Summary: DeepSeek shipped the V4-Flash-0731 update on July 31, keeping the preview’s architecture and size while re-running post-training for agent performance. The company reports strong agent benchmark gains and adds Responses API support tuned for Codex, without changing V4-Pro API or app/web models. External testing cited a 35 Intelligence Index and 210.1 output tokens per second, while legacy aliases retired on July 24 need migration.
DeepSeek shipped a 0731 update to V4-Flash on July 31. The official changelog is blunt about scope: DeepSeek-V4-Flash-0731 “keeps the same model architecture and size as DeepSeek-V4-Flash-Preview, and was only re-post-trained.” Engineering effort concentrated on agent performance—the company reports numbers that “far exceed V4-Pro-Preview”—while adding first-class Responses API support tuned explicitly for Codex. The rollout covers the V4-Flash API only; V4-Pro API endpoints and the models powering APP and WEB surfaces stay as they were.
Read that last point carefully. When a Flash-class checkpoint claims to clear its Pro sibling on benchmarks, it signals where agent competition actually happens: flagship SKUs no longer own the category, and the cheapest upgrade path is not more parameters.
Same Blueprint, Retrained Tail
DeepSeek published hard agent scores: Terminal Bench 2.1 at 82.7; NL2Repo 54.2; CyberGym 76.7; DeepSWE 54.4; Toolathlon (verified) 70.3; Agent Last Exam 25.2. Evaluation settings are disclosed alongside—DeepSeek Harness in minimal mode, max effort level, top_p 0.95, temperature 1.0—so the figures are not artifacts of hidden tuning. Publishing the harness recipe next to the leaderboard is its own quiet comment on how benchmark claims should be audited.
Responses API support deserves its own paragraph. Responses is the interface spec OpenAI has been advancing, and DeepSeek states plainly that the build was adapted for Codex. When an open-weights shop lands compatibility inside Codex itself, the fight moves off headline scores and onto wiring friction: what used to require a bespoke adapter often reduces to swapping one model identifier in configuration.
Independent Benchmarks: Index 35, 210 Tokens Per Second
Artificial Analysis places the 0731 build on a sharper grid. It earns 35 on the Intelligence Index—9th among 113 tracked models—comfortably above the 18-point median for peers. Throughput is the louder story: 210.1 output tokens per second, third site-wide, with 0.98 seconds to first token. Comparable open-weight models median at 66.8 tokens per second, so DeepSeek is delivering roughly triple the category norm.
Pricing remains aggressively DeepSeek: $0.44 per million input tokens and $1.32 per million output tokens on the first-party API, about $0.23 blended at the usual mix, plus a 97% cache discount when hits land. Under the hood it is a 284B-parameter MoE with 13B active parameters, a 1M-token context window—roughly fifteen hundred A4 pages—and MIT-licensed weights cleared for commercial use across 17 API providers.
Budget carefully, though. During evaluation it emitted on the order of 240 million output tokens versus a 140 million median, meaning the model runs verbose. Per-token rates are real, yet production invoices tend to sit above napkin math unless prompts enforce brevity or you post-process aggressively.
Developer Takeaways
Three lines to keep. First, agent economics moved again: a top-ten Intelligence Index model whose output tokens cost a sliver of closed frontier pricing—and that you can self-host under MIT—reframes what a production agent budget should assume. At 210 tokens per second, wall-clock latency on tool-heavy loops also compresses, which often separates agents teams ship from agents users abandon.
Second, copy the release mechanics. Freeze architecture, re-run post-training only, ship. It is the leanest bridge from research artifact to productized API, and because the surface contract stays fixed, downstream integrations do not snap. DeepSeek has repeated this playbook across the V4 family often enough that the cadence itself is becoming product identity.
Third, interfaces track ecosystems, not press releases. Native Responses API plus explicit Codex targeting shows open-weights rivalry reaching into toolchain fit. Leaderboards attract attention, but the boring deploy question—does this slot into my existing harness without new glue—increasingly decides what gets production traffic.
Housekeeping: legacy aliases deepseek-chat and deepseek-reasoner, introduced at the April V4 launch, were retired July 24. Any codebase still referencing those strings needs migration now, not later.