DeepSeek Retires deepseek-chat and deepseek-reasoner on July 24: Your V4 Migration Guide

Last updated: September 26, 2026

If you have deepseek-chat or deepseek-reasoner anywhere in your codebase, you have a hard deadline. DeepSeek's V4 announcement states that both legacy model names "will be fully retired and inaccessible after Jul 24th, 2026, 15:59 (UTC Time)." The official pricing page confirms the same cutoff: deprecation lands at 2026/07/24 15:59 UTC.

The good news: this is one of the gentler migrations in recent memory. The base URL does not change, both legacy names already route to the new model, and the replacement is cheaper per token than almost anything else on the market. The catch is that thinking mode moved from a model name to a request parameter, and a few legacy deepseek-reasoner quirks need cleanup on the way through. This guide is the what-maps-to-what checklist. (If you migrated in July, one more rename landed in September - deepseek-v4-flash became deepseek-flash, below.)

What Changed on September 26#

The target model changed names and got cheaper (September 10). DeepSeek shipped DeepSeek-V4.1-Flash, the first model in a new architecture family: a 552B-parameter MoE with asymmetric causal encoder-decoder design (8B active for input, 16B for output), native visual understanding, and a quarter of the previous KV cache footprint (release news, verified September 26, 2026). It is served as deepseek-flash, and the old deepseek-v4-flash and deepseek-v4-flash-vision-exp names are retired as canonical models - requests still work because they temporarily route to V4.1-Flash, but the documented name going forward is deepseek-flash. V4-Pro is being phased out: since September 14, 2026 at 04:00 UTC, all deepseek-v4-pro requests route to V4.1-Flash and bill at V4.1-Flash rates until V4.1-Pro launches.

The repricing also cut rates on the Flash tier, on top of the peak/off-peak policy from August 16. (All numbers verified September 26, 2026 on the pricing page; off-peak is everything outside 01:00-04:00 and 06:00-10:00 UTC, peak is 2x):

If you built your migration in July around deepseek-v4-flash at $0.14/$0.28, two things changed: the name you should pin is deepseek-flash, and the rate sheet you should budget against is the table above. The frontier pricing tracker has the full September landscape.

What Changed on July 25#

The deprecation deadline has passed. As of July 24, 2026 at 15:59 UTC, the legacy deepseek-chat and deepseek-reasoner model names stopped resolving. Requests using those names now return errors rather than routing to V4-Flash. (The pricing in this section is the July snapshot: V4-Flash at $0.14/$0.28 per MTok. September replaced that model with deepseek-flash at $0.15/$0.60 off-peak - see above.)

If your code was still using the legacy names after the cutoff, every API call is failing. The fix is a one-line model name swap in each request. No base URL change, no API key rotation, no endpoint migration.

What Actually Happens on July 24#

Since the V4 preview shipped on April 24, 2026, the legacy names have been aliases. Per the official announcement, deepseek-chat and deepseek-reasoner are "currently routing to deepseek-v4-flash non-thinking/thinking" respectively. So if your app still sends the old names, you have already been on V4-Flash for weeks - the July 24 date is when the aliases stop resolving entirely.

That means two separate facts worth internalizing:

• Your model already changed. If you noticed output drift since late April, this routing is why. Any evals you ran against the old V3-era models no longer describe what you are serving.
• Your code will hard-break on July 24. After 15:59 UTC, requests using the legacy names become errors, not redirects.

DeepSeek's own migration advice is one line: "Keep base_url, just update model to deepseek-v4-pro or deepseek-v4-flash." The endpoint stays https://api.deepseek.com, and the API remains compatible with the OpenAI ChatCompletions format, with an Anthropic-compatible surface as a second option (more on that below).

What Maps to What#

The mapping is mechanical for Flash, with one structural change: thinking is no longer a separate model name.

Per the API quick start, thinking mode on V4 models is controlled by a request parameter, not a model suffix. With the OpenAI Python SDK that looks like extra_body={"thinking": {"type": "enabled"}}; in the Node.js SDK, thinking passes as a direct parameter. The docs also show a reasoning_effort parameter (for example "high") for tuning how hard the model thinks.

Note what is new here: V4-Pro had no legacy alias at all. The retirement only forces you onto V4-Flash equivalents. Moving up to Pro is a deliberate choice, which is what the next section is for.

V4.1-Flash vs V4-Pro: Head-to-Head#

All pricing below is from the official pricing page, verified September 26, 2026 (peak = 2x off-peak; off-peak is everything outside 01:00-04:00 and 06:00-10:00 UTC). The V4.1-Flash architecture and benchmark claims are from the September 10 release news, which reports third-party tests placing V4.1-Flash ahead of V4-Pro on performance, cost, and speed - the Artificial Analysis figures in the table below describe the retired V4-Flash, its predecessor. Open-weights info is from the DeepSeek-V4.1-Flash and DeepSeek-V4-Pro Hugging Face repositories.

