News

DeepSeek Releases V4 Flash Update with Stronger Agent Scores and Unchanged Pricing

• By John K. Waters
• 08/07/2026

DeepSeek has released an updated version of its V4 Flash model through an official API that remains in public beta, pairing a steep price discount relative to its flagship model with agent-benchmark scores that the Chinese AI lab says beat its larger Pro model.

The update, dated July 31, 2026, changes neither the architecture nor the model size. It is, by DeepSeek’s account, a post-training upgrade, and the distinction matters for organizations deciding whether to route production traffic to it.

What changed, and what did not
V4 Flash keeps the same footprint as the earlier preview build, a mixture-of-experts model with 284 billion total parameters and 13 billion active per token, along with a context window of up to one million tokens. DeepSeek says only the post-training process was revised.

The V4 Pro API, along with the DeepSeek app and web chat, was not changed as part of the update. DeepSeek says the official release of V4 Pro “will follow soon,” but it has not provided a date. Although the company describes the new Flash build as its official release, its API changelog says the service remains in public beta.

Pricing held steady. DeepSeek’s published rate card lists V4 Flash at $0.14 per million input tokens on a cache miss and $0.28 per million output tokens. The cache-hit rate for repeated prompt prefixes is $0.0028 per million input tokens.

Those figures put Flash at roughly one-third of the list price DeepSeek charges for V4 Pro, which costs $0.435 per million input tokens and $0.87 per million output tokens. DeepSeek says it expects to raise its overall API prices significantly in the future, but it has not announced the new rates or an effective date.

The benchmark case
DeepSeek reports that the new Flash build outscores its V4 Pro Preview on the coding and agent tests included in the company’s published comparison. The results include a Terminal-Bench 2.1 score of 82.7, up from 61.8 for the earlier Flash build and above the 72.1 reported for V4 Pro Preview.

The company also reports a DeepSWE score of 54.4, compared with 7.3 for the previous Flash version and 12.8 for V4 Pro Preview. Other reported gains include a Cybergym score of 76.7 and a Toolathlon-Verified score of 70.3, according to the DeepSeek model card.

Those numbers come from DeepSeek’s evaluation process. The company says it tested public coding-agent tasks using the minimal mode of its DeepSeek Harness at the maximum reasoning-effort setting, with a temperature of 1.0 and top_p set to 0.95. DeepSeek describes that version of the harness as “to be released soon.”

Two of the cited tests, DSBench-FullStack and DSBench-Hard, are internal DeepSeek evaluations. The published results therefore describe what the company measured rather than independently reproduced performance.

Outside testing provides a partial check. Artificial Analysis initially gave DeepSeek V4 Flash 0731 a score of 50 on its Intelligence Index. As of August 7, its current model page reported a score of 52 for the reasoning model at the maximum effort setting.

Artificial Analysis measured output speed at about 102 tokens per second. It also described the model as unusually verbose, reporting that it generated 210 million output tokens during the Intelligence Index evaluation, compared with a median of 100 million among comparable models.

That finding matters for cost comparisons. DeepSeek’s token rates are low, but a model that produces longer responses can consume some of those savings. Actual costs will depend on the prompts, reasoning settings, and output controls used in production.

Independent evidence remains less extensive than DeepSeek’s own benchmark table. The available results support the view that the updated model offers strong performance for its price, but they do not independently reproduce every agent score in the company’s announcement.

What it's built for
The release adds native support for the Responses API format and includes adaptations for OpenAI’s Codex, changes aimed primarily at developers building coding and tool-calling agents rather than ordinary chatbot users.

DeepSeek has also released the July checkpoint as open weights under an MIT license. The DeepSeek-V4-Flash-0731 repository, published through DeepSeek’s Hugging Face account, is listed at 167 GB and includes the model’s safetensor files.

The release allows researchers and infrastructure providers to run the updated model independently. It also makes outside reproduction of the company’s results more feasible, although publishing the weights does not validate DeepSeek’s benchmark claims.

For buyers weighing model choice against cost, the practical takeaway is narrower than the headline figures suggest. The API pricing and benchmark gains apply to the V4 Flash endpoint, not to DeepSeek’s consumer products. The comparisons anchoring the announcement were produced by the vendor, while independent testing shows strong value but also flags high token usage.

Organizations considering the model for production would still need to test it against their own tools, latency requirements, and usage patterns. The revised model appears inexpensive and capable, but the available evidence does not eliminate the need for workload-specific evaluation.