DeepSeek V4 Pro Launches Unexpectedly: Agent Capabilities Leap Forward, Performance Approaches Fable 5

Chinese AI startup DeepSeek made an unexpected move late at night. On August 12, 2024, the company's official website API documentation was quietly updated, reflecting a new model version labeled "DeepSeek-V4-Pro-0813." This indicates that the official release of V4 Pro has likely already gone live via API.

More notably, benchmark results circulating within the community have drawn significant attention. Multiple tech media outlets citing data from DeepSeek's official group report that DeepSeek-V4-Pro-0813 performs close to Anthropic's Claude Fable 5 across various AI agent tests, surpassing it in some categories. Coupled with its near-simultaneous release alongside xAI's Grok 4.6, the performance and pricing race among cutting-edge AI models has intensified once again.

Judging from changes on DeepSeek's official site, this rollout of the V4 Pro official version has been extremely low-key. Currently, the "Models & Pricing" page on the official API site lists deepseek-v4-pro as corresponding to DeepSeek-V4-Pro-0813, supporting a 1 million token context window, up to 384,000 token output, and features such as JSON Output, Tool Calls, Responses API, Anthropic API compatibility, and FIM (Fill-in-the-Middle).

The official "First API Call" guide page also lists deepseek-v4-pro as a callable model and provides example calls using reasoning mode, indicating the new model has entered active service.

However, DeepSeek has not yet released an official changelog for V4-Pro-0813. The latest announcement on the official updates page still dates back to July 31, regarding the official release of V4 Flash, during which the company explicitly stated that only the V4 Flash API was upgraded, while the V4 Pro API, app, and web models remained unchanged, and announced that the "official version of DeepSeek-V4-Pro will be released soon."

Therefore, strictly speaking, it is more accurate to say that the V4 Pro official version has launched via API, rather than DeepSeek having officially announced a full product release.

The most notable advancement lies in agent-related capabilities. Compared to the earlier V4 Pro preview version, evaluations show significant improvement, particularly in how closely it approaches Fable 5. These enhancements focus on the AI's actual ability to "get work done," including tool calling, multi-step task execution, code writing and editing, and sustained handling of complex workflows—precisely the direction DeepSeek has consistently emphasized with its V4 series.

When DeepSeek launched the V4 preview version in April, it highlighted that V4 Pro achieved leading performance among open-source models in Agentic Coding benchmarks. It also optimized compatibility with mainstream agent products like Claude Code, OpenClaw, OpenCode, and CodeBuddy, while supporting a 1 million token context and reasoning mode.

By late July, the V4 Flash official version further strengthened agent capabilities. DeepSeek published strong results on multiple benchmarks including Terminal Bench 2.1, NL2Repo, DeepSWE, and Toolathlon verified, stating that the official version's agent capabilities had significantly surpassed those of V4-Pro-Preview. Now, with the V4 Pro official version approaching Fable 5 even more closely, it demonstrates that DeepSeek's accumulated post-training and engineering optimizations are translating into stronger real-world task execution abilities.

It should be noted that current benchmark data primarily stems from internal group leaks and community circulation. DeepSeek has not yet officially published the complete benchmark scores for V4-Pro-0813 in its update logs. As such, these scores remain unofficial disclosures and cannot be fully equated with formally announced official benchmarks.

Despite performance upgrades, pricing remains unchanged. According to DeepSeek's current API pricing page, the new V4 Pro version maintains the previous low-price strategy following earlier significant reductions, without any immediate price adjustment accompanying the capability boost.

However, this low pricing may not last long. DeepSeek has clearly announced on its pricing page its plan to raise prices for its entire API service in the near future, noting that the "expected increase will be substantial," with actual adjustments pending further official notice.

In other words, the current state of V4 Pro official version presents "upgraded performance with temporarily unchanged pricing," but this situation may not be sustainable long-term. This release appears strategically timed to complete a model upgrade and product transition before adjusting prices based on compute availability, call volume, and commercialization progress.

Interestingly, the launch timing of DeepSeek V4 Pro-0813 almost directly coincides with that of Grok 4.6.

Just hours before DeepSeek's new model quietly appeared, Grok 4.6 was officially released. According to documentation, Grok 4.6 achieved 1753 Elo on Artificial Analysis' GDPVal-AA v2 agent evaluation, ranking first, and scored 61 points on the Artificial Analysis Intelligence Index.

Pricing is also a key weapon for Grok. Grok 4.6's API is priced at $2 per million input tokens and $6 per million output tokens, emphasizing significantly lower costs compared to other frontier models. Now, DeepSeek V4 Pro-0813 also approaches Fable 5 in multiple tests while maintaining low pricing, causing both models to send nearly identical signals to the market simultaneously: the capability gap between frontier models continues to narrow, and price is becoming the next critical competitive factor.

Past AI model competition centered around "who has the strongest model," but as models like DeepSeek and Grok continuously approach the top tier, the market's comparison criteria are shifting. If performance differences shrink to a certain degree, whether developers are willing to pay multiples for more expensive models is becoming a new competitive focal point.

The significance of this V4 Pro upgrade extends beyond a simple version iteration.

When DeepSeek launched V4 in April, it had already shifted its focus from traditional chat models to Agents, positioning million-token context, code engineering, and tool-calling capabilities as core selling points. Subsequently, the V4 Flash official version further enhanced capabilities like Code Agent and Search Agent.

Now, with the V4 Pro official version closely approaching top-tier closed models like Fable 5, it shows DeepSeek's competitive goal is no longer merely creating a "cheaper general-purpose large model," but attempting to secure a position as the foundational model for AI agents during the transition from conversational AI to practical work.

This aligns closely with Grok 4.6's upgrade direction. Grok 4.6 also emphasizes long-running agents, complex programming, and knowledge work as key areas, delivering services through developer tools.

The near-simultaneous emergence of DeepSeek V4 Pro-0813 and Grok 4.6 reflects a new phase in AI industry competition: the frontier model battlefield is shifting from simple comparisons of parameters, chat, and benchmark performance toward comprehensive competition encompassing agent execution capability, token cost, and developer ecosystem strength.