DeepSeek V4-Flash GA: Agent Benchmarks, Responses API, and What Didn't Change

DeepSeek's V4-Flash 2026-07-31 GA: big agent-benchmark gains, native Responses API plus Codex support, and the same architecture as preview. How to read the numbers and what stays on V4-Pro.

On 2026-07-31, DeepSeek moved V4-Flash to its GA API endpoint (announced as a public beta). The release note leads with agent benchmarks that sit well above the V4-Pro-Preview numbers we tracked earlier, plus two integration changes that matter more than the scores: native Responses API support and a Codex-specific adaptation.

The part that should stop you from overreacting: V4-Flash-0731 is the same model size and structure as V4-Flash-preview, only re-post-trained. This is not a new architecture; it is a tuned checkpoint. The other fine print is that only the V4-Flash API changed — V4-Pro API and the APP/WEB models are untouched, and DeepSeek says a V4-Pro GA "will follow soon."

What the announcement actually changed

AreaBefore (preview)GA (2026-07-31)
Model IDdeepseek-v4-flashSame deepseek-v4-flash; deepseek-chat / deepseek-reasoner retired 2026-07-24
ArchitectureV4-Flash-previewSame size and structure, re-trained only
Responses APINot mentioned in GA noteNative support
CodexGeneric tool callingAdapted for Codex
V4-Pro API / APP / WEBCurrentUnchanged

One detail to check on your end: the old aliases deepseek-chat and deepseek-reasoner were scheduled to stop three months after the April V4 launch, around 2026-07-24. If your CI still points at either name, this GA note is your reminder to migrate to the explicit model ID before retries start hitting an error.

The agent benchmark table (read it in two halves)

DeepSeek reports nine scores, and they split into public and internal sets:

BenchmarkScoreSource
Terminal Bench 2.182.7Public
NL2Repo54.2Public
Cybergym76.7Public
DeepSWE54.4Public
Toolathlon verified70.3Public
Agent Last Exam25.2Public
Automation Bench (Public)25.1Public
DSBench-FullStack68.7Internal
DSBench-Hard59.6Internal

DevCove view: The public tasks still ran in the DeepSeek Harness minimal mode (not yet released) at max tier, topp=0.95, temperature=1.0. A vendor harness that ships after the score is a warning sign for procurement: match the harness, not the number.

The first seven are reproducible on public sets, so independent verification becomes possible once the harness ships. The last two — DSBench-FullStack (full-stack development) and DSBench-Hard (hard coding-agent problems) — are internal and cannot be checked by outsiders. Treat them as directional, not evidence.

What we will not do with these numbers

  • We will not call V4-Flash "the best agent model" from one vendor table; the same rule as Best AI coding models for real projects applies — snapshot date, harness, and task mix first.
  • We will not swap a model just because the headline says "far above V4-Pro-Preview." Preview versus GA is an unfair baseline; compare against what you actually run today.
  • We will not assume the Codex adaptation is zero-config. DeepSeek's note points to a dedicated Codex setup doc; that is a real configuration step, and it stays outside our ship checklist until we run it ourselves.

Quick way to confirm the GA endpoint

Point a small request at your existing base URL with the explicit model ID (use a placeholder key, never a real one):

curl -s https://api.deepseek.com/chat/completions \
  -H "Authorization: Bearer $DEEPSEEK_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model":"deepseek-v4-flash","messages":[{"role":"user","content":"ping"}]}'

If the old deepseek-chat alias now errors while deepseek-v4-flash responds, you are on the GA endpoint. base_url in your SDK config stays unchanged — DeepSeek kept it stable.

Bottom line

The Responses API + Codex support and the post-training-only note are the parts to act on; the benchmark spread is context until a released harness lets you reproduce it. If your team already runs DeepSeek via deepseek-v4-flash, check the Codex doc and move off any retired aliases. If you are on V4-Pro, nothing changed for you yet — wait for the GA note rather than planning a migration today.

Related: AI coding model rankings · Claude Opus 5 tiers · Kimi K3 for coding agents · Verification gates for AI coding agents

Source: DeepSeek API update log, 2026-07-31 (api-docs.deepseek.com/zh-cn/updates); verify against the live page before production.

In this topic

Related articles

Complete guideHow to Choose AI Coding Models for Real ProjectsUse public model leaderboards without confusing benchmark scores for permanent coding quality. Learn what Artificial Analysis metrics mean for cost, latency, context, and real delivery.Claude Opus 5 for Coding: Tiers, Latency, and When to DownshiftOpus 5 is a tier ladder, not one model. Use DevCove's 2026-07-25 snapshot to match max, high, or medium to your repo—and stop paying max prices for autocomplete work.Kimi K3 for Coding Agents: Harness, Context, and VerificationK3 in production is a harness problem first: permissions, what you inject into 1M context, and gates before merge—not another benchmark recap.AI Coding Agents Need Verification Gates, Not Just Better PromptsAI coding agents can write larger patches than autocomplete tools, but production teams need tests, review, sandboxing, and release gates around their output.

Related tools

Use the tools from this article

AI Coding Ship ChecklistAI coding checklist / AI app launch checklist / vibe coding checklist

Learn the format

AI Literacy for DevelopersPractical AI basics for developer workflows: models, prompts, coding assistants, verification, privacy, and reliable AI-assisted work.

Back to articles