On 31 July, DeepSeek pushed out a new build of its open-weight V4 Flash model, called V4-Flash-0731, and the headline is almost rude. On the company’s own numbers, this free download now beats DeepSeek’s more expensive flagship, V4-Pro, on all nine agent and coding tests they published. Same size, same price, same speed. Just retrained. And it costs $0.14 per million input tokens.
Let us translate that, because the frontier labs would rather you did not do the maths.
What actually changed
Nothing you can see in the architecture. V4 Flash is still a Mixture-of-Experts model, meaning it holds 284 billion parameters but only wakes up about 13 billion of them for any given word (that is what keeps it cheap to run). It still handles a million tokens of context, roughly a fat novel plus your entire codebase.
What DeepSeek altered is the post-training, the finishing school phase that teaches a model how to use what it already knows: when to call a tool, how to plan a multi-step job, how not to give up halfway through a task. On DeepSeek’s Terminal-Bench 2.1 test, which measures agent work in a real terminal, the new build scored 82.7, up from 61.8 for the earlier Flash preview and ahead of the Pro model’s 72.1. On DeepSWE, a coding-agent benchmark, it went from a miserable 7.3 to 54.4. That is not an upgrade, that is a different animal wearing the same coat.
Why the little guy should care
Because pricing is the whole story. A model this capable at $0.14 in and $0.28 out per million tokens is roughly a tenth to a twentieth of what the big American labs charge for comparable agent work. The weights are on Hugging Face, so if you have the hardware you can run it yourself and pay nothing per token at all. Developer Simon Willison, who tracks this stuff obsessively, flagged the same odd combination: a large open checkpoint, aggressive hosted pricing, and scores above much bigger open models.
For a freelancer building coding agents, or a small shop automating support tickets, that gap is the difference between a viable product and a monthly bill that eats the margin.
The part they hope you skip
Here is the honest bit. Every single benchmark number above comes from DeepSeek, run by DeepSeek, on DeepSeek’s own test harness. As of the 31 July release there was no independent reproduction from a neutral third party. Vendor benchmarks are marketing until someone else confirms them, and the history of this industry is littered with launch-day figures that shrank under scrutiny.
So treat the leaderboard claims as a promissory note, not a receipt. The sensible move is to run V4-Flash-0731 on your own workload, the boring real one, and see if it holds up. The good news is that testing it costs almost nothing, which is rather the point.
The bigger signal is what it does to everyone else’s price list. When a free retrain can credibly claim flagship performance, the labs charging premium rates for the same job are the ones who now have to explain themselves.
Did you know: a Mixture-of-Experts model can have hundreds of billions of parameters on paper while only using a small fraction for each token, which is how a “284 billion parameter” model can cost less to run than a dense model a quarter its size.
Related on Top Tool Stack: Moonshot’s Kimi K3 open weights · Grok’s 8-cent voice model