DeepSeek quietly pushed the finished version of its flagship V4 Pro model on Wednesday, without a blog post or a formal launch announcement. The clearest sign was a small change on its API pricing page: the version listed for “deepseek-v4-pro” now shows DeepSeek-V4-Pro-0813.
The price did not move. Input and output still cost about $0.435 and $0.87 per million tokens, the basic unit of information an AI model processes. What changed, according to the article, was the model underneath those rates.
That detail matters because DeepSeek V4 Pro has been available since April, and it was already priced 98% below GPT-5 Pro. Yet independent labs that tested it were working with a preview build, not the finished release. DeepSeek said as much on July 31, when it made V4-Flash generally available and noted that the Pro API was “unchanged,” with the official release set to “follow soon.” The model card on Hugging Face still describes the V4 line as “a preview version.”
So the benchmark numbers widely shared so far reflect a build DeepSeek itself did not consider complete. No independent party outside the company has benchmarked the 0813 version yet.
What DeepSeek says the new model can do
DeepSeek published a comparison covering 10 agent benchmarks. On the eight tests where Claude Fable 5 or another competing model comes out ahead, the gaps are described as narrow.
Using the figures cited in the article, Fable 5’s average relative lead across the benchmark set comes to 5.3%. Remove Humanity’s Last Exam without tools, where DeepSeek scored 42.7 and Fable 5 scored 53.3, a 10.6% difference, and the average gap across the remaining rows drops to 2.8%.
The pricing gap is where the comparison widens
Both sides publish their pricing, and this is where the contest stops looking close. Fable 5 is priced at $10 per million input tokens and $50 per million output tokens. V4 Pro is listed at $0.435 and $0.87, with cached input priced at $0.003625.
On blended rates, that works out to about $30 for Fable 5 against $0.65 for V4 Pro. In other words, Fable 5 costs roughly 46 times as much, or about 4,600% of V4 Pro’s cost. For companies running AI systems at scale, that spread can outweigh a single-digit benchmark lead.
The gap grows on a per-task basis because Fable 5 spends longer thinking and generates more output. Artificial Analysis measured Fable 5 at $3.15 per benchmark task, versus $0.03 for V4-Flash, making the latter about 105 times cheaper. Hugging Face CEO Clément Delangue put the spread at more than $31 per task against roughly $0.04 on the other side. There is still no per-task figure for the 0813 build.
Anthropic’s own stack complicates the premium
The article also points to Anthropic’s own lineup. Claude Opus 5 reportedly scores above Fable 5 on most benchmarks while costing half as much, which makes the premium attached to Fable 5 harder to justify on price alone.
Independent verification is still missing
DeepSeek generated these scores itself, using infrastructure it has not yet released. Its July note said DeepSeek Harness minimal mode was “to be released soon,” and that the model was run at max effort with high creativity.
Two of the 10 benchmark suites, DSBench-FullStack and DSBench-Hard, are internal test sets with no public leaderboard, so outside researchers cannot cross-check those results today.
A broader pricing trend remains in place
Even without an outside benchmark of 0813, the direction described in the article is consistent: Chinese open-weight labs are getting within a few points of leading U.S. models while charging a fraction of the price. The piece cites Kimi K3, which it says beat Fable 5 and GPT-5.6 Sol at launch, and says DeepSeek and Xiaomi have been pushing frontier-model costs down by 99% while U.S. labs move in the opposite direction.
DeepSeek’s weights are released under an MIT license and are available on Hugging Face, which means independent verification is close at hand for anyone willing to download and test them.