China’s DeepSeek Upgrades V4 Pro: Claude Fable Is Only 5% Better at 4,500% the Price

DeepSeek shipped the finished version of its flagship on Wednesday with no blog post and no announcement. The tell was a table cell: the model version listed for “deepseek-v4-pro” on the API pricing page now reads DeepSeek-V4-Pro-0813.

The cost to use this model is around $0.435 and $0.87 per million tokens (the basic unit of information a model can handle) of input and output. So the pricing structure remains the same, but what changed is the weights underneath.

Deepseek V4 Pro has been in the wild since April, priced 98% below GPT-5 Pro, and every independent lab that tested it was testing a preview. DeepSeek said so itself on July 31, when it pushed V4-Flash to general availability and noted that the Pro API was “unchanged” with the official release to “follow soon.” The model card on Hugging Face, the go-to repository for open-source AI projects, still describes the V4 series as “a preview version.”

So the widely circulated scores describe a build DeepSeek considered unfinished. Nobody outside the company has independently benchmarked 0813 yet.

What DeepSeek’s table claims

The company published a comparison across 10 agent benchmarks. On the eight where Fable 5 or another model has the advantage, the gaps are small.

Average Fable 5’s relative lead across the benchmarks and you get 5.3%. Strip out Humanity’s Last Exam without tools—where DeepSeek scores 42.7 against 53.3, a 10.6% gap that skews everything—and the remaining rows average 2.8%.

Pricing is public on both sides, and this is where the comparison stops being close. Fable 5 runs $10 per million input tokens and $50 per million output. V4 Pro runs $0.435 and $0.87, with cached input at $0.003625. On blended rates that’s $30 against $0.65—roughly 46 times, or 4,600% of the cost. That’s the kind of spread that matters when you’re running a business and using AI tools at scale.

Cost per completed task runs wider, because Fable 5 thinks longer and writes more. Artificial Analysis measured it at $3.15 per benchmark task against 3 cents for V4-Flash, about 105 times cheaper. Hugging Face CEO Clément Delangue put the spread at over $31 per task against roughly $0.04. No per-task figure exists for 0813 yet.

Anthropic’s own lineup complicates the premium. Claude Opus 5 outscores Fable 5 on most benchmarks at half the price.

DeepSeek scored these itself, on infrastructure it hasn’t released. Its July note specified DeepSeek Harness minimal mode “to be released soon,” running at max effort with high creativity. Two of the ten benchmarks, DSBench-FullStack and DSBench-Hard, are internal test sets with no public leaderboard to check them against.

The direction of travel is consistent, though. Chinese open-weights labs keep landing within a few points of the American frontier at a fraction of the price—Kimi K3 beat Fable 5 and GPT-5.6 Sol on release, and DeepSeek and Xiaomi have been cutting frontier costs by 99% while U.S. labs go the other way. DeepSeek’s weights are MIT-licensed and on Hugging Face, so independent verification is a download away.