DeepSeek has quietly launched the finalized version of its V4 Pro model, identified as DeepSeek-V4-Pro-0813, updating its API pricing page without a formal announcement. This new iteration retains the existing pricing structure while incorporating updated model weights. The cost-effectiveness of this release is a key highlight, with DeepSeek V4 Pro priced at approximately $0.435 per million input tokens and $0.87 per million output tokens. This represents a substantial cost advantage compared to competitors like Fable 5, which is priced at $10 per million input tokens and $50 per million output tokens.
Previously, a preview version of DeepSeek V4 Pro had been available since April. The model card on Hugging Face continues to describe the V4 series as a preview version, suggesting that performance scores circulated previously may have been based on this earlier, unfinished build. Crucially, no independent benchmarks of the new 0813 version have been conducted by entities outside of DeepSeek.
DeepSeek’s internal comparisons indicate that the new V4 Pro model performs comparably to or slightly better than Fable 5 on several agent benchmarks. However, these internal results are accompanied by caveats. Two of the ten benchmarks utilized, DSBench-FullStack and DSBench-Hard, are internal test sets lacking public leaderboards for verification. Additionally, the Humanity’s Last Exam benchmark showed a more significant gap, with DeepSeek scoring 42.7 against Fable 5’s 53.3. When this specific benchmark is excluded, the average relative lead for Fable 5 across the remaining benchmarks narrows to 2.8%.
The cost disparity between DeepSeek V4 Pro and Fable 5 is considerable. On blended rates, DeepSeek V4 Pro costs approximately $0.65 per million tokens, whereas Fable 5 is around $30 per million tokens. This approximately 46-fold difference, or a 4,600% cost advantage for DeepSeek, is a significant consideration for businesses employing AI at scale. For context, Artificial Analysis measured the cost per completed task at approximately 3 cents for DeepSeek’s V4-Flash model compared to $3.15 for Fable 5, a difference of about 105 times. Hugging Face CEO Clément Delangue reportedly estimated the cost per task at around $0.04 for DeepSeek versus over $31 for Fable 5. While a specific per-task figure for the new V4 Pro-0813 is not yet available, the trend points towards a continued emphasis on cost-efficiency.
The release of DeepSeek V4 Pro aligns with a broader trend observed among Chinese open-weights AI labs. These entities, including DeepSeek and Xiaomi, are reportedly reducing frontier AI costs by up to 99%, contrasting with the trajectory of U.S. labs. Models such as Kimi K3 have also demonstrated performance competitive with leading U.S. models at a substantially lower price point.
DeepSeek’s model weights are MIT-licensed and accessible on Hugging Face, meaning independent performance verification is technically feasible once researchers conduct their own benchmarks. Until such independent validation is available, the significant cost advantage of DeepSeek V4 Pro remains its most verifiable and impactful feature, while its performance claims await independent confirmation.
Why This Matters
The materials describe a narrow update: DeepSeek quietly updated its API pricing page to reflect the release of the finished version of its DeepSeek V4 Pro model, now designated as DeepSeek-V4-Pro-0813. Nobody outside the company has independently benchmarked the DeepSeek-V4-Pro-0813 version yet.
Broader Context
Source materials place the factual news in this context: DeepSeek shipped the finished version of its flagship on Wednesday with no blog post and no announcement.
Leave a Reply