DeepSeek V4 Pro Launches: Performance Approaches Fable 5 at Just 2% of the Price

The official release of DeepSeek V4 Pro went live on Wednesday (August 12). Currently, DeepSeek has not issued any tweets—the only indication being a single line change on the API pricing page of its official website: the version number following 'deepseek-v4-pro' has shifted from preview to 'DeepSeek-V4-Pro-0813'.

This signifies that the flagship model, offered in preview since April 2026, has finally been promoted to full production status.

DeepSeek V4 Pro is a Mixture-of-Experts (MoE) model with a total parameter count of 1.6 trillion, activating 49 billion parameters per token. It features a context window of 1 million tokens and can output up to 384,000 tokens—placing it among the top tier of global AI models in scale. Third-party testing has primarily relied on the preview version, and the model has already seen extensive downloads on platforms like Hugging Face.

Pricing for the official release remains consistent with off-peak rates. This rate reflects the permanent price reduction implemented in May 2026, when prices were cut to one-quarter of their original level. Specifically, input costs $0.435 per million tokens, output costs $0.87, and cached input hits cost just $0.003625.

By comparison, Anthropic's Claude Fable 5 charges $10 per million input tokens and $50 for output. Calculated by output token price, the two differ by nearly 58 times, meaning DeepSeek V4 Pro's price is under 2% of Claude Fable 5's.

When measuring cost per individual task, the gap may be even more pronounced. One organization tested completing a benchmark task: using Claude Fable 5 cost $3.15, while DeepSeek V4 Flash required only 3 cents—a 100-fold difference. While comprehensive third-party testing for the DeepSeek-V4-Pro-0813 version is still lacking, given no changes in the API pricing structure, the actual cost disparity is likely still substantial.

Such a dramatic price gap stems largely from architectural design. DeepSeek V4 Pro employs two variants of attention mechanisms: Compressed Sparse Attention and Heavily Compressed Attention.

According to DeepSeek, this architecture reduces single-token inference computation to 27% of V3.2's level and cuts KV cache usage to just 10% of prior versions. For a model with a 1-million-token context window, these savings are considerable. Improved inference efficiency also grants greater flexibility in pricing.

DeepSeek conducted its own benchmark tests, comparing against Claude Fable 5 across 10 agent evaluations. After excluding the most divergent item (Humanity's Last Exam, where DeepSeek V4 Pro trailed by 10.6 percentage points), DeepSeek averaged 2.8% behind across the remaining nine tasks, winning two of them.

However, these results come from DeepSeek's internal testing, using a proprietary evaluation framework not yet publicly disclosed. The internal test sets also lack public leaderboards for verification. Therefore, the real-world performance of DeepSeek-V4-Pro-0813 still awaits confirmation from third-party evaluations. These figures currently only suggest that DeepSeek believes V4 Pro approaches Fable 5 in certain tests—not definitive proof of overall parity.

The DeepSeek V4 Pro API offers varying inference modes, including non-reasoning and different reasoning intensity settings. Higher reasoning modes further enhance the model's inferential capabilities. The API supports both OpenAI and Anthropic interface formats, allowing developers already working within these ecosystems to switch seamlessly without major code modifications. It also supports tool calling and structured JSON output.

With the official launch of DeepSeek V4 Pro, the product lineup within the DeepSeek V4 series has become clearer: the Flash version handles high-frequency, high-concurrency use cases, emphasizing low cost, while the Pro version concentrates more computational power on complex reasoning, long-context encoding, and agent tasks. Some AI development companies have already begun combining both versions.

Nonetheless, DeepSeek's pricing page has already announced an upcoming 'significant price adjustment.' Specific increases and timing remain undisclosed, leaving uncertainty about how long this current low-price window will last.

Meanwhile, data from multiple aggregation platforms shows that by July 2026, DeepSeek's model token call volume ranked among the global AI elite, trailing only Anthropic's models on some leaderboards. DeepSeek's business growth relies heavily on actual API usage by developers. To date, it has not held large media launch events, instead announcing product updates primarily through its official WeChat account and occasionally publishing technical blogs.

It is widely known that DeepSeek's compute reserves are modest compared to leading Western AI firms. According to leaked content from a July investor meeting, Liang Wenfeng mentioned DeepSeek then had approximately 20,000 H100-equivalent GPUs. This figure pales in comparison to overseas giants operating clusters of tens or even hundreds of thousands of GPUs.

Nevertheless, Liang stated the number is dynamic, and DeepSeek continues actively procuring more compute, planning to 'buy as many cards as possible within reasonable pricing.' Simultaneously, reports indicate DeepSeek is planning GW-scale in-house compute infrastructure, with recent job postings for Inner Mongolia data center roles seen as signs of expanded compute deployment.

DeepSeek also collaborates with Huawei. Per the investor meeting leaks, DeepSeek received around 16,000 Ascend 950 chips from Huawei for software ecosystem validation and deployment. Third-party institutions have already completed full-parameter post-training of DeepSeek V4 Pro on domestic Chinese compute clusters (Ascend 910C), indicating domestic hardware is increasingly handling heavier AI model training workloads.

Currently, the DeepSeek V4 series adopts an open-weights model under the MIT license. The timeline for releasing the weights of the DeepSeek-V4-Pro-0813 official version remains subject to official announcements. Following the V4 Flash pattern of 'API first, weights follow,' the market continues to await further news on official weight releases.

Third-party evaluation agencies are now progressively testing the DeepSeek-V4-Pro-0813 version. Once more independent benchmark data becomes available, a clearer picture will emerge regarding the extent of DeepSeek V4 Pro's performance improvements and its actual gap relative to top-tier models like Claude Fable 5.