DeepSeek released the final version of DeepSeek-V4-Pro (build 0813) on Wednesday without prior notice, by updating the API pricing page. In a benchmark table, the company points out that the model trails Claude Fable 5, from Anthropic, by an average of 5.3%, but costs about 46 times less.
The change was noticed by the swap of the listed version to deepseek-v4-pro, which now appears as DeepSeek-V4-Pro-0813. The model had been in preview since April, with a price 98% below GPT-5 Pro. On July 31, DeepSeek said the Pro API was "unchanged" and that the official release would come next. The model page on Hugging Face still describes it as a "preview version."
The table released by DeepSeek covers ten agent benchmarks. In eight of them, Fable 5 or another model has the advantage; DeepSeek wins in two. In the nine in which both score, Fable 5's average advantage is 5.3%. Excluding the Humanity's Last Exam test without tools, where the difference reaches 10.6 pp, the average advantage drops to 2.8%.
Public prices show the biggest gap. Claude Fable 5 costs US$ 10 per million input tokens and US$ 50 per million output tokens. The V4 Pro costs US$ 0.435 and US$ 0.87, with cached input at US$ 0.003625. At combined rates, using Fable 5 costs US$ 30, versus US$ 0.65 for V4 Pro — about 46 times more.
Without independent verification
DeepSeek stresses that the results were obtained by the company itself, on infrastructure not yet released. Two of the ten tests, DSBench-FullStack and DSBench-Hard, are internal and have no public leaderboard. No external lab has evaluated version 0813 so far.
Cost-per-task measurements also reinforce the difference. Artificial Analysis estimated US$ 3.15 per task for Fable 5 versus US$ 0.03 for V4-Flash. Hugging Face CEO Clément Delangue calculated a variation of more than US$ 31 per task versus about US$ 0.04.
Anthropic, however, has Claude Opus 5, which beats Fable 5 on most benchmarks for half the price. DeepSeek's move follows the trend of Chinese open-weights labs, such as Kimi K3 and Xiaomi, which approach American models at a fraction of the cost. The V4 Pro weights have an MIT license and are available on Hugging Face for independent verification.
