China's AI company DeepSeek has swapped its flagship model, DeepSeek V4 Pro, to the official release build 0813. Since previously published performance evaluations were based on a preview version, some assessments say the official version’s actual performance still needs external verification.

Blockchain outlet Decrypt reported on Aug. 12 that DeepSeek changed the model name on its API pricing page to DeepSeek-V4-Pro-0813, without a separate announcement or blog post.

DeepSeek V4 Pro has been offered in preview form since April. API pricing remained unchanged with the shift to the official version. It costs $0.435 per 1 million input tokens and $0.87 per 1 million output tokens, while cached input costs $0.003625 per 1 million tokens.

The focus of this update is not price but changes to the model’s internal weights. DeepSeek said when it officially released V4-Flash on July 31 that the V4 Pro API was "unchanged," and it had signalled that an official release would soon follow.

However, the model card on Hugging Face still shows the V4 series as a preview version. External tests will be needed to confirm how much the official 0813 version’s performance differs from earlier evaluations.

In benchmarks released by DeepSeek, the performance gap with leading U.S. AI models was not large. DeepSeek published results for 10 agent benchmarks, and on nine items where both models had scores, Anthropic’s Claude Fable 5 led by an average of 5.3 percent.

DeepSeek led on 2 items. Excluding some items with large gaps, the remaining average performance difference narrows to about 2.8 percent.

The gap in price competitiveness was much wider. Claude Fable 5 costs $10 per 1 million input tokens and $50 per 1 million output tokens. Based on a blended input-output rate, the price is about $30, about 46 times higher than DeepSeek V4 Pro’s blended rate of about $0.65.

When calculated by actual work units, the cost difference could be larger. That is because Claude Fable 5 tends to run inference for longer and generate more output than the DeepSeek model.

Artificial Analysis put the cost per benchmark task at about 3 cents for V4-Flash and about $3.15 for Claude Fable 5. Clement Delangue (클레망 들랑그), Hugging Face's CEO, also mentioned that the per-task cost could widen to about $31 versus $0.04. A per-task cost based on the official V4 Pro 0813 version has not yet been disclosed.

The benchmark results themselves also have limits. The comparison scores were measured by DeepSeek, and it did not disclose detailed testing infrastructure. DeepSeek said in a July notice that it would soon release its in-house testing environment, "DeepSeek Harness minimal mode."

Also among the 10 benchmarks, DSBench-FullStack and DSBench-Hard are internal test sets with no external leaderboard. As a result, some point out that it is difficult to confirm DeepSeek V4 Pro’s performance advantage based only on currently available results.

Even so, the official release is expected to again highlight the price competitiveness of China’s open-weight AI models. As Chinese AI institutes narrow the performance gap with leading U.S. models to single digits while offering much lower prices, some are raising the possibility that the AI model market’s pricing structure itself could change.

DeepSeek is releasing model weights on Hugging Face under the MIT licence, and direct verification by external developers and researchers is expected to follow. Through this, the actual performance of V4 Pro 0813 and differences from the preview version are expected to emerge more concretely.

Competition among Anthropic’s internal products is also a variable. Claude Opus 5 is known to have posted higher performance than Claude Fable 5 on multiple benchmarks while being priced at about half the level.

Ultimately, competition around DeepSeek V4 Pro 0813 goes beyond which model scores higher. As high-priced premium AI models and low-priced open-weight models face off over performance and cost efficiency, the criteria for companies and developers choosing AI models is increasingly likely to shift from top performance to cost relative to performance.