DeepSeek did not launch a new model on July 31. It launched a new brain inside an old one — and then left the price tag exactly where it was.
The deepseek-v4-flash endpoint now serves DeepSeek-V4-Flash-0731, a public beta that keeps the same API name, the same base URL, the same 284B-total/13B-active Mixture-of-Experts architecture, and the same $0.14 per million input tokens. What changed is post-training. And according to independent measurement, that change bought roughly a generation's worth of capability for free.
What actually shipped
This is the least glamorous kind of AI release and quite possibly the most useful one. Per DeepSeek's official change log, the 0731 update is API-only: no new architecture, no new parameter count, no new context window.
| Spec | Value |
|---|---|
| Total / active parameters | 284B / 13B (MoE) |
| Context window | 1M tokens |
| Max output | 384K tokens |
| Modality | Text in, text out |
| Input price (cache miss) | $0.14 / 1M |
| Output price | $0.28 / 1M |
| Input price (cache hit) | $0.0028 / 1M |
The V4 family's architecture is unusual enough to be worth restating: it combines Compressed Sparse Attention with Heavily Compressed Attention, adds mHC connections, and was trained with the Muon optimizer. The official open-weight checkpoint uses mixed FP4 expert parameters with mostly FP8 non-expert parameters — a detail that matters if you plan to self-host.
The API also picked up real plumbing. It now speaks the Responses API alongside OpenAI Chat Completions and the Anthropic API, and supports tool calls, JSON output, prefix completion, and FIM completion in non-thinking mode.
The benchmark jump is not subtle
Artificial Analysis scored 0731 at 50 on its Intelligence Index, a 10-point leap over the April 2026 DeepSeek V4 Flash, which sat at 40. That is enough to put the cheap Flash model 6 points ahead of DeepSeek V4 Pro — an awkward result for DeepSeek's own product ladder.
Where it lands relative to everyone else:
| Model | AA Intelligence Index |
|---|---|
| Kimi K3 (max) | 57 |
| GPT-5.6 Luna (max) | 51 |
| GLM-5.2 (max) | 51 |
| DeepSeek V4 Flash 0731 | 50 |
| Gemini 3.6 Flash | 50 |
| DeepSeek V4 Flash (April) | 40 |
Every single evaluation in the index improved. CritPt gained 9 points to 17%, SciCode 5 points to 50%, Humanity's Last Exam 5 points to 37%, AA-LCR 3 points to 66%, and GPQA Diamond 1 point to 91%.
The agentic numbers are where the story lives. On GDPval-AA v2, Artificial Analysis's evaluation of real-world work tasks, 0731 posts an Elo of 1559 — up from 1189. That is not an incremental bump; that is a different class of model. Terminal-Bench 2.1 rose 17 points to 79% on AA's harness, and τ³-Bench Banking gained 8 points to 31%.
DeepSeek's own reported agent figures run a little higher, as vendor numbers usually do. The company cites Terminal Bench 2.1 at 82.7, NL2Repo at 54.2, CyberGym at 76.7, DeepSWE at 54.4, Toolathlon Verified at 70.3, Agents' Last Exam at 25.2, and Automation Bench (Public) at 25.1, with internal DSBench-FullStack and DSBench-Hard results of 68.7 and 59.6.
Read the fine print on those: DeepSeek states its public Code Agent evaluations used a forthcoming minimal harness at max effort, temperature 1.0, and top_p 0.95. Harness choice moves agentic scores by several points. Treat the 79% and the 82.7% as measurements of two different setups, not as a contradiction.
The price is the headline
Here is the number that should make procurement teams pause. At the end of July, OpenAI cut GPT-5.6 Luna's price by 80%, taking it from $1/$6 to $0.20/$1.20 per million input/output tokens. Even after that cut, Artificial Analysis calculates DeepSeek V4 Flash 0731's Cost per Task on DeepSeek's first-party API at roughly 60% below GPT-5.6 Luna (max) — a model one single index point ahead of it.
Note what that repricing did and did not achieve. At $0.20 input, Luna is still above DeepSeek's $0.14, and OpenAI left flagship Sol untouched at $5/$30. The 80% cut was a defensive move aimed squarely at this price band.
The mechanism is a cache discount most people are underrating. DeepSeek charges $0.0028 per million cache-hit input tokens, a 98% discount, against the 90% that most of the industry offers. For agentic workloads — where the same system prompt, tool schema, and file context get replayed on every turn — that 8-point gap compounds into something enormous.
There is a second efficiency win. Running the full Intelligence Index, 0731 burned ~206M output tokens versus ~234M for its predecessor. It got smarter and 12% less verbose, which is close to the opposite of what reasoning-model scaling usually delivers.
One caution: DeepSeek has announced future peak pricing at twice the regular rate. No effective date has been published, so it is not in the active price table — but budget for it.
The honest asterisk
The intelligence gain is real. The reliability gain is narrower than it looks.
On AA-Omniscience, 0731 improves 7 points to an index score of -16. But that improvement comes entirely from a lower hallucination rate — the frequency with which the model produces a confident wrong answer instead of declining. That rate fell roughly 11 points to 84%. Raw accuracy is unchanged at 37%.
In plain terms: it has not learned more facts. It has gotten somewhat better at not inventing them. For a 284B model that is exactly what you would predict, and it is the right kind of progress — but do not read a 10-point index jump as a knowledge upgrade.
The weights question
This is the part people are getting wrong on social media. The 0731 weights are not open.
The April 24, 2026 DeepSeek-V4-Flash checkpoint remains on Hugging Face under the MIT License, commercial use permitted. DeepSeek explicitly states the 0731 update applies only to the API — the app, the web model, and the published weights were not replaced. Artificial Analysis notes full weights are expected in the coming weeks, and if they land, this becomes the second-highest open-weights score on GDPval-AA v2, behind only Kimi K3 (1687) and ahead of GLM-5.2 (1510).
Until then, anyone quoting 0731 benchmarks while pointing at the Hugging Face repo is describing two different models.
The Bottom Line
DeepSeek V4 Flash 0731 is the clearest evidence yet that post-training, not parameter count, is where the cheap wins now live. Same 284B/13B architecture, same $0.14, ten more index points, and a 370-Elo jump on real work tasks.
If you are building agents and paying frontier prices for tool-calling loops, run your eval suite against deepseek-v4-flash this week — the 98% cache discount alone may rewrite your unit economics. If you need reliable factual recall, the flat 37% accuracy says look elsewhere. And if your deployment story depends on open weights, wait for the checkpoint. It is coming, but it is not here.
