DeepSeek V4-Flash Just Outscored Its Own Flagship on Nine Benchmarks

DeepSeek moved its V4-Flash API out of preview and into public beta on July 31, 2026. The build carries the internal tag V4-Flash-0731, and per DeepSeek’s own agent benchmarks, it beats the company’s higher-tier V4-Pro-Preview on all nine published coding and agent tests — including Terminal-Bench 2.1 and DeepSeek’s internal DeepSWE suite.

Nothing about the model’s size changed. DeepSeek confirms the architecture is identical to the April preview: 284 billion total parameters, 13 billion active per token, and a 1-million-token context window. The gains came from a fresh post-training pass, not a bigger model.

What Actually Shipped

Calling the API still works the same way developers already know. Set the model field to deepseek-v4-flash and the endpoint now serves the 0731 build automatically — no new base URL, no migration step. DeepSeek’s changelog describes the update as touching only the Flash API; V4-Pro’s API, the DeepSeek chat app, and the web client stay on their previous builds for now.

Two developer-facing features arrived alongside the benchmark jump. The model now natively supports OpenAI’s Responses API format, and DeepSeek says it tuned the build specifically for Codex compatibility. Both moves target teams already running agent pipelines built around OpenAI’s tooling, lowering the switching cost for anyone testing DeepSeek as a cheaper backend.

The Benchmark Story

DeepSeek’s headline claim is specific: V4-Flash-0731 beats V4-Pro-Preview, the company’s own flagship, across every agent and coding benchmark it published for this release. That is not a marginal edge on one test — it is a smaller, cheaper model outscoring the larger one it was supposed to sit beneath.

DeepSeek attributes the jump entirely to re-post-training: applying a new training cycle to a model whose weights and architecture never changed. The pre-training phase fixed what the model knows. This second pass changed how it acts on that knowledge inside agent loops — tool calls, multi-step coding tasks, and the kind of extended reasoning chains that agent benchmarks are built to test.

Why This Matters for Pricing

OpenRouter lists V4-Flash at $0.0896 per million input tokens and $0.1792 per million output tokens, positioning it well below V4-Pro on cost. But cheaper per-token pricing doesn’t automatically mean cheaper per-task pricing. Agent workloads that trigger longer reasoning chains can push effective cost per completed task higher than the sticker price suggests.

Teams migrating agent pipelines from V4-Pro-Preview should benchmark cost per finished task, not cost per token, before locking in infrastructure decisions. The token price looks like a downgrade in cost. The task price might not be.

The Open-Weights Gap

Here’s the detail that changes who actually benefits from this release: 0731 currently exists only behind DeepSeek’s API. For the three months since the April preview, anyone running the open-weights version of V4-Flash had the same model as API customers. As of July 31, that’s no longer true — the API customer is running a newer generation than the person who downloaded the weights.

DeepSeek has a track record of eventually open-sourcing preview builds, and the company hinted that an official V4-Pro release is coming soon. But nothing published so far promises open weights for the 0731 build specifically. Anyone building on the self-hosted version should treat the timeline as unconfirmed rather than assumed.

The broader lesson extends past DeepSeek. Any model served through a hosted API can change on the vendor’s schedule, not the developer’s. DeepSeek did this one transparently — announced, versioned, dated, and benchmarked. Not every vendor discloses a swap this clearly, which makes the open-weights download still the only version a developer fully controls. That tension has shaped the wider competition among Chinese open-weight models gaining share on OpenRouter, and it echoes the hardware independence story behind DeepSeek’s Huawei Ascend deployment earlier this month.

What Developers Should Check First

Before shifting production traffic to V4-Flash-0731, a few things are worth confirming directly rather than assuming from the announcement. Check whether the third-party providers hosting deepseek-v4-flash have actually rolled out the 0731 weights — as of launch day, only DeepSeek’s own endpoint served the new build. Run your own agent workload against both V4-Flash and V4-Pro-Preview before switching, since DeepSeek’s benchmarks reflect its own test suite, not necessarily your specific pipeline. And confirm Responses API and Codex compatibility against your existing integration code, since native support doesn’t guarantee zero-change migration.

This release lands in the same week that Moonshot AI’s Kimi K3 reshuffled expectations for what a single open-weight release can do to the coding-model leaderboard. DeepSeek’s move keeps that competitive pressure aimed squarely at agent and coding workloads — the same segment enterprise buyers are watching most closely right now.

Frequently Asked Questions

What is DeepSeek V4-Flash-0731?

It’s the official public beta build of DeepSeek’s V4-Flash API, released July 31, 2026. It keeps the same 284-billion-parameter architecture as the April preview but ships with new post-training that DeepSeek says improves agent and coding benchmark scores past its own V4-Pro-Preview.

Do I need to change my API code to use it?

No. The model name stays deepseek-v4-flash and the base URL is unchanged. Existing API calls automatically route to the 0731 build.

Is V4-Flash-0731 available as open weights?

Not yet, as of this release. DeepSeek has only confirmed the API build. The company has previously open-sourced preview models on its own timeline, but no open-weights date has been published for 0731 specifically.

Is V4-Flash cheaper than V4-Pro?

Per-token pricing is lower — OpenRouter lists $0.0896 per million input tokens and $0.1792 per million output tokens for V4-Flash. But effective cost per completed agent task can run higher for workloads that trigger long reasoning chains, so it’s worth testing against your actual use case before assuming savings.