On July 31, DeepSeek shipped a 0731 refresh of V4-Flash. The official changelog is blunt about the scope: DeepSeek-V4-Flash-0731 “keeps the same model architecture and size as DeepSeek-V4-Flash-Preview, and was only re-post-trained.” The changes concentrate on agent capabilities — the company says the new benchmark results “far exceed V4-Pro-Preview” — plus native Responses API support, specifically adapted for Codex. Only the V4-Flash API was upgraded; the V4-Pro API and the APP and WEB models are unchanged.
That last sentence contains the real headline. A Flash-tier model claiming to leap past its own Pro-tier sibling is a signal about where agent competition has moved: it is no longer a flagship-only game, and the cheapest lever is not more parameters.
Same Architecture, Retrained Back Half
The agent benchmarks DeepSeek published are specific. Terminal Bench 2.1: 82.7. NL2Repo: 54.2. CyberGym: 76.7. DeepSWE: 54.4. Toolathlon (verified): 70.3. Agent Last Exam: 25.2. The test conditions are public too — DeepSeek Harness in minimal mode, max effort level, top_p 0.95, temperature 1.0. These are not numbers that depend on cherry-picked settings, and publishing the harness configuration alongside the scores is itself a small statement about how model claims should be read.
Native Responses API support deserves its own note. Responses is the interface spec OpenAI has been pushing, and DeepSeek explicitly says the model was adapted for Codex. When an open-weights lab wires compatibility straight into the Codex toolchain, the competition moves from leaderboard scores to the developer’s wiring cost. What used to require an extra adapter layer now amounts to changing one model name in a config file.
Third-Party Numbers: 35 on the Intelligence Index, 210 Tokens per Second
Artificial Analysis’ evaluation puts the 0731 release on a clearer map. It scores 35 on the Intelligence Index, ranking 9th of 113 models — well above the 18-point median among comparable models. Speed is the louder headline: 210.1 output tokens per second, third overall across the entire ranking, with a 0.98-second time to first token. The median for similarly sized open-weight models is 66.8 tokens per second, so DeepSeek is shipping roughly three times the class norm.
Pricing stays aggressively DeepSeek: $0.44 per million input tokens and $1.32 per million output tokens through the first-party API, about $0.23 blended at their standard mix, with a 97% cache discount on hits. The model is a mixture-of-experts network with 284B total and 13B active parameters, a 1M-token context window — roughly fifteen hundred A4 pages — and MIT-licensed open weights cleared for commercial use, served by 17 API providers.
There is one catch worth knowing before you migrate a cost model onto it. During evaluation it generated about 240 million output tokens against a 140 million median — the model is verbose. The per-token price is real, but real-world bills will run above the sticker math unless you prompt for brevity or post-process aggressively.
What It Means for Developers
Three takeaways. First, the cost curve for agent workloads dropped again. A top-ten model on the Intelligence Index that costs a fraction of the closed frontier flagships per output token — and that you can self-host under MIT — resets what a production agent budget should look like. If your agent loop burns tokens on tool calls and retries, throughput at 210 tokens per second also shortens wall-clock latency, which is often the difference between an agent that ships and one that annoys.
Second, the update pattern is worth copying inside your own stack. Freeze the architecture, retrain only the post-training, ship. It is the cheapest path from research model to product model, and because the interface does not move, integrations do not break. DeepSeek has now done this across the V4 line, and the cadence is becoming part of its product identity.
Third, interfaces follow ecosystems. Native Responses API plus explicit Codex adaptation shows that open-weights competition now extends to toolchain compatibility. Benchmark rankings get the attention, but the boring question — does this drop into my existing agent harness without new glue code — increasingly decides what teams actually deploy.
One housekeeping note: the old deepseek-chat and deepseek-reasoner aliases, introduced when V4 launched in April, were retired on July 24. If any of your code still points at those names, migration is no longer optional.
Sources
AI-assisted summary compiled from the sources above, reviewed by a human before publishing.