DeepSeek V4 Flash Goes Official: Agent Benchmarks That Beat V4 Pro Preview

DeepSeek just pushed the official version of its V4 Flash API into public beta — and the headline isn't a new model, it's a retrain. The build, tagged DeepSeek-V4-Flash-0731, keeps the exact same architecture, size, and price as the preview, but its agent benchmark scores now far exceed V4-Pro-Preview, the larger, pricier model in the same family. It also adds native Responses API support and is adapted for Codex. If you already call deepseek-v4-flash, the model behind your API just got dramatically better for free.

What Actually Changed

Per DeepSeek's official API changelog, the official V4 Flash API is now in public beta, and the calling method is unchanged — you set the model name to deepseek-v4-flash exactly as you did during the preview. (DeepSeek changelog)

Three facts define this release:

• It's a retrain, not a new model. DeepSeek states that V4-Flash-0731 keeps the same architecture and size as V4-Flash-Preview and was only re-post-trained. New weights on the same skeleton — which is why integration is a drop-in change. (DeepSeek changelog)
• The agent numbers are the story. DeepSeek describes "significantly enhanced agent capabilities," with benchmark results far exceeding V4-Pro-Preview. The upgrade targets autonomous execution and tool-calling in complex tasks. (DeepSeek changelog)
• This is API-only, for now. The upgrade applies only to the V4 Flash API. The V4 Pro API and the models on DeepSeek's app and website are unchanged, and DeepSeek says the official V4 Pro release will follow soon. (DeepSeek changelog)

The strategic signal is what stands out. For most of the V4 generation the hierarchy was simple: Pro is strong, Flash is cheap and weaker. This retrain flips that story for agentic workloads — the lightweight model now posts numbers above the preview flagship. (TechNode)

The Benchmark Numbers

DeepSeek published a full benchmark table with the 0731 build. These are the official self-reported scores from the changelog:

Source: DeepSeek API changelog. DSBench-FullStack and DSBench-Hard are DeepSeek's internal test sets — a full-stack development suite and a hard coding-agent suite, respectively.

Two caveats before you read too much into the "beats V4 Pro Preview" line:

• The comparison isn't perfectly apples-to-apples. V4-Flash-0731 scores 82.7 on Terminal Bench 2.1, while the V4-Pro-Preview figure cited in coverage (67.9) is on Terminal Bench 2.0 — different versions of the test set, so a direct comparison isn't fully fair. (36Kr)
• Code Agent tasks used a specific harness. DeepSeek notes the public Code Agent benchmarks were run using its "DeepSeek Harness minimal mode" (not yet released) at max effort, top-p 0.95, temperature 1.0. Your production numbers depend on your own scaffolding. (DeepSeek changelog)

Still, the magnitude is notable: a model with roughly 13B activated parameters landing agent scores in this range shows how much post-training alone can move the needle. (36Kr)

Why "Same Size, Better Agent" Matters

V4 Flash is a 284B-parameter MoE with ~13B active parameters, versus V4 Pro's 1.6T total / ~49B active. Both share a 1M-token context window. The whole point of Flash was throughput and cost — you traded some capability for speed. (For the full spec breakdown, see our DeepSeek V4 Pro guide.)

By pushing agent performance up without changing the parameter count, DeepSeek is targeting the exact pain point most teams hit in production: they want low-cost, high-concurrency automation but can't afford to run a 1.6T flagship on every agent step. A cheaper model that handles tool-calling and multi-step execution well changes the economics of running agents at scale. (BigGo Finance)

For anyone building agent pipelines, the practical takeaway is a routing one: more of your everyday agent steps can now default to Flash, with escalation to a heavier model reserved for the genuinely hard branches.

Responses API and Codex Support

The other half of this release lowers the integration barrier. The official V4 Flash natively supports the Responses API format and is specifically adapted for Codex. (DeepSeek changelog)

That matters because the Responses API is the format OpenAI's newer agentic tooling — including Codex — expects. Native support means teams already wired into a Codex-style workflow can point it at deepseek-v4-flash with configuration changes rather than a rewrite. DeepSeek says the goal is to lower the barrier for developer adoption, and the configuration details live in its API docs. (BigGo Finance)

Combined with the fact that the V4 API is already reachable through both OpenAI-style ChatCompletions and Anthropic-style interfaces, DeepSeek is clearly optimizing for drop-in compatibility across the major agent ecosystems. (DeepSeek changelog)

What This Means for Your Stack

If you build on DeepSeek, here's the short version:

• Already on the preview? You get the upgrade with no code changes — same model name, same price, better agent behavior. Re-run your evals to confirm the improvement shows up on your tasks, not just the published benchmarks.
• Running Codex-style workflows? V4 Flash is now a genuine candidate as a cheaper backend, thanks to native Responses API support and Codex adaptation.
• Waiting on V4 Pro? DeepSeek says the official Pro release is coming soon; this Flash retrain is a preview of the post-training gains you'll likely see there too.
• Treat vendor benchmarks as directional. The scores are self-reported, some sets are internal, and the Terminal Bench version mismatch means the "beats Pro Preview" framing needs an asterisk until independent evals land.

Do This With Your Own AI Workforce

A model getting cheaper and better is exactly why a model-agnostic agent platform pays off — you want to slot in the best model for each step without re-architecting. Eigent is an open-source, multi-agent "Cowork" desktop app that orchestrates specialized models across real workflows, so you can route routine agent steps to a fast, cheap model like V4 Flash and escalate the hard branches elsewhere. If you're wiring up agentic coding tasks, see how Eigent handles workflows like reviewing GitHub PRs, or download Eigent to build your own AI workforce locally.

Frequently Asked Questions

What is DeepSeek V4 Flash?

DeepSeek V4 Flash is the lightweight variant of DeepSeek's V4 model family — a 284B-parameter Mixture-of-Experts model with roughly 13B active parameters and a 1M-token context window, tuned for fast, low-cost, high-throughput workloads. The official API (build 0731) is now in public beta.

What changed in the official V4 Flash release?

The 0731 build keeps the same architecture, size, and price as the preview but was re-post-trained, delivering significantly stronger agent benchmark scores. It also adds native Responses API support and Codex adaptation. The calling method is unchanged — you still set the model name to deepseek-v4-flash.

Do V4 Flash's benchmarks really beat V4 Pro Preview?

DeepSeek's self-reported agent scores exceed V4-Pro-Preview on several benchmarks. But the comparison isn't fully apples-to-apples — for example, the Flash score is on Terminal Bench 2.1 while the cited Pro Preview score is on Terminal Bench 2.0. Treat the numbers as directional until independent evals confirm them.

Does V4 Flash work with Codex?

Yes. The official V4 Flash natively supports the Responses API format and is specifically adapted for Codex, so Codex-style workflows can use it as a backend with configuration changes rather than a rewrite.

Is V4 Pro getting the same upgrade?

Not yet. This update applies only to the V4 Flash API. The V4 Pro API and DeepSeek's app and web models are unchanged, though DeepSeek says the official V4 Pro release will follow soon.