Both models support thinking and non-thinking modes, JSON output, tool calls, and chat prefix completion (beta). FIM completion works in non-thinking mode only. V4.1-Flash also adds native multimodal (visual) understanding - a first for the Flash tier - while V4-Pro does not list vision support.

One number worth dwelling on: the 1M-token context window is standard on both models, which Artificial Analysis notes is an 8x expansion from V3.2's 128K. If you built chunking or context-compression layers to live inside the old window, the migration is a chance to delete code.

A pricing caveat, because honesty beats tidiness: rates moved twice since this guide launched in June - down, both times. July's launch-window coverage listed V4-Pro at $1.74 input / $3.48 output per million tokens; the standing rates after August 13's GA priced Pro at $0.435 / $0.87. September's repricing cut the Flash tier and folded V4-Pro requests into V4.1-Flash billing (since September 14, 04:00 UTC): deepseek-v4-pro requests are served by V4.1-Flash at V4.1-Flash rates, so the $0.66 / $1.98 off-peak column above is the ceiling, not the default bill. Treat the official page as canonical and re-check it before you commit a budget.

The Migration Checklist#

Work through these in order. For most codebases this is an afternoon, not a sprint.

1. Find every legacy reference#

Grep for both names across code, config, env files, and infrastructure-as-code. The sneaky ones live in evaluation harnesses, fallback chains, and notebooks.

2. Swap the model names#

deepseek-chat becomes deepseek-flash. deepseek-reasoner becomes deepseek-flash with thinking: {"type": "enabled"} in the request body. Base URL stays https://api.deepseek.com. (If you migrated in July to deepseek-v4-flash, that name still resolves today - it temporarily routes to V4.1-Flash - but pin deepseek-flash going forward, since the older name is retired and will stop being documented as a target.)

3. Keep stripping reasoning content in multi-turn flows#

The reasoning model guide documents that the chain of thought comes back in a separate reasoning_content field and must be removed from assistant messages before you send the conversation back - including it triggers errors. If your deepseek-reasoner integration already handles this, keep that code. If you are adding thinking mode for the first time, build this in from day one.

4. Revisit max_tokens#

Legacy deepseek-reasoner defaulted to 32K output with a 64K maximum, including the chain of thought, per the reasoning model guide. V4 models support up to 384K output tokens per the pricing page. If you hardcoded conservative limits to dodge truncation, you have new headroom - but remember output is the expensive direction, so raise limits deliberately.

5. Re-test tool calls in thinking mode#

Legacy deepseek-reasoner did not support function calling at all. On V4, the pricing page lists tool call support on both models. If you previously split traffic between a reasoning model and a tool-calling model to work around this, you can likely collapse that into a single V4-Flash call with thinking enabled. That is a genuine architectural simplification, so re-run your agent evals to confirm.

6. Audit your cache hit rate#

The cache-hit discount on V4.1-Flash is the strongest on any pricing page in this guide: $0.003 versus $0.15 per million input tokens off-peak, verified September 26, 2026 on the pricing page. Stable system prompts and shared context prefixes pay for themselves at this ratio. We dug into cache-first agent design in DeepSeek V4 for budget coding agents if you want patterns.

7. Check your concurrency assumptions#

V4.1-Flash allows 2,500 concurrent requests; V4-Pro allows 500. If you are migrating a high-throughput pipeline and considering the Pro upgrade at the same time, the 5x concurrency difference may matter more than the per-token price difference - and note that deepseek-v4-pro currently serves V4.1-Flash at V4.1-Flash rates, so the Pro upgrade buys concurrency limits, not a different model, until V4.1-Pro lands.

8. Decide whether the Anthropic format helps you#

DeepSeek runs a second API surface at https://api.deepseek.com/anthropic that speaks the Anthropic Messages format, per the Anthropic API guide. You point the Anthropic SDK at it with ANTHROPIC_BASE_URL and your DeepSeek key in ANTHROPIC_API_KEY. Model name mapping is automatic: claude-opus variants map to deepseek-v4-pro (served by V4.1-Flash since September 14), claude-sonnet and claude-haiku variants map to the Flash tier, and anything unmapped defaults to Flash. Streaming, temperature, stop sequences, top_p, tool use, web search result content, and thinking mode are supported, though budget_tokens is ignored. Images, documents, code execution results, and MCP content types are not supported through this bridge.

Decision Guide: Which Path for Which Team#

High-volume, cost-sensitive API users. Move to deepseek-flash non-thinking and stop there. It is the literal successor to deepseek-chat, the behavior change already happened back in April, and at $0.15/$0.60 per million tokens off-peak (verified September 26, 2026) it still sits at the open-weights budget floor. Our frontier pricing notes cover why open-weights economics keep this floor low.

Reasoning-heavy pipelines. Migrate to deepseek-flash with thinking enabled first and measure. The Pro step-up is currently in flux: DeepSeek's own release notes report third-party tests placing V4.1-Flash ahead of V4-Pro on performance, cost, and speed, and deepseek-v4-pro has routed to V4.1-Flash at V4.1-Flash rates since September 14 pending V4.1-Pro. Budget at Flash rates, re-run quality gates after the V4.1-Pro launch, and see the full V4.1 developer guide for where quality gaps show up in practice.

