DeepSeek has quietly transitioned its deepseek-v4-pro model to the 0813 build, marking the general-availability release of a system that had been running as a preview since April. The Chinese AI lab shipped the finished version of its flagship on Wednesday with no blog post or formal announcement—the only signal was a table cell on the API pricing page, where the model version for "deepseek-v4-pro" now reads DeepSeek-V4-Pro-0813.

Pricing remains unchanged at approximately $0.435 per million input tokens and $0.87 per million output tokens, with cached input at $0.003625. What changed is the underlying model weights. The API pricing page is the sole public indicator of the update.

DeepSeek V4 Pro has been available since April, priced 98% below GPT-5 Pro, but every independent lab that tested it was evaluating a preview build. DeepSeek acknowledged this 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 leading repository for open-source AI projects, still describes the V4 series as "a preview version." This means widely circulated benchmark scores describe a build DeepSeek itself considered unfinished. No one outside the company has independently benchmarked the 0813 release yet.

Benchmark Comparison

DeepSeek published a comparison across 10 agent benchmarks—tests that measure how well models autonomously complete multi-step tasks like tool use, code execution, and data analysis rather than simply answering questions. On the eight where Claude Fable 5 or another competing model holds the advantage, the gaps are modest. Averaging Fable 5's relative lead across all benchmarks yields 5.3%. Excluding Humanity's Last Exam without tools—where DeepSeek scores 42.7 against 53.3, a 10.6% gap that disproportionately affects the average—the remaining rows average just 2.8%. On two of the nine benchmarks where both models are scored, DeepSeek wins outright.

Pricing is public on both sides, and the divergence is stark. Fable 5 costs $10 per million input tokens and $50 per million output tokens. V4 Pro costs $0.435 and $0.87 for the same volumes. On blended rates, that is roughly $30 versus $0.65—approximately 46 times, or 4,600% of the cost. According to Reuters, research firms have identified DeepSeek's model as by far the cheapest among well-known AI models.

Cost per completed task runs even wider, because Fable 5 generates longer outputs and spends more compute per response. Artificial Analysis measured Fable 5 at $3.15 per benchmark task versus 3 cents for V4-Flash—roughly 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 the 0813 build yet.

Anthropic's own product lineup complicates the premium positioning of Fable 5. Claude Opus 5 outscores Fable 5 on most benchmarks at half the price.

DeepSeek conducted these benchmark evaluations itself, on infrastructure it has not 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 for independent verification.

Broader Industry Context

The trajectory is consistent across Chinese open-weights labs, which continue to land within a few percentage points of the American frontier at a fraction of the cost. Kimi K3 beat both Fable 5 and GPT-5.6 Sol on release, and DeepSeek and Xiaomi have been cutting frontier AI costs by 99% while U.S. labs move in the opposite direction. DeepSeek's weights are MIT-licensed and available on Hugging Face, making independent verification readily accessible. The MIT license permits unrestricted commercial use, including modification, redistribution, and integration into proprietary products without royalty obligations.