specific model ID values

The migration is mechanical — replace deepseek-chat with deepseek-v4-flash and deepseek-reasoner with deepseek-v4-flash (thinking mode) — but the July 24 deadline is hard. There is no grace period in the documentation.

2. The cost structure shifts with explicit model names

Teams using deepseek-chat today are using V4 Flash at its non-thinking price. After migration to deepseek-v4-flash, the default thinking mode is enabled, which uses more tokens per request. The pricing table shows V4 Flash at $0.14/1M input (cache miss) and $0.28/1M output — the same as before, but the thinking-mode token overhead means actual spend per request may increase if the routing layer does not explicitly set thinking: { "type": "disabled" } for tasks that do not need reasoning.

Conversely, V4 Pro at $0.435/1M input and $0.87/1M output represents a meaningful step up in both capability and cost. Teams that route by task complexity (cheap model for simple tasks, expensive model for hard problems) now have a clearer two-tier pricing story with DeepSeek.

3. Anthropic API compatibility creates a new fallback vector

For operators running Claude as the primary model, the Anthropic-compatible endpoint on DeepSeek becomes a structural fallback option. Instead of maintaining a separate OpenAI-format integration for DeepSeek fallback, teams can now point the same Anthropic SDK client at DeepSeek with only a base URL and API key change.

This matters for:

• Latency failover: If the Anthropic API is slow or rate-limited, rerouting to DeepSeek's Anthropic endpoint provides the same SDK experience at a different cost/latency profile.
• Cost optimization: Routing Claude Haiku/Sonnet-bound traffic to DeepSeek V4 Flash at $0.14/1M input versus Anthropic's published pricing.
• Geography: DeepSeek's API endpoints may have lower latency from mainland China infrastructure.

However, the compatibility limitations (no image messages, no MCP, no redacted thinking) mean this is not a universal drop-in. Teams need feature-flag routing logic that checks whether the request uses Anthropic-only features before falling back to DeepSeek.

4. Coding-agent ecosystem implications

DeepSeek's own agent integration guide shows explicit support for Claude Code, GitHub Copilot, and OpenCode as supported tools. The Anthropic API endpoint makes the Claude Code integration particularly simple: set ANTHROPIC_BASE_URL=https://api.deepseek.com/anthropic and ANTHROPIC_API_KEY to a DeepSeek key, and Claude Code sends traffic to DeepSeek V4 models.

For teams that want Claude Code's UX but DeepSeek's pricing — especially in development environments — this is a viable configuration. The risk is feature parity: Claude Code may invoke MCP tools or image analysis that DeepSeek's Anthropic endpoint does not support, causing silent failures.

What to watch in your routing layer

Here is a checklist for AI engineering teams:

1. Audit model ID references. Search your codebase, gateway configs, and infrastructure-as-code for deepseek-chat and deepseek-reasoner. Replace with deepseek-v4-flash (or deepseek-v4-pro for capability-tiered routing). Target: before July 24, 2026.

2. Test thinking-mode cost impact. If you use deepseek-chat (non-thinking) for high-volume cheap tasks, explicitly disable thinking in your requests after migration to avoid unexpected token overhead.

3. Evaluate Anthropic API fallback. If you run Claude as primary, test DeepSeek's Anthropic endpoint as a fallback for non-image, non-MCP workloads. Validate tool-call parity and streaming behavior before enabling in production.

4. Update cost dashboards. Billing systems that aggregate by model ID will need the new deepseek-v4-flash and deepseek-v4-pro identifiers. Historical data under deepseek-chat and deepseek-reasoner should be mapped to V4 Flash for trend continuity.

5. Monitor concurrency limits. V4 Flash allows 2500 concurrent requests; V4 Pro allows 500. If your routing layer load-balances across both tiers, the concurrency asymmetry matters for burst traffic.

The broader pattern: namespace deprecations are accelerating

This is the third major model-namespace deprecation in June 2026. Anthropic deprecated Claude Opus 4.1 on June 5, setting an August 5 retirement. OpenAI is retiring legacy model IDs on a rolling schedule through October. The common thread: providers are consolidating their model catalogs from friendly aliases to explicit versioned model IDs, and the transition windows are short.

For teams running multi-provider routing, the operational lesson is clear: routing layers that hard-code model-name strings without a deprecation-aware migration strategy will break repeatedly as providers standardize their naming. The fix is not to chase every alias — it is to build routing logic on capability tiers (fast/cheap, balanced, max-capability) and map to whatever model ID the provider currently supports, with automated deprecation detection from the provider's pricing or model-list endpoint.