DeepSeek's Flash model just beat its own bigger sibling on every agent benchmark. The bigger-is-better rule is fraying.
Jump to a section
See the story at a glance
DeepSeek shipped V4 Flash 0731 on July 31, and there's one comparison worth pausing on before getting into the feature list.
The Flash model — 284B parameters total, 13B active at inference, priced at $0.14 per million input tokens — scored 82.7 on Terminal Bench 2.1. The V4-Pro-Preview, DeepSeek's larger and more expensive model, scored 72.1 on the same benchmark. The Flash Preview from April scored 56.9.
A model activating 13B parameters just outscored a model using considerably more compute on every one of the nine agent benchmarks DeepSeek published. That's the MoE (Mixture of Experts) architecture doing what it's supposed to do: rather than running all parameters on every pass, DeepSeek trained specialized sub-networks for different task types. At inference, only the relevant slice activates. The result is a model that costs far less to run but performs like a bigger one on the tasks those sub-networks cover well — in this case, agentic coding work.
The Flash Preview shipped in April at 56.9 on Terminal Bench. The 0731 build posts 82.7. Same architecture. Only post-training was redone.
Benchmarks in detail
| Model | Terminal Bench 2.1 | DeepSWE | Cybergym | Price (input/output) |
|---|---|---|---|---|
| V4 Flash 0731 (official) | 82.7% | 54.4% | 76.7% | $0.14 / $0.28 per M |
| V4-Pro-Preview (larger) | 72.1% | — | — | higher |
| V4 Flash Preview (April) | 56.9% | — | — | $0.14 / $0.28 per M |
Terminal Bench 2.1 is the one I weight most here. It measures realistic tool-use in command-line agent tasks — the kind of thing you'd actually run in a CI pipeline or an automated coding agent. DeepSWE and Cybergym fill in the picture on software engineering and adversarial evaluation respectively. None of these are DeepSeek's own invented benchmarks; they're field-standard or independent.
The 0731 build also ships with native Codex and Responses API support, meaning existing Codex-pattern workflows drop in without changes. Weights are MIT-licensed on Hugging Face, ungated — commercial use without asking permission.
Source spread
- MarkTechPost — DeepSeek upgrades V4 Flash 0731 [builder] — Clean benchmark summary, confirms architecture and post-training-only change.
- Kingy — V4 Flash 0731 adds Codex, Terminal Bench 82.7 [builder] — Most detailed on API support and agentic configuration specifics.
- OpenRouter — DeepSeek V4 Flash 0731 pricing and benchmarks [builder] — Live API entry; pricing and availability confirmation.
- Digital Applied — V4 Flash 0731 agent benchmarks [skeptic] — Notes that DeepSWE and Cybergym results for V4-Pro-Preview aren't published for a direct apples-to-apples comparison.
Pros & cons
What's worth getting excited about:
- 82.7 on Terminal Bench 2.1 is a real number on a real benchmark, and it's better than the bigger model at the same price. That's the line that matters.
- Post-training quality, not raw parameter count, produced the gains. Same architecture, same weights size as the preview — just better post-training. That has implications for how competitive this space is going to be. You don't have to build a bigger model to stay ahead; you have to train it better.
- MIT license is clean. Apache 2.0 is technically friendlier for some patent-protection scenarios, but MIT is fine for the vast majority of commercial use. The weights are real weights you can actually download and run.
- Codex/Responses API compatibility means migration from existing Codex setups is a configuration change, not a rewrite.
What deserves skepticism:
- The Pro model comparison is incomplete. DeepSeek published DeepSWE and Cybergym only for Flash 0731, not for V4-Pro-Preview, so the "beats every benchmark" claim is technically "beats every benchmark where both were measured." The Terminal Bench comparison is clean; the others aren't.
- 82.7 on Terminal Bench is outstanding for a model at this price. For complex reasoning tasks — not agent throughput — V4-Pro-Preview likely still has the edge. Flash wins on agentic speed; don't default to it for deep reasoning chains.
- MIT license covers the weights, not the training recipe. You get the artifact. DeepSeek's curation and RLHF process are still proprietary.
What builders need to know
- Upgrade path from Flash Preview is a drop-in. Same API endpoint (
deepseek-v4-flash), same pricing, dramatically better agent performance. Run your eval suite and flip. - $0.14/$0.28/M is the number that matters for cost modeling. At this price, running 100M input tokens daily costs $14. That's within reach for builders who previously couldn't afford agentic workflows at scale.
- Self-hosting via MIT weights is real but still compute-heavy. 284B MoE needs serious hardware to run at low latency — 13B active doesn't mean 13B inference costs in VRAM terms. Plan accordingly.
- Codex/Responses API parity means zero migration work if you're running Codex-pattern loops. Native support confirmed; no adapter needed.
- For complex multi-step reasoning — chains where you need the model to hold and manipulate a lot of context over many steps — V4-Pro-Preview is still the right call. 0731 wins on throughput-oriented agent tasks, not depth.
- Don't conflate Terminal Bench 82.7 with "best model." It's the best cheap model for agent command-line tasks. That's a specific, valuable niche. It's not a universal ranking.
Further reading
- MarkTechPost — DeepSeek upgrades V4 Flash 0731 with agentic gains — benchmark detail and architecture notes
- Kingy — V4 Flash 0731 Codex support and Terminal Bench — most complete on API surface and agent configuration
- OpenRouter — V4 Flash 0731 live pricing — current pricing and availability status
- Digital Applied — V4 Flash 0731 official release and benchmarks — skeptical comparison notes
Liked this? Get the weekly digest.
Free. Monday mornings. The week's stories, synthesized. Unsubscribe anytime.
Your take
How'd I do on this one?
What did I miss?
Tell Samwise (and Sam).
Disagree with the take? Spotted a fact I got wrong? Have context I should have included? Drop it here. Anonymous unless you leave an email.
Keep following the thread
A little more to explore.
Google opened Carbon to everyone. The price doubles in three weeks.
Gemini 4 Carbon became available to all paying API developers on November 10, 2026, at an introductory rate of $2 per million input tokens — 75% cheaper than Claude Fable 5.2. The intro rate expires November 30. Terminal-Bench 4.0 scores range from 63% to 65% across evaluators, and Google previewed /swarm, a multi-agent coordination feature still in limited access.
Fable 5.2 leads the coding benchmarks. The science numbers depend on who's running them.
Anthropic shipped Claude Fable 5.2 and Mythos 5.2 on November 9 at $8/$40 per million tokens — 20% cheaper than Fable 5.1 and GPT-6 Astra. On coding and agentic benchmarks the lead holds across evaluators. On science tasks, Anthropic's first-party and Vals's independent harness disagree by 19 points.
Claude Haiku 5.5 stopped being the compromise option
Claude Haiku 5.5 landed October 7 at $0.10/M input tokens — a tenfold cut from Haiku 4.5's $1.00/M. Terminal-Bench 4.0 went from 0% to 39.2%. The model tier you routed around is worth reconsidering.