DeepSeek V4 Flash 0731: The Small Model That Beat Its Big Sibling
DeepSeek's V4 Flash graduated from preview to official release on July 31, 2026, badged V4-Flash-0731. Nothing changed about its architecture — same 284-billion-parameter mixture-of-experts design, same roughly 13 billion active parameters per token. What changed was the post-training pass, and the resulting jump in agentic benchmark scores was large enough that the smaller Flash model now beats DeepSeek's own larger V4 Pro on nearly every published agent test, at about a third of the price.
Same silicon, same architecture — DeepSeek's gains here came entirely from retraining, not a bigger model.
Same Architecture, Completely Retrained
DeepSeek's July 31, 2026 update moved V4 Flash out of preview and into general availability as DeepSeek-V4-Flash-0731, replacing the earlier preview API endpoint. The company's own changelog is explicit that the model "was only re-post-trained" — the underlying architecture, a 284-billion-parameter mixture-of-experts design with roughly 13 billion parameters active per token and a 1-million-token context window, is identical to the preview build. Every capability gain in this release comes from what the model was trained to do after pretraining, not from any change to its size or shape.
That distinction matters because the benchmark jump is large enough that it would normally suggest a bigger model. Terminal-Bench 2.1 rose from 61.8 to 82.7. DeepSWE, an agentic software-engineering benchmark, jumped from 7.3 to 54.4 — roughly a 7.5x improvement. On DeepSeek's own published table, the retrained Flash-0731 now beats its own larger sibling, V4-Pro-Preview, on the large majority of published agentic and coding benchmarks, despite activating fewer than a third of V4 Pro's parameters per token.
What Improved, and By How Much
V4 Flash 0731 supports five times the concurrent request volume of its larger V4 Pro sibling.
Terminal-Bench 2.1
82.7, up from 61.8 on the preview — a 20.9-point gain, though it still trails Claude Opus 4.8's reported 85.0 on the same test.
DeepSWE
54.4, up from 7.3 — the largest relative jump in the release, and now ahead of V4-Pro-Preview's 12.8.
Cybergym
76.7 on DeepSeek's published table, part of the broader agentic-capability improvement across the board.
Concurrency
2,500 concurrent requests supported, versus 500 for V4 Pro — a five-times headroom advantage that matters more to production agent fleets than any single benchmark score.
It's worth being precise about what these numbers do and don't prove. Every figure above comes from DeepSeek's own evaluation harness, and independent verification is still catching up — Artificial Analysis's separate Intelligence Index showed a roughly ten-point improvement over the preview, which corroborates the direction of the gain without matching DeepSeek's exact methodology. Across the nine benchmarks DeepSeek published comparing V4 Flash 0731 to Claude Opus 4.8, Flash-0731 trailed on every single one — the closest gaps were Terminal-Bench 2.1 (82.7 vs 85.0) and Agents' Last Exam (25.2 vs 25.7), while wider gaps showed up on tasks like NL2Repo (54.2 vs 69.7). The honest framing: V4 Flash 0731 doesn't set a new capability ceiling, it moves the price-to-capability curve substantially for anyone who doesn't need the absolute top score.
Where the Real Value Is
Pricing carried over unchanged from the preview: $0.14 per million input tokens on a cache miss, $0.0028 per million on a cache hit, and $0.28 per million output tokens — a fraction of most agent-tier APIs, and roughly a third of what DeepSeek charges for its larger V4 Pro model. The release also added native support for OpenAI's Responses API format and Codex compatibility, meaning tooling built around OpenAI's interface can call V4 Flash 0731 without a compatibility shim.
The concurrency figure is arguably the most operationally significant detail in the whole release. Supporting 2,500 concurrent requests against V4 Pro's 500 matters enormously for anyone running high-throughput agent pipelines — code review bots, browser agents, batch document processing — where the bottleneck is often how many requests you can run in parallel, not how smart any single request is. Combined with the price, that positions V4 Flash 0731 less as a benchmark-chasing flagship and more as the practical default for cost-sensitive, high-volume agentic workloads.
Who Actually Benefits From This Update
- High-throughput agentic pipelines where concurrency matters more than peak score
- Cost-sensitive coding and tool-calling workloads at scale
- Teams already using OpenAI's Responses API format who want a cheaper drop-in
- Tasks that genuinely need the highest available accuracy, where Opus 4.8 still leads
- Workloads where independent, third-party benchmark verification is a hard requirement
- Anything where DeepSeek's data handling and hosting terms don't fit your compliance needs
Worth Testing on Your Own Workload
A genuinely large jump from post-training alone — verify it on your own tasks before migrating fully.
The direction of the improvement is real and corroborated by at least one independent benchmark, and the price-to-throughput math is compelling enough that it's worth a serious look for any high-volume agentic pipeline. But every headline number here comes from DeepSeek's own harness, comparing against its own older model, so the responsible move is the same one that applies to any vendor-published benchmark chart: run your own task suite through both the model you're on now and V4 Flash 0731, and compare cost per successfully completed task rather than trusting the leaderboard alone.
Frequently Asked Questions
What changed in DeepSeek V4 Flash 0731 versus the preview?
The architecture is identical — same 284-billion-parameter mixture-of-experts design with about 13 billion active parameters per token. The only change was the post-training pass, which was retuned for agentic and coding workloads, producing large jumps on benchmarks like Terminal-Bench 2.1 and DeepSWE.
How much does DeepSeek V4 Flash 0731 cost?
$0.14 per million input tokens on a cache miss, $0.0028 per million on a cache hit, and $0.28 per million output tokens — unchanged from the preview pricing and roughly a third of what DeepSeek charges for its larger V4 Pro model.
Is DeepSeek V4 Flash 0731 better than Claude Opus 4.8?
No. Across every benchmark DeepSeek published comparing the two, V4 Flash 0731 trailed Opus 4.8, though the gap is narrow on some tests like Terminal-Bench 2.1. Its real advantage is price and concurrency, not raw peak capability.
More AI Guides, Written Clearly
Browse our full library of AI tool comparisons and explainers, updated as the category evolves.
More from the creator
Explore two other projects from the same creator.