Foreign media report that DeepSeek has switched its flagship model, deepseek-v4-pro, to build 0813, ending its preview status since April. The company did not release a blog post or announcement; the change first appeared in the version identifier on the API pricing page.
The official version has quietly launched.
DeepSeek’s pricing remains unchanged, with input and output still priced at $0.435 and $0.87 per million tokens, respectively. The change lies in the underlying weight updates, meaning external developers are now calling the official released version rather than the previous preview version.
Comparison with Claude Fable 5
DeepSeek compared its V4 Pro with Claude Fable 5 in its own benchmark table. Across 10 agent benchmarks, Fable 5 led by an average of 5.3%; however, DeepSeek outperformed in two of those tests.
- The input price for Fable 5 is $10 per million tokens, and the output price is $50.
- Based on the hybrid rate, the two parties differ by approximately 46 times.
- DeepSeek also lists the cached input price at $0.003625.
The cost gap remains a priority.
The article points out that the real differentiator is usage cost: Fable 5 has a higher cost per task, while DeepSeek’s V4 Pro continues to maintain a low-price strategy. DeepSeek also emphasizes that its weights are already open on Hugging Face, making external verification straightforward.
Additional information: The article notes that DeepSeek’s recent comparisons are still based on internal testing, with two of the test sets consisting of proprietary data that cannot be independently verified by external parties at this time.