DeepSeek-V4-Flash Is Now Official: Agent Benchmarks Beat V4-Pro-Preview, and Your Code Does Not Change
By the Deep Seek AI editorial desk · August 2, 2026 · 7 min read
DeepSeek pushed the official deepseek-v4-flash API into public beta on July 31, 2026. Same model string, same endpoint, same architecture — but a fresh post-training pass that lifts agent scores clean past V4-Pro-Preview. Here is what actually changed, what did not, and what it means for the cheap tier of the DeepSeek API.
The 60-second version
- The official (正式版) DeepSeek-V4-Flash API is live in public beta. Keep calling
deepseek-v4-flash— you get the newest version automatically. - Agent capability is the headline. DeepSeek reports the official V4-Flash substantially exceeding V4-Pro-Preview across nine agent benchmarks.
- Same model, new post-training. V4-Flash-0731 keeps the exact structure and size of V4-Flash-Preview; only post-training was redone.
- Native Responses API + Codex. The official Flash speaks the Responses API format natively and is specifically adapted for Codex.
- V4-Pro and the app are untouched. Only the Flash API was upgraded. An official V4-Pro release is promised "as soon as possible".
The agent benchmarks DeepSeek published
All figures below are DeepSeek's own reported numbers from the July 31, 2026 changelog entry. For public-benchmark code-agent tasks, DeepSeek says the official V4-Flash was evaluated using DeepSeek Harness minimal mode (announced as forthcoming) at the max tier, with top_p = 0.95 and temperature = 1.0. Those settings matter: agent scores are extremely sensitive to the harness, so treat them as vendor-reported until independently reproduced.
| Benchmark | V4-Flash (official) | What it measures |
|---|---|---|
| Terminal Bench 2.1 | 82.7 | Long-horizon terminal/shell agent tasks |
| NL2Repo | 54.2 | Building a working repository from a natural-language spec |
| Cybergym | 76.7 | Security/CTF-style agentic reasoning |
| DeepSWE | 54.4 | Real-world software-engineering issue resolution |
| Toolathlon (verified) | 70.3 | Multi-tool orchestration under verification |
| Agent Last Exam | 25.2 | Hard, open-ended agent reasoning |
| Automation Bench (Public) | 25.1 | End-to-end workflow automation |
| DSBench-FullStack | 68.7 | Internal full-stack development suite |
| DSBench-Hard | 59.6 | Internal hard coding-agent suite |
Two of those suites are internal: DSBench-FullStack is DeepSeek's in-house full-stack development set and DSBench-Hard its hard coding-agent set. They are not third-party comparable — useful as a signal of intent, not as a leaderboard.
Why "same structure, new post-training" is the most important line
DeepSeek states plainly that DeepSeek-V4-Flash-0731 has the same model structure and size as DeepSeek-V4-Flash-Preview, and that only the post-training was rerun. That has three practical consequences:
- Latency and cost profiles should be stable. No parameter-count change means no reason to expect a different throughput or price tier, and none was announced.
- Your prompts may still shift behaviour. Post-training is exactly what changes tool-calling style, refusal behaviour and verbosity. If you have prompt-brittle agent scaffolding, re-run your evals this week.
- Self-hosters are unaffected for now. This was an API-side upgrade; the changelog does not announce new open weights.
Native Responses API and Codex adaptation
The official V4-Flash natively supports the Responses API format and has been specifically adapted for Codex. Practically, that means you can point a Codex-style agent harness at DeepSeek without a translation shim between chat-completions and Responses-style items. DeepSeek documents the configuration in its agent integrations guide.
Combined with the agent benchmarks, the positioning is hard to miss: DeepSeek wants Flash — the cheap tier — to be the default model behind coding agents, not the fallback you use when Pro is too expensive.
What this costs you
No pricing change accompanied the July 31 update. Verified rates per 1,000,000 tokens in USD (last verified 2026-07-25):
| Model | Input (cache miss) | Input (cache hit) | Output |
|---|---|---|---|
| DeepSeek-V4-Flash | $0.14 | $0.0028 | $0.28 |
| DeepSeek-V4-Pro | $0.435 | $0.003625 | $0.87 |
A peak-hour surcharge has been announced but is not active yet — the official rate card still shows one flat tier. We re-check the official rate card weekly and publish every change on our DeepSeek pricing page.
What did not change
DeepSeek is explicit: only the DeepSeek-V4-Flash API was upgraded. The DeepSeek-V4-Pro API and the models serving the app and web chat are unchanged. If you benchmarked Pro last week, those numbers still stand — and the official V4-Pro release is still pending, promised "as soon as possible".
Your migration checklist (it is short)
- Keep model: "deepseek-v4-flash" — no string change, no endpoint change.
- Re-run your agent and tool-calling evals: post-training changed, so behaviour can shift.
- If you run a Codex-style harness, try the native Responses API path and drop your shim.
- Do not switch to V4-Pro for agent work on price grounds alone — Flash now leads on the published agent suites.
- Re-check the official rate card before you commit budget; no price change was announced with this release.
Frequently asked questions
Do I need to change my code for the official DeepSeek-V4-Flash release?
No. DeepSeek states the API call is unchanged: keep model set to deepseek-v4-flash and you are served the latest version. There is no new model string, no new endpoint and no schema change.
Is DeepSeek-V4-Flash-0731 a new architecture?
No. DeepSeek says the model structure and size of DeepSeek-V4-Flash-0731 are identical to DeepSeek-V4-Flash-Preview — only the post-training was redone. The gains are alignment and agent-behaviour gains, not a new base model.
How do the agent benchmarks compare to V4-Pro-Preview?
DeepSeek reports that the official V4-Flash substantially exceeds V4-Pro-Preview on agent benchmarks: Terminal Bench 2.1 82.7, NL2Repo 54.2, Cybergym 76.7, DeepSWE 54.4, Toolathlon verified 70.3, Agent Last Exam 25.2, Automation Bench (Public) 25.1, DSBench-FullStack 68.7 and DSBench-Hard 59.6. Code-agent scores were measured with DeepSeek Harness minimal mode at the max tier, top_p 0.95, temperature 1.0.
Did DeepSeek-V4-Pro or the app change on July 31, 2026?
No. DeepSeek explicitly notes that only the DeepSeek-V4-Flash API was upgraded. The DeepSeek-V4-Pro API and the models behind the app and web chat are unchanged. An official V4-Pro release is promised as soon as possible.
Did prices change with the V4-Flash GA release?
No price change was announced with the July 31, 2026 update. DeepSeek-V4-Flash remains $0.14 per 1M input tokens on a cache miss, $0.0028 on a cache hit and $0.28 per 1M output tokens. A peak-hour surcharge has been announced but is not active yet — the official rate card still shows one flat tier.
Does V4-Flash support the Responses API and Codex?
Yes. The official V4-Flash natively supports the Responses API format and has been specifically adapted for Codex; DeepSeek publishes the configuration steps in its agent-integrations documentation.
Source: DeepSeek's official API changelog entry dated 2026-07-31 (api-docs.deepseek.com/updates). This is an independent fan-run guide and is not affiliated with DeepSeek.