Simon Willison frames DeepSeek-V4-Flash-0731 as a value-per-intelligence jump

This adds a broader value-per-intelligence read to the earlier Arena price-performance note, and points at reasoning-effort settings as a practical quality lever.

Logged at IST: 2026-08-01 12:48 IST

What it is: Simon Willison's link-blog note on deepseek-ai/DeepSeek-V4-Flash-0731, pointing to the Hugging Face release and Artificial Analysis' pricing/intelligence view.

Gist: Simon highlights DeepSeek-V4-Flash-0731 as the latest V4-family release with “substantially enhanced agentic capabilities.” The model card says it is a 304B-parameter release, about 167GB on Hugging Face, that outperforms the V4-Pro preview on listed agent/code benchmarks despite a much smaller activated-parameter count.

The important framing is value, not just model size. Simon notes that Artificial Analysis ranks it ahead of MiniMax M3, a larger 428B model, and that pricing around $0.14/M input and $0.27/M output may make it the best value-per-intelligence model currently available.

There is also a practical harness detail. Simon tried his standard pelican prompt through OpenRouter and got a weak result at the default reasoning level, then a much better one after setting reasoning_effort=high. That is a useful reminder that cheap agent/coding models may need the right reasoning-effort configuration before judging their quality.

Newsletter angle: Pairs well with the Arena.ai note. Together they say: DeepSeek is showing up not only as a cheap frontend-code arena outlier, but as a broader price/intelligence story with enough API compatibility and reasoning-control surface to matter in real coding-agent setups.