DeepSeek’s V4 Migration Is Complete. Now the Legacy Endpoints Are Gone.

# DeepSeek’s V4 Migration Is Complete. Now the Legacy Endpoints Are Gone.

The launch happened in April. The deadline was yesterday.

At 15:59 UTC on July 24, 2026, DeepSeek permanently retired the legacy `deepseek-chat` and `deepseek-reasoner` model identifiers, completing its API transition to the DeepSeek-V4 series.[[1]](api-docs.deepseek.com↗ [[2]](api-docs.deepseek.com↗

That timing matters because this is not the launch of a brand-new model family. DeepSeek-V4 Preview arrived on April 24, introducing V4-Flash and V4-Pro.[[3]](api-docs.deepseek.com↗ [[1]](api-docs.deepseek.com↗ The news of the past 24 hours is operational: compatibility aliases that gave developers time to migrate have now reached the end of the road.

Before retirement, calls to `deepseek-chat` were mapped to V4-Flash in non-thinking mode, while `deepseek-reasoner` calls were mapped to V4-Flash in thinking mode.[[4]](api-docs.deepseek.com↗ [[1]](api-docs.deepseek.com↗ That bridge reduced immediate breakage, but it also made procrastination easy. After the cutoff, callers must explicitly select a V4 model and configure the intended behavior.

The headline is API migration, not a new weight release. DeepSeek previewed V4 three months ago; July 24 was the moment its old endpoint vocabulary stopped being a supported abstraction.

The practical first stop is DeepSeek’s official API updates page↗, followed by the main developer documentation↗.[[3]](api-docs.deepseek.com↗ [[1]](api-docs.deepseek.com↗ Access is through a DeepSeek account and the API at `api.deepseek.com`.↗ The interface supports both OpenAI-compatible and Anthropic-compatible behavior, so many integrations can preserve their existing client structure even while changing model selection.[[4]](api-docs.deepseek.com↗ [[5]](api-docs.deepseek.com↗ There is no basis here to promise a free tier or universal regional availability; teams should verify access directly in their accounts.

For developers, the immediate changes are straightforward but consequential:

• Replace `deepseek-chat` or `deepseek-reasoner` with an explicit V4 model identifier, normally `deepseek-v4-flash` or `deepseek-v4-pro`.
• Reproduce the old behavior deliberately: non-thinking for applications formerly using `deepseek-chat`, thinking for those that relied on `deepseek-reasoner`.
• Keep the API base at `api.deepseek.com`,↗ unless your own gateway or provider configuration requires something different.[[4]](api-docs.deepseek.com↗ [[5]](api-docs.deepseek.com↗
• Revalidate structured output, tool execution and latency rather than assuming compatibility means identical runtime behavior.

This is the less glamorous side of rapid model progress. A model upgrade can be exciting; an alias retirement appears in incident channels, evaluation dashboards and procurement spreadsheets. Any team that treated a generic-looking model name as a permanent contract has just received a reminder that it was really a vendor-controlled routing layer.

What developers are actually migrating to

The V4 series offers two sharply differentiated capacity and cost profiles. V4-Flash has 284 billion total parameters, with 13 billion active, while V4-Pro has 1.6 trillion total parameters, with 49 billion active.[[1]](api-docs.deepseek.com↗ Because these are sparse architectures, “active” parameters identify the smaller portion used during a given operation rather than the model’s full parameter pool.

Both models provide:

• A 1 million-token context window.
• Thinking and non-thinking modes, allowing behavior to be chosen separately from the model family.
• DeepSeek Sparse Attention and token-wise compression, designed to make very long context more efficient.
• JSON output, tool calls and prefix completion for structured and agentic applications.
• API behavior compatible with established OpenAI and Anthropic request patterns.[[4]](api-docs.deepseek.com↗ [[1]](api-docs.deepseek.com↗ [[5]](api-docs.deepseek.com↗

The official chat-completion documentation↗ should therefore be treated as a contract to test, not merely a syntax page.[[4]](api-docs.deepseek.com↗ Compatibility can reduce code changes, but it does not remove the need to examine generated JSON, tool-call arguments, stop behavior, reasoning settings and application-specific quality.

