Skip to content
NewsModel release

DeepSeek V4-Flash-0731 enters public beta, outscoring V4-Pro-Preview on all nine agent benchmarks through retraining alone

· by Pondero Newsdesk

The short version

DeepSeek moved its V4-Flash API to public beta on July 31, 2026 with a build that beats V4-Pro-Preview on every published agent benchmark at $0.14 per million input tokens, through re-post-training on an unchanged 284B-parameter architecture.

DeepSeek V4-Flash-0731 enters public beta, outscoring V4-Pro-Preview on all nine agent benchmarks through retraining alone

DeepSeek's budget Flash tier now outscores its own premium Pro-Preview on every agent benchmark the company published, at the same price it shipped in April. V4-Flash-0731 moved to public beta on July 31, per DeepSeek's API changelog. Every benchmark gain came from a second round of post-training. The parameter count, architecture, and knowledge base stayed untouched; only the behavioral post-training was rerun.

What changed

V4-Flash-0731 runs the same 284 billion total parameters with 13 billion active per token and a 1 million token context window as the April 2026 preview, per Artificial Analysis. What DeepSeek ran was a targeted re-post-training cycle for agent and coding tasks, not a new base model.

On the nine benchmarks DeepSeek published, the 0731 build outscores V4-Pro-Preview on every entry, per DeepSeek's changelog. Terminal Bench 2.1 shows the clearest gap: 82.7 for V4-Flash-0731 against 72.1 for V4-Pro-Preview, confirmed by TechTimes. DeepSWE, which evaluates real-world GitHub issue resolution across 113 repositories, rose from 7.3 on the Flash Preview to 54.4 on the 0731 build. Other published scores include NL2Repo at 54.2, Cybergym at 76.7, and Toolathlon Verified at 70.3.

Artificial Analysis rated the model 50 on its Intelligence Index, up 10 points from the prior V4-Flash's score of 40, per its July 31 evaluation. Hallucination rate dropped 12 percentage points to 84%, with overall accuracy steady at 37%.

Pricing holds at $0.14 per million input tokens and $0.28 per million output tokens, per DeepSeek's pricing page. Teams already calling deepseek-v4-flash received the upgrade automatically. Build 0731 also natively supports the Responses API format and carries specific training for the Codex endpoint, per DeepSeek.

Why it matters

V4-Pro charges $0.87 per million output tokens; V4-Flash charges $0.28. A coding-agent pipeline consuming 100 million output tokens monthly costs roughly $28 on Flash against roughly $87 on Pro, at a point where Flash now scores higher on agent benchmarks. That pricing gap was easier to justify before July 31.

Three caveats apply before treating the agent scores as settled. First, DeepSeek's benchmark harness for coding evaluations has not been released publicly, so independent replication is not yet possible, per notes in DeepSeek's own changelog. DSBench-FullStack and DSBench-Hard are explicitly internal test sets. Second, full open weights for the 0731 build were not yet available at publication. Artificial Analysis noted DeepSeek expects to release them in the coming weeks. Third, thinking mode is on by default in V4-Flash, and the reasoning tokens the model generates internally are billed at the output token rate even though they do not appear in the final response, per TechTimes' analysis. Developers running extended agent reasoning chains will see effective cost per task above the headline per-token figures.

For API routing, as of the release announcement only DeepSeek's own provider served the 0731 build. Teams using OpenRouter or similar aggregators may still receive the April preview build until those providers update their routing.

What to watch next

DeepSeek stated in the same July 31 changelog that an official V4-Pro release will follow. If V4-Pro undergoes comparable re-post-training, its benchmark profile could shift substantially from current preview numbers. Third-party evaluation of V4-Flash-0731 on Terminal Bench 2.1 and DeepSWE under independent harnesses will determine how much of the July 31 improvement holds under contamination-controlled conditions.

Sources