On July 31, 2026, DeepSeek updated its deepseek-v4-flash model to the DeepSeek-V4-Flash-0731 build. Unusually for a release with gains this large, it is not a new model generation: DeepSeek states the architecture and parameter count are unchanged from the preview build. Only the post-training was redone, and almost all of the improvement lands on agentic and coding work.
The Benchmark Jump
Three of the nine published benchmarks have preview-build comparisons, and they show how lopsided the retraining was:
- Terminal Bench 2.1: 61.8 to 82.7
- DeepSWE: 7.3 to 54.4
- DSBench-FullStack: 37.0 to 68.7
The remaining published scores for the official build: Cybergym 76.7, Toolathlon (verified) 70.3, DSBench-Hard 59.6, NL2Repo 54.2, Agent Last Exam 25.2, and Automation Bench (Public) 25.1. For scale, V4-Pro-Preview scores 72.1 on Terminal Bench 2.1 — meaning the far cheaper Flash model now runs ahead of the previous flagship preview on that test. DeepSeek reports the official Flash build matching or beating V4-Pro-Preview across several of the nine benchmarks.
Worth reading with care: these are DeepSeek own published numbers on agent-flavored benchmarks rather than independent evaluations, and a jump from 7.3 to 54.4 on DeepSWE says as much about how weak the preview build was on that test as about how strong 0731 is. On the hardest long-horizon coding work, frontier models still lead. What the release genuinely moves is the price-to-capability floor for routine agent loops.
The Model ID Did Not Change
The most practical detail for anyone already running DeepSeek: the calling convention is untouched. Per DeepSeek release note, you still request deepseek-v4-flash and are served the latest build. There is no migration, no new model string, and no redeploy — existing agents were moved onto 0731 automatically.
Responses API and Pricing
The 0731 build natively supports the Responses API format and is advertised as compatible with Codex-style clients. Pricing is unchanged and stays aggressive: on the direct DeepSeek API, Flash is $0.14 per million input tokens on a cache miss, $0.0028 on a cache hit, and $0.28 per million output tokens, against $0.435 / $0.87 for V4 Pro. Flash also carries a 2,500 concurrency limit versus Pro 500, which matters for busy always-on bots.
DeepSeek V4 Flash is available in the model picker on OpenClaw Launch, and any instance already using it is on the 0731 build. Our full write-up, including the complete benchmark table and setup steps, is in the DeepSeek V4 Flash guide.