The price table hides a caching strategy

DeepSeek’s official July rates, per 1 million tokens, are:[[4]](api-docs.deepseek.com↗

| Model | Input: cache hit | Input: cache miss | Output | |---|---:|---:|---:| | V4-Flash | $0.0028 | $0.14 | $0.28 | | V4-Pro | $0.003625 | $0.435 | $0.87 |

These figures make V4-Flash the natural first candidate for high-volume applications, while V4-Pro carries a higher price for teams seeking the flagship tier. But the larger strategic distinction is between cache hits and cache misses. Reused prompt material is billed far below fresh input under both models, making prompt stability and cache observability part of application economics—not just infrastructure housekeeping.

The current rates are published on DeepSeek’s official pricing page↗, which teams should monitor because pricing can change.[[4]](api-docs.deepseek.com↗ A 1M context window does not mean every request should use 1 million tokens. Long-context testing should measure how much input is fresh, how much is cached and how much output the workflow generates. A retrieval pipeline with stable instructions has a different cost profile from an agent that continually rewrites its working history.

For buyers, this complicates simplistic “price per million” comparisons. Procurement should ask about expected cache-hit rates, output length, traffic bursts and whether the application needs thinking mode. Developers, meanwhile, need separate evaluation tracks for Flash versus Pro and thinking versus non-thinking. A single blended benchmark can conceal the operational choice that the endpoint retirement now forces teams to make.

A crowded market, without apples-to-apples certainty

DeepSeek became globally prominent through a combination of performance and cost, reinforced by historical releases under the MIT License, which permitted commercialization, distillation and fine-tuning.[[6]](api-docs.deepseek.com↗ Its prior model catalog can be explored through DeepSeek’s Hugging Face organization↗.[[7]](huggingface.co↗

That history must not be misapplied to V4. The evidence here supports historical MIT-licensed releases; it does not establish that V4-Flash or V4-Pro has open weights or a particular license. Today’s development concerns a hosted API migration, not a V4 weight drop.

The commercial comparison is nevertheless becoming more international and more interesting. Chinese providers are not only competing on headline model capability; they are offering different combinations of context, sparse activation, caching and API pricing.

Three useful reference points

Zhipu AI’s GLM-5.2, released in June—reported as June 16 in the platform update—also offers a 1M context window.[[8]](docs.bigmodel.cn↗ Its listed price is ¥8 per million input tokens and ¥28 per million output tokens.[[9]](bigmodel.cn↗ Zhipu reports strong coding and long-range task performance, but the supplied evidence does not establish a license for GLM-5.2. Developers can consult the Z.ai documentation↗ for current platform details.[[10]](docs.z.ai↗

MiniMax-M3, released June 1, likewise supports 1M context and uses MiniMax Sparse Attention.[[11]](minimax.io↗ [[12]](minimax.io↗ [[13]](platform.minimax.io↗ Standard pay-as-you-go pricing is $0.30 per million input tokens for input at or below 512K, and $1.20 per million output tokens. Beyond 512K, the corresponding standard rates double to $0.60 and $2.40.[[14]](platform.minimax.io↗ MiniMax’s text-model quickstart↗ provides the practical access path.[[14]](platform.minimax.io↗

MiniMax reports that M3 scored 83.5 on BrowseComp, compared with 79.3 for Opus 4.7.[[11]](minimax.io↗ That is a vendor-reported benchmark, not an independently reproduced result in the supplied evidence. It is useful as a claim to test, not a settled league table.

