DeepSeek V4 Pro Launches with Agent Boost and Performance Near Anthropic’s Fable 5

On the night of August 12, DeepSeek updated its API documentation to reference the model "DeepSeek‑V4‑Pro‑0813," indicating that the V4 Pro official version is now accessible through the API despite the absence of a formal product announcement.

The new model supports up to 1 million tokens of context, a maximum output of 384 k tokens, and adds features such as JSON output, tool calls, the Responses API, Anthropic‑compatible API, and FIM.

DeepSeek has not yet published an official changelog for V4‑Pro‑0813; the latest official update page still only mentions the July 31 V4 Flash release, which clarified that only the V4 Flash API was upgraded while the V4 Pro API and web/app models remained unchanged.

Community‑circulated benchmark tables, sourced from DeepSeek’s official group and various tech media, claim that V4‑Pro‑0813’s performance on multiple agent‑centric tests now approaches Anthropic’s Claude Code Fable 5, with some items even surpassing it. Agent tests emphasize practical "work‑doing" abilities such as tool invocation, multi‑step tasks, code generation and modification, and sustained execution of complex workflows.

When DeepSeek launched the V4 preview in April, it highlighted that V4 Pro achieved leading open‑source results on Agentic Coding benchmarks and was optimized for agents like Claude Code, OpenClaw, OpenCode, and CodeBuddy, while also supporting 1 M token context and a thinking mode.

By the end of July, the V4 Flash official version further strengthened agent capabilities, publishing results for Terminal Bench 2.1, NL2Repo, DeepSWE, and Toolathlon verified, and stating that its agent abilities had dramatically outperformed the V4‑Pro preview. The current V4‑Pro‑0813 release continues this trajectory, narrowing the gap to Fable 5.

All benchmark data for V4‑Pro‑0813 currently originates from community reports; DeepSeek has not released a formal, comprehensive benchmark suite for the model, so the figures remain unofficial.

Pricing-wise, DeepSeek’s API price page shows that the upgraded V4 Pro retains the low‑price strategy introduced after a previous steep discount. However, the site also warns of an upcoming substantial price increase for the overall DeepSeek API service.

The timing of the V4‑Pro‑0813 launch coincides with the release of Grok 4.6. Grok 4.6 achieved a top Elo score of 1753 in the AI Analysis GDPVal‑AA v2 agent evaluation and scored 61 on the Intelligence Index, while being priced at $2 per million input tokens and $6 per million output tokens, positioning cost as a key differentiator.

These parallel releases illustrate a broader industry shift: as front‑line models close the performance gap, developers increasingly weigh token‑cost efficiency and agent execution capabilities alongside raw model strength, turning the competition from pure “best model” battles to a more nuanced trade‑off between ability and price.

Overall, DeepSeek’s V4 Pro upgrade signals a strategic move away from merely offering cheaper general‑purpose models toward establishing a strong agent‑oriented foundation model, aligning with the same direction taken by Grok 4.6.