DeepSeek releases V4-Flash-0731 with top value per dollar
DeepSeek has released DeepSeek-V4-Flash-0731, the latest model in its V4 family, described as having "substantially enhanced agentic capabilities." The model has 304 billion parameters and downloads as a 167GB file on Hugging Face. Despite that size, benchmark tracker Artificial Analysis ranks it ahead of MiniMax M3, a competing model with 428 billion parameters, roughly 40% more than DeepSeek-V4-Flash-0731 carries. Pricing is $0.14 per million input tokens and $0.27 per million output tokens. Commentator Simon Willison, testing the model via OpenRouter, called this combination of ranking and price possibly the best value-per-intelligence available right now, pointing to the model's strong position on Artificial Analysis's Intelligence Index versus Cost per Intelligence Index Task chart. Willison also ran his standard test of asking the model to draw an SVG pelican riding a bicycle. At the default reasoning level the result was disappointing, but raising the reasoning effort to "high" (via the command llm -m openrouter/deepseek/deepseek-v4-flash-0731 -t pelican -o reasoning_effort high) produced a noticeably better output. The source text does not give a release date, specify what the enhanced agentic capabilities consist of, confirm whether the quoted pricing is DeepSeek's own or OpenRouter's, or provide numeric benchmark scores beyond the ranking claim.
Key facts
- DeepSeek-V4-Flash-0731 has 304 billion parameters and is a 167GB download on Hugging Face
- Artificial Analysis ranks it ahead of MiniMax M3, a 428 billion parameter model
- Pricing is $0.14 per million input tokens and $0.27 per million output tokens
- Reviewer Simon Willison called it possibly the best value-per-intelligence model currently available
- Willison's pelican-drawing test looked poor at default reasoning level but improved markedly at high reasoning effort
Why it matters
DeepSeek keeps pushing the price-to-capability curve down. V4-Flash-0731 is being noticed not for raw size (at 304 billion parameters it is smaller than the MiniMax M3 model it reportedly outranks) but for landing near the top of Artificial Analysis's Intelligence Index while charging a fraction of what larger rivals ask per token. That combination, rather than any single benchmark score, is what is driving the attention.
Who it affects
Developers and companies choosing which model to route agentic or high-volume workloads through, since token cost compounds quickly at scale. It also affects other Chinese model makers such as MiniMax, whose M3 model DeepSeek-V4-Flash-0731 is said to outrank despite being smaller.
How to use it
The model is available on Hugging Face as a 167GB download and can be accessed through OpenRouter, as in the command llm -m openrouter/deepseek/deepseek-v4-flash-0731 -t pelican -o reasoning_effort high. Pricing is $0.14 per million input tokens and $0.27 per million output tokens; it is not stated in the source whether this is DeepSeek's direct pricing or specific to OpenRouter. Output quality reportedly depends heavily on the reasoning effort setting: the default level gave a weaker result than the "high" setting in informal testing.
How solid is it
The ranking above MiniMax M3 comes from Artificial Analysis, a third-party benchmark tracker, which lends it some independence, but the source gives no numeric scores, only the ranking and a chart reference. The pelican test cited is a single informal, non-standardized check by one reviewer, not a rigorous benchmark, and no release date or architecture detail is provided to corroborate the claims further.
Risks and caveats
"Substantially enhanced agentic capabilities" is not broken down into specifics in the source. The pricing figures lack confirmation of their source (DeepSeek directly versus OpenRouter), and output quality appears sensitive to configuration, meaning default settings may understate or overstate real-world performance depending on how a deployment is tuned.
“this may currently be the best value-per-intelligence model out there”
— Simon Willison