Twelve days after the release of the DeepSeek V4 Flash official version, the full-fledged Pro version of the DeepSeek V4 family has finally arrived. On the evening of August 12, the model and pricing page in DeepSeek’s official API documentation was quietly updated. The official DeepSeek V4 Pro version is now designated as “DeepSeek-V4-Pro-0813”, with pricing and specifications announced at the same time. OpenRouter also listed a new model page in sync. This update was very low-key: the banner on DeepSeek’s official homepage still shows the previous Flash official version public beta and the notice that V4-Pro remains temporarily unchanged, but the API documentation clearly lists the new version number and complete price table. The model calling method remains the same; using deepseek-v4-pro automatically points to the latest 0813 version.
Version Confirmation: Flash 0731 and Pro 0813 Dual-Track Finalized
According to API documentationDeepSeek V4 currently consists of two official versions: DeepSeek-V4-Flash-0731 and DeepSeek-V4-Pro-0813. They share exactly the same base URL — https://api.deepseek.com in OpenAI format and https://api.deepseek.com/anthropic in Anthropic format — with the only difference being the model version number. OpenRouter’s model page also lists Pro 0813 as the GA version, with a release date of August 12, 2026.
Key specifications: 1M context, 384K output, dual thinking modes
Pro 0813 and Flash 0731 remain consistent in core specifications, with the main differences being performance positioning and pricing. Both support a 1M context length, with a maximum output of 384K, and support both non-thinking and thinking modes, defaulting to thinking mode. For handling long codebases, large-scale documents, and Agent tasks requiring multi-step continuous execution, 1M input paired with 384K output means a single call can accommodate and generate a considerable amount of content. This has also been one of the selling points that the DeepSeek V4 series has consistently emphasized since its preview version.
In terms of feature support, both official versions cover JSON Output, Tool Calls, the Responses API, and the Anthropic API, and support the conversation prefix continuation beta. The FIM infill beta is only available in non-thinking mode. API access is compatible with both OpenAI and Anthropic formats, making migration costs for existing workflows relatively low.
Price officially announced: Pro is three times the price of Flash, with the official also hinting at a further price increase.
The most notable aspect of this update is pricing. DeepSeek had previously signaled an overall increase in API service prices, and the hike could be substantial. The Pro version’s official pricing is now seen as confirming that direction. According to the RMB-to-USD comparison published in the API documentation, the prices for the two models are as follows:
- Cache hit input: Flash 0.02 yuan per million tokens, 0.0028 US dollars; Pro 0.025 yuan per million tokens, 0.003625 US dollars.
- Cache miss input.Flash: 1 yuan ($0.14) per million tokens; Pro: 3 yuan ($0.435) per million tokens.
- OutputFlash is 2 yuan ($0.28) per million tokens; Pro is 6 yuan ($0.87) per million tokens.
- Concurrency limit：Flash 2500，Pro 500
This shows that for the two main billing items, cache miss input and output, Pro’s prices are both 3 times those of Flash, while its concurrency limit is one-fifth of Flash’s. OpenRouter’s pricing page shows the same figures: Pro 0813 at $0.435 per million input tokens and $0.87 per million output tokens, while Flash 0731 is priced at $0.14 and $0.28, respectively. The actual weighted average input price is lower depending on the cache hit rate. The documentation also notes that there are still plans to raise overall service pricing in the future, with specific details subject to official announcements. From a product positioning standpoint, Flash is clearly aimed at high-frequency, large-scale call scenarios, where its lower price and higher concurrency are better suited for massive parallel tasks. Pro, on the other hand, concentrates resources on higher-spec tasks; although its unit price is higher, it still represents a relatively low call cost compared with models in the same tier.
Benchmark tests: a significant leap over the preview version, with mixed results against top-tier models.
A benchmark comparison table for Agent models leaked from the official WeChat group shows that Pro 0813 delivers quite significant improvements over the two preview versions. The table covers 10 Agent benchmarks, including HLE, Terminal Bench 2.1, NL2Repo, Cybergym, DeepSWE, Toolathlon-Verified, Agents’ Last Exam, AutomationBench, DSBench-FullStack, and DSBench-Hard, with comparisons against DeepSeek’s own Flash 0731, the two preview versions, as well as GLM-5.2, Kimi-K3, Opus-4.8, and Fable 5.
In a comparison of DeepSeek’s own versions, Pro 0813 ranks best across all 10 benchmarks. Among them, DeepSWE jumped from 12.8 in the preview version to 62.7, DSBench-Hard from 31.1 to 67.2, Cybergym from 52.7 to 83.3, and AutomationBench from 12.8 to 31.8—all representing multi-fold growth. Flash 0731 also outperforms the two preview versions, but still trails Pro 0813 in most categories.
Compared with external models, Pro 0813 scored 87.9 on Terminal Bench 2.1, close to Kimi-K3’s 88.3; on Cybergym it scored 83.3, roughly matching Fable 5’s 83.1; and it achieved the highest score on the table with 31.8 on AutomationBench. Fable 5 still leads on benchmarks such as HLE, DeepSWE, Toolathlon-Verified, and DSBench-FullStack, while Opus-4.8 posted the highest scores on NL2Repo and DSBench-Hard, and Kimi-K3 leads with 27.6 on Agents’ Last Exam. Overall, Pro 0813 has reached a level where it trades wins and losses with current top-tier agent models, rather than lagging behind across the board.
The product structure is clearer: Flash focuses on scaling, while Pro focuses on high-end specs.
With the launch of Pro 0813, the product division of DeepSeek V4 has become clearer. Official documentation presents the two models side by side, with fully aligned specifications and features, differentiated only in price and concurrency: Flash is geared toward cost and scale, while Pro targets higher-spec model services.
In terms of ecosystem integration, both models already support extension to Agent and software development scenarios through Tool Calls, the Responses API, and the Anthropic API. For users who require long contexts and high concurrency, the official side also offers cache hit discounts, encouraging the use of context caching to reduce the cost of repeated input. OpenRouter’s data also shows that Pro 0813’s cache hit rate has reached approximately 85.3%, bringing the weighted average input price down to around $0.067 per million tokens (still so cheap it might as well be free).
For developers waiting for the official release, the calling method remains unchanged—just continue using deepseek-v4-pro to get the latest 0813 version, while Flash maps to deepseek-v4-flash. What’s worth watching next is whether the official announcement and updated technical report are published, and whether the open-source weights on Hugging Face are also updated to 0813.
Conclusion
DeepSeek-V4-Pro-0813’s release didn’t come with a grand launch event—it quietly arrived via a file update, but the pricing and benchmark data already speak to its positioning. This flagship model with 1.6T total parameters shows a major improvement over the preview version on Agent benchmarks, trades blows with current top-tier models across multiple tests, while maintaining a relatively low API price. As for how capable it truly is, we may have to wait for more hands-on feedback from the open-source community over the next couple of days.
Source: KOCPC Chinese