China’s DeepSeek Upgrades V4 Professional: Claude Fable Is Solely 5% Higher at 4,500% the Worth – Decrypt

DeepSeek shipped the completed model of its flagship on Wednesday with no weblog publish and no announcement. The inform was a desk cell: the mannequin model listed for “deepseek-v4-pro” on the API pricing web page now reads DeepSeek-V4-Professional-0813.

The associated fee to make use of this mannequin is round $0.435 and $0.87 per million tokens (the fundamental unit of knowledge a mannequin can deal with) of enter and output. So the pricing construction stays the identical, however what modified is the weights beneath.

Deepseek V4 Professional has been within the wild since April, priced 98% under GPT-5 Professional, and each unbiased lab that examined it was testing a preview. DeepSeek stated so itself on July 31, when it pushed V4-Flash to common availability and famous that the Professional API was “unchanged” with the official launch to “observe quickly.” The mannequin card on Hugging Face, the go-to repository for open-source AI initiatives, nonetheless describes the V4 collection as “a preview model.”

So the extensively circulated scores describe a construct DeepSeek thought of unfinished. No one exterior the corporate has independently benchmarked 0813 but.

What DeepSeek’s desk claims

The corporate printed a comparability throughout 10 agent benchmarks. On the eight the place Fable 5 or one other mannequin has the benefit, the gaps are small.

Common Fable 5’s relative lead throughout the benchmarks and also you get 5.3%. Strip out Humanity’s Final Examination with out instruments—the place DeepSeek scores 42.7 towards 53.3, a ten.6% hole that skews every thing—and the remaining rows common 2.8%.

Pricing is public on either side, and that is the place the comparability stops being shut. Fable 5 runs $10 per million enter tokens and $50 per million output. V4 Professional runs $0.435 and $0.87, with cached enter at $0.003625. On blended charges that’s $30 towards $0.65—roughly 46 occasions, or 4,600% of the fee. That’s the form of unfold that issues whenever you’re working a enterprise and utilizing AI instruments at scale.

Price per accomplished activity runs wider, as a result of Fable 5 thinks longer and writes extra. Synthetic Evaluation measured it at $3.15 per benchmark activity towards 3 cents for V4-Flash, about 105 occasions cheaper. Hugging Face CEO Clément Delangue put the unfold at over $31 per activity towards roughly $0.04. No per-task determine exists for 0813 but.

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

DeepSeek scored these itself, on infrastructure it hasn’t launched. Its July be aware specified DeepSeek Harness minimal mode “to be launched quickly,” working at max effort with excessive creativity. Two of the ten benchmarks, DSBench-FullStack and DSBench-Onerous, are inside take a look at units with no public leaderboard to examine them towards.

The course of journey is constant, although. Chinese language open-weights labs preserve touchdown inside a number of factors of the American frontier at a fraction of the worth—Kimi K3 beat Fable 5 and GPT-5.6 Sol on launch, and DeepSeek and Xiaomi have been slicing frontier prices by 99% whereas U.S. labs go the opposite means. DeepSeek’s weights are MIT-licensed and on Hugging Face, so unbiased verification is a obtain away.