On July 31, 2026, DeepSeek posted a number I did not expect this late in the V4 cycle: the official V4 Flash build, DeepSeek-V4-Flash-0731, scores 82.7 on Terminal Bench 2.1 — and per DeepSeek's own changelog, that "far exceeds" V4-Pro-Preview, the model that is nearly six times its size. A 284B open-weights model topping its own bigger sibling on agentic work is the story here.

What makes it stranger: 0731 is not a new model. The changelog is explicit that it "was only re-post-trained." Same 284B-parameter MoE as the April 24 preview, same 13B active per token, same 1M-token context, same MIT license. All of the gain below comes from post-training and a wider API surface.

What actually shipped

Three things, per the official changelog:

  1. Re-post-training, not a new architecture. Same size, same context. The benchmark jumps are purely a post-training result.
  2. Native Responses API support, and the build is "specifically adapted for Codex" — harnesses built on OpenAI's Responses format work without an adapter layer.
  3. Public beta on the deepseek-v4-flash model id. The V4-Pro API and app/web models are untouched; DeepSeek says the official V4-Pro release is "coming soon."

The benchmarks

From the official changelog:

Benchmark DeepSeek-V4-Flash-0731
Terminal Bench 2.1 82.7
Cybergym 76.7
Toolathlon (verified) 70.3
DeepSWE 54.4
NL2Repo 54.2

For scale from what I have tracked on this blog: 82.7 on TB2.1 is above the 81.0 that GLM-5.2 posted back in June, and above the 73.0 from Qwen 3.8-27B — the strongest open-weights model you can run on a single 24 GB card. Cross-lab, same-benchmark comparisons always carry the usual harness caveats, but this is the first time a mid-size open-weights model has sat on top of the TB2.1 board I track.

Independent measurement is directionally in line. Artificial Analysis scores the 0731 reasoning build (max effort) at 50 on Intelligence Index v4.1 — #2 of 162 models measured, against a median of 17. Their caveat matters, and I will come back to it: the model is verbose, generating 210M tokens across the eval suite versus a 62M median.

Pricing and context

First-party API pricing, unchanged by this release:

Price per 1M tokens
Input (cache miss) $0.14
Input (cache hit) $0.003
Output $0.28

That output rate is roughly a third of V4 Pro's $0.87, and the cache-hit input rate is a 98% discount. Context is 1M tokens with 384K max output — whole repositories and long agent traces fit without aggressive compaction.

How I would run it

Two reasoning variants are exposed, high and max. In practice, the split is clean:

  • max for the multi-step loops this model is being built for — plan, run a command, read the output, self-correct. That is the Terminal-Bench shape of work.
  • high for single-file edits, review passes, and anything short-horizon. It is faster, cheaper, and given the verbosity numbers above it is the single easiest cost lever you have.

Through OpenCode it is one line:

opencode run --model opencode-go/deepseek-v4-flash --variant max \
  "find the flaky test in this repo and explain why it fails"

And because of the Responses API adaptation, the same model id also slots into Codex-style harnesses without translation glue — switching harness no longer means switching model.

The catch

I want to be straight about what I have and have not verified:

  • The 82.7 and the rest of the table are self-reported. DeepSeek ran them on their own harness; I have not independently reproduced any of them. The Artificial Analysis ranking is independent but measures a different thing, and it is reassuring rather than confirmatory.
  • Verbosity is a real cost, not a footnote. 210M generated tokens versus a 62M median means "cheap per token" is not "cheap per task." On a long-horizon agent loop, budget on tokens generated, not on the sticker price. A 3.4× token multiplier eats a meaningful chunk of the per-token advantage.
  • "Flash beats Pro" is a snapshot, not a law. V4-Pro-Preview has not had its post-training pass. The moment the official Pro release lands, the ranking in this family can flip, and DeepSeek's own note says it is imminent.
  • This is an API model, not a local one. A 284B MoE is not going on a 5090. For local inference the story remains Qwen/GLM-class — this post is about what the API tier does to pricing and expectations.

What it means

An MIT-licensed 284B model at $0.14/$0.28 per million tokens posting a Terminal Bench score above every other model I track sets a floor. Any lab charging a premium for agentic coding now has to explain what the premium buys, and the answer "bigger model" gets harder to sustain once the small model passes it on exactly the work the premium was supposed to cover. The pattern is the same one the Qwen 3.8-27B post documented from the local side of the market: the gap keeps closing, and now it is closing on price as well as capability.

Sources

Related reading