StepFun’s Step 3.7 Flash offers a different balance: 198B total parameters, 11B active, and a 256K context window.[[15]](platform.stepfun.com↗ Its prices per million tokens are CNY 1.35 for cache-miss input, CNY 0.27 for cache-hit input and CNY 8.1 for output.[[15]](platform.stepfun.com↗ [[16]](platform.stepfun.com↗ The StepFun LLM guide↗ documents the platform.[[15]](platform.stepfun.com↗

Vendor benchmarks answer, “What result does the provider report under its setup?” Independent evaluations answer, “Can another party reproduce the advantage?” Those are related questions, but they are not interchangeable.

The fair comparison is therefore not “which model won?” It is which service fits a specific workload:

• DeepSeek V4-Flash emphasizes unusually low listed token prices, especially for cached input, alongside 1M context.
• DeepSeek V4-Pro offers the larger active and total parameter profile at higher—but still explicitly tiered—API pricing.
• GLM-5.2 provides another 1M-context option with renminbi-denominated input and output rates.
• MiniMax-M3 combines 1M context with pricing that rises after 512K, making prompt length a visible commercial threshold.
• Step 3.7 Flash has a shorter 256K window but provides a useful contrast in sparse activation and cache-aware pricing.

One genuinely recent adjacent update arrived on July 22: MiniMax reduced its API-vlm price to $0.01 per call.[[14]](platform.minimax.io↗ That is a related last-72-hours commercial move, but it should not distract from the main development. DeepSeek’s change can interrupt existing text applications because model identifiers have been retired; MiniMax’s update changes the price of a separate service.

The migration playbook

What to do Monday morning

Start with a controlled migration rather than a blind string replacement:

• Inventory every caller. Search application code, environment variables, orchestration tools, notebooks, scheduled jobs and internal gateways for `deepseek-chat` and `deepseek-reasoner`.
• Choose explicit targets. Map former chat traffic to V4-Flash non-thinking and former reasoner traffic to V4-Flash thinking as the closest documented continuity path; evaluate V4-Pro separately rather than silently upgrading everything.[[4]](api-docs.deepseek.com↗ [[1]](api-docs.deepseek.com↗
• Pin evaluation traffic. Send a stable test set to each intended model-and-mode combination, and record the exact model string, settings and date.
• Validate tools and JSON. Check schema adherence, argument names, malformed-call handling, prefix completion and downstream parser behavior.
• Measure long-context costs. Separate cache-hit input, cache-miss input and output tokens; do not extrapolate from short prompts to million-token workflows.
• Watch the provider documentation. Recheck the updates, pricing and API pages before deployment and after any unexplained behavioral or billing change.

Testing thinking modes deserves special care. A workflow previously aimed at `deepseek-reasoner` should not merely return a plausible final answer; it should meet the application’s quality, latency and cost requirements under the explicit V4 thinking configuration. Conversely, enabling thinking in a high-volume classification or extraction pipeline may add behavior the product never needed.

Tool use needs end-to-end validation. Confirm that generated calls match your schemas, that the application executes the correct function, and that tool results return to the model in the expected format. For JSON workloads, test invalid and adversarial inputs rather than checking only a happy path. The V4 documentation lists support for these features, but production reliability belongs to the complete system.[[4]](api-docs.deepseek.com↗ [[1]](api-docs.deepseek.com↗

Finally, retain a rollback route through your own routing layer, even though rollback to the retired names is no longer the answer. Dependency concentration is the broader lesson: compatible APIs make providers easier to adopt, but model aliases, pricing and lifecycle dates remain under provider control. A small internal abstraction around model selection can make the next retirement a configuration change instead of an emergency.

Chinese API releases are increasingly shaping the global developer market through concrete choices about long context, caching and sparse-model economics—not through a vague geographic horse race. The buyer’s challenge is to compare reproducible workload performance and total cost. The developer’s challenge is to keep integrations portable while still using provider-specific advantages.

The decisive takeaway: replace the legacy DeepSeek identifiers now, test V4 by model and mode, and treat endpoint lifecycle monitoring as part of production engineering—not release-note reading for later.