Frontend · Backend · DeepSeek

DeepSeek V4-Flash Ships Post-Training Overhaul That Doubles Agent Benchmarks

By 勇宝趣学前端 · Jul 31, 2026

Read original on juejin.cn ↗ Google Translate ↗ Alt translation

A model that costs ¥2 per million output tokens and now scores competitively on agent benchmarks changes the economics of running autonomous coding loops at scale. Developers building CI-integrated repair bots, repo-analysis pipelines, or multi-step automation can get Pro-preview-beating performance at Flash pricing, with native tool-calling interfaces that skip the usual adapter work.

Summary

The V4-Flash-0731 release keeps the model architecture and parameter count unchanged but retrains the model to execute multi-step agent tasks more reliably. Terminal Bench 2.1 jumps from 61.8 to 82.7, Cybergym nearly doubles to 76.7, and DeepSWE climbs from 7.3 to 54.4 — all exceeding the earlier Pro preview. Native Responses API support and explicit Codex adaptation make the model a drop-in for programming assistants and automated workflows. Pricing stays at ¥0.02 per million cached input tokens, ¥1 uncached, and ¥2 output, undercutting the Pro tier by 3× on generation cost. The update is API-only; the web app, mobile app, and V4-Pro remain on the previous versions, so end users won't see a change yet.

Takeaways

— Model architecture and parameter count are identical to the preview; all gains come from post-training optimization.

— Terminal Bench 2.1 rose from 61.8 to 82.7, approaching Opus-4.8 in the same comparison table.

— Cybergym jumped from 38.7 to 76.7, and DeepSWE from 7.3 to 54.4.

— NL2Repo improved from 39.4 to 54.2, Toolathlon Verified from 49.7 to 70.3, and DSBench-Hard from 25.8 to 59.6.

— V4-Flash now natively supports the Responses API and has been specifically adapted for Codex.

— Pricing remains ¥0.02/M cached input tokens, ¥1/M uncached input, and ¥2/M output — roughly one-third the output cost of V4-Pro.

— The update applies only to the V4-Flash API; the web app, mobile app, and V4-Pro API are unchanged.

— Official V4-Pro release is still pending with no fixed date beyond 'as soon as possible.'

Conclusions

Doubling Cybergym and multiplying DeepSWE by 7× through post-training alone suggests the base model had latent agent capability that earlier fine-tuning failed to surface.

Shipping a Flash-tier model that beats the previous Pro preview on agent benchmarks while keeping the same architecture is an unusual pricing-capability inversion — the cheaper model now leads on the metrics that matter for automation.

Explicit Codex adaptation and Responses API support signal that DeepSeek is targeting integration into existing agent frameworks rather than trying to own the orchestration layer itself.

The API-only rollout treats the developer channel as the primary surface for agent models, while consumer-facing apps lag behind — a pattern that mirrors how GPT-4's API led its ChatGPT deployment by months.

Concepts & terms

Responses API

An API format designed for multi-turn agent interactions that handles tool calls, state management, and structured output natively, as opposed to simple chat-completion endpoints that return only text.

Post-training optimization

Additional training applied after a model's initial pre-training and fine-tuning, often using reinforcement learning or curated task data to improve specific behaviors — here, multi-step code execution and tool use — without changing the underlying architecture or parameter count.

Terminal Bench / Cybergym / DeepSWE

Agent evaluation benchmarks that measure a model's ability to use a terminal, navigate cybersecurity tasks, and solve real-world software engineering problems respectively, as opposed to static code-generation accuracy.

Source: juejin.cn ↗ Google Translate ↗ Backup ↗