DeepSeek deepseek-chat Retires July 24: The Gateway Routing Audit Every Operator Must Run This Week

Seven days from now — July 24, 2026 at 15:59 UTC — DeepSeek's legacy model aliases deepseek-chat and deepseek-reasoner will stop accepting requests. Any routing rule, provider config, or hardcoded SDK call that still points to these names will return errors. The fix is a one-line model-name change, but the audit to find every place those strings appear across a production AI gateway is not.

This is not a speculative announcement. The retirement date has been documented on DeepSeek's official API changelog since April and is now confirmed on the live API landing page. The 90-day grace period is ending.

What Is Actually Changing

When DeepSeek launched V4 in April, the two legacy names were remapped as soft aliases:

• deepseek-chat → deepseek-v4-flash (non-thinking mode)
• deepseek-reasoner → deepseek-v4-flash (thinking mode)

After July 24, both aliases are removed. Calls using these model identifiers will fail — they will not silently route somewhere else. The replacement model identifiers are:

• deepseek-v4-flash — fast, economical, 1M context; supports both thinking and non-thinking modes via the thinking parameter
• deepseek-v4-pro — 1.6T parameter frontier model; optimized for complex reasoning, agentic coding, and long-context workloads

Both support the OpenAI ChatCompletions interface and the Anthropic API interface, both support 1M token context, and both expose the reasoning_effort parameter introduced with V4.

Why This Matters for AI Engineering Teams

The operational risk here is not the migration itself — deepseek-chat → deepseek-v4-flash is a string substitution. The risk is incomplete discovery: production systems that reference the deprecated alias in more places than engineers expect.

Common sources of hidden alias references:

• Routing table entries — gateway or proxy configs that specify the model name as a routing key, separate from the SDK call
• Fallback chains — routing policies that fall back from a primary model to deepseek-chat or deepseek-reasoner as a "safe default"
• Caching layers — prompt caches keyed by (model, messages) tuples where the model name is part of the cache key
• Observability dashboards — usage tracking, cost attribution, and alert thresholds that filter by model name; post-cutoff these filters silently return zero results
• Load test and eval harnesses — any benchmark or regression test that calls the deprecated names will break after July 24, masking real eval failures as fixture errors

The Router/Operator Angle: Three-Layer Audit Before Thursday

An AI gateway that abstracts model selection from the application layer has a structural advantage here — there is typically one canonical place where model identifiers are configured. But that advantage is only realized if the audit is thorough. Three layers to check:

Layer 1 — Provider configuration. Search your gateway config for the literal strings deepseek-chat and deepseek-reasoner. In most routers this is the YAML, TOML, or JSON provider spec. Replace both with deepseek-v4-flash (or deepseek-v4-pro for workloads that justify the cost delta). Confirm the base_url is unchanged: https://api.deepseek.com.

Layer 2 — Fallback and routing policies. Check every fallback definition. If deepseek-reasoner was configured as a lower-cost fallback for reasoning workloads, verify that the replacement is deepseek-v4-flash with "thinking": {"type": "enabled"} and the appropriate reasoning_effort value ("low", "medium", or "high"). The thinking-mode behavior is now a parameter, not a model alias.

Layer 3 — Downstream clients. Enumerate SDK clients and agents that call DeepSeek endpoints directly, bypassing the gateway. This includes coding agents like Claude Code, Cursor, or OpenCode if they were configured to hit DeepSeek directly with a hardcoded model name. Check any .env, settings.json, or provider config file that each agent reads. The DeepSeek agent integrations guide confirms that tool-level configs must use the new model identifiers.

What TheRouter Users Should Watch or Try

If your TheRouter routing profile routes any workload to DeepSeek, this week is the time to:

1. Open your provider config and confirm deepseek-v4-flash or deepseek-v4-pro are the specified model IDs, not the legacy aliases.
2. Check your fallback chain — if deepseek-reasoner was a fallback for reasoning tasks, replace it with deepseek-v4-flash and add reasoning_effort: high to the request parameters.
3. After updating, validate with a test request against the DeepSeek models listing to confirm the new identifiers are recognized.
4. Set a calendar alert for July 23 as a final pre-cutoff verification window.

The July 24 cutoff is not the kind of deadline that sends a warning email when it passes. Requests will start failing silently for any operator who does not complete this audit before 15:59 UTC on Thursday.