Tech
DeepSeek V4-Flash-0731 beats its own 1.6T Pro model on all nine agent benchmarks at one-third the output price
MarkTechPost · 2026-07-31 · Source ↗
DeepSeek V4-Flash-0731 beats its own 1.6T Pro model on all nine agent benchmarks at one-third the output price
MarkTechPost · 2026-07-31 · Source ↗
DeepSeek released V4-Flash-0731 as the official post-preview version of its April V4-Flash model on Friday. The gap between "preview" and "official" is not a wrapper change: nine benchmarks moved substantially without a single parameter or attention change to the underlying architecture. Terminal Bench 2.1 rose from 61.8 to 82.7. DeepSWE moved from 7.3 to 54.4, a 7.5x lift on the pure software-engineering agentic evaluation. NL2Repo went from a preview number DeepSeek did not publish to 54.2, beating V4-Pro's 38.5. Cybergym climbed from 38.7 to 76.7, edging V4-Pro's 52.7 by 24 percentage points on the same task set. Toolathlon-Verified reached 70.3, up from V4-Pro's 55.9.
The architecture stays a 284B/13B mixture-of-experts with a 1 million token context window, one shared expert and 256 routed experts per layer, and hybrid attention combining Compressed Sparse Attention with Heavily Compressed Attention. Six routed experts fire per token. The V4-Pro Preview built on the same substrate, and DeepSeek's technical note makes clear the delta is entirely post-training reinforcement work — a claim the community can now audit against the released checkpoint on Hugging Face.
Pricing is the second half of the story
V4-Flash-0731 is priced at $0.14 per million input tokens on the API with cache miss, $0.0028 per million on cache hit, and $0.28 per million output tokens. V4-Pro output pricing runs approximately $0.87 per million, so the Flash upgrade is delivering better benchmark numbers at roughly one-third the output cost of its sibling. On Artificial Analysis Intelligence Index with maximum effort, V4-Flash-0731 posts a score of 50.
"V4-Flash-0731 beats V4-Pro (Preview) on every agentic benchmark." — DeepSeek release note
The reading inside the developer community is that DeepSeek has decided to price the Flash line as a wedge under Anthropic's Sonnet 5 and OpenAI's Sol Terra, using the aggressive $0.14/$0.28 rate to move volume onto their inference at the moment enterprise procurement is rechecking the vendor stack ahead of the EU AI Act enforcement window. The MIT license means the weights are portable to any inference stack that supports the DeepSeek architecture, and lm.studio and DeepInfra listings went live within 12 hours of release.
What the number does not answer
The gain on DeepSWE from 7.3 to 54.4 is large enough to raise the question of test-set contamination. DeepSeek's release note addresses this by publishing per-instance predictions for a sampled subset alongside the score, but external verification is still catching up. Terminal Bench 2.1's public leaderboard has independent submission slots open, and community reruns should land within the next week. If the numbers hold, V4-Flash-0731 becomes the first sub-$0.30 output-per-million-tokens model above 80 on Terminal Bench, which resets the price/performance floor for coding-agent inference.