Teams standardized on the Anthropic SDK or Claude Code. The Anthropic-compatible endpoint means migration can be a two-environment-variable change. The automatic claude-to-deepseek model mapping is convenient, but pin explicit deepseek-flash / deepseek-v4-pro names in production so a mapping change upstream cannot silently swap your model. If you are weighing DeepSeek against staying on Claude entirely, see our Fable 5 vs DeepSeek V4 cost-quality comparison.

Self-hosters. Both models are MIT-licensed with weights on Hugging Face (V4.1-Flash is the new one; the official announcement rosters the older family). The API retirement does not touch self-hosted deployments. Flash-class weights at ~160GB are the realistic target for most setups; Pro's 865GB download is datacenter territory, and V4.1-Flash's quarter-size KV cache improves the self-host cost picture further.

When to Do Nothing, and When to Leave#

There is no "stay on the legacy endpoints" option here - July 24 was a hard wall. But there are two honest non-migration paths.

Doing nothing was mostly fine, with one asterisk. Because the aliases already route to the Flash model, swapping names changed nothing about behavior. The asterisk: re-run evals, because the model behind your existing calls already changed - in April to V4-Flash, then again in September to V4.1-Flash.

Leaving DeepSeek entirely makes sense in a few specific cases. The old V4 was text-in, text-out only, per Artificial Analysis - if your roadmap needs vision or document understanding in the same call, this migration window was the natural moment to consolidate on a multimodal provider. That caveat shrinks with V4.1-Flash, which adds native visual understanding to the Flash tier; it still applies to FIM completion inside thinking-mode flows (FIM is non-thinking only) and to data-governance requirements that rule out the hosted API.

If you are coming from even older code - R1-era patterns and V3 assumptions - our earlier DeepSeek R1 and V3 guide is a useful snapshot of how much the platform has shifted under the same base URL.

FAQ#

Is deepseek-chat deprecated?#

Yes. The legacy model names deepseek-chat and deepseek-reasoner were fully retired on July 24, 2026 at 15:59 UTC. Requests using those names now return errors. The replacement is deepseek-flash (with thinking mode where you used reasoner).

What replaces deepseek-reasoner after the July 24 deprecation?#

DeepSeek-V4.1-Flash with thinking mode enabled. Instead of selecting a separate reasoning model, you send model: "deepseek-flash" with thinking: {"type": "enabled"} in the request body. A reasoning_effort parameter tunes thinking depth, and the Pro tier offers the same dual modes at a higher-quality tier (currently served by V4.1-Flash until V4.1-Pro launches).

Will my API calls break now that the deadline passed?#

If your code still uses the legacy model names, yes - requests fail with errors. The base URL, API key, and OpenAI-compatible request format all stay the same, so the fix is updating the model name string and, for reasoner traffic, adding the thinking parameter. This is a one-line change per request.

How much does DeepSeek V4 cost compared to other frontier APIs?#

Verified September 26, 2026 on the official pricing page, with peak/off-peak rates in effect since August 16: V4.1-Flash is $0.15 input / $0.60 output per million tokens off-peak ($0.003 on cache hits), doubling to $0.30 / $1.20 during peak hours (01:00-04:00 and 06:00-10:00 UTC). V4-Pro lists at $0.66 / $1.98 off-peak but has routed to V4.1-Flash at V4.1-Flash rates since September 14. Both include a 1M-token context window and up to 384K output tokens. For context, GPT-6 Luna now undercuts even this floor at $0.10/$0.50 - see the frontier pricing tracker.

Can I use DeepSeek V4 with the Anthropic SDK?#

Yes. DeepSeek serves an Anthropic-compatible API at https://api.deepseek.com/anthropic. Set ANTHROPIC_BASE_URL to that endpoint and put your DeepSeek key in ANTHROPIC_API_KEY. Claude model names map automatically (opus-class to Pro, sonnet and haiku-class to Flash), though images, documents, and MCP content types are not supported through the bridge.

Official Sources#

Continue Reading#

• DeepSeek 4.1 Flash as a Daily Coding Model - cache math, peak-hour windows and what the HN thread says about the cost claims
• Frontier Model API Pricing (September 2026) - live pricing snapshot across all major providers
• GPT-6 Sol vs Claude Opus 5.5 vs Grok 4.7: The September 2026 Price War - where DeepSeek's floor sits against the new closed-provider rates
• DeepSeek V4 Developer Guide - deep dive into V4 capabilities and API usage
• DeepSeek V4 for Budget Coding Agents - cost-optimized patterns for agentic workloads
• Fable 5 vs DeepSeek V4 Cost-Quality Comparison - head-to-head benchmark analysis
• CC Switch Guide: Manage Claude Code and Codex Providers From One App (2026) - switch Claude Code between providers without hand-editing env vars