DeepSeek V4 Pro 0813 Stable Release: Agent Benchmark Ties with Fable 5

DeepSeek today promoted V4 Pro from preview to stable release, with the model versioned DeepSeek-V4-Pro-0813, now available on the DeepSeek API and OpenRouter.

V4 Pro Parameter Specifications

💡 What You Will Learn

DeepSeek V4 Pro 0813 Stable Release: Agent Benchmark Ties with Fable 5 DeepSeek today promoted V4 Pro from preview to stable release, with the model versioned DeepSeek-V4-Pro-0813, now available on t

DeepSeek V4 Pro 0813 Official Release: Agent Benchmarks Match Fable 5

DeepSeek today moved V4 Pro from preview to official release, with model version DeepSeek-V4-Pro-0813, now available on the DeepSeek API and OpenRouter.

V4 Pro's parameter specs have long been public knowledge — 1.6 trillion total parameters (MoE mixture-of-experts architecture), 4.9 billion activated parameters per token, 1 million token context window, 384K maximum output tokens, and 32 trillion tokens of pretraining data. This architecture carries over from the April preview version with no major changes.

The key shift is in Agent capabilities

DeepSeek's official comparison table lists 9 Agent benchmarks, where V4-Pro-0813 trails Claude Fable 5 by an average of just 2.8%, and actually surpasses it on 2 of them. This is the first time DeepSeek has placed its flagship model on the same comparison table as a top-tier closed-source model like Fable.

Looking at specific core benchmarks

Terminal Bench 2.1 hit 87.9 (terminal operations), up 15.8 points from the preview's 72.1. Cybergym rose to 83.3 (cybersecurity), up 30.6 from the preview. DeepSWE jumped from 12.8 to 62.7 (software engineering), a 49.9-point gain — the largest increase and the core narrative behind this official release. DeepSeek's messaging has also shifted, from "better at answering questions" during the Preview period to "better at getting things done." DeepSWE tests whether an AI can work continuously like an engineer.

Benchmarks are up, but what does the user actually feel?

The improvement you'll notice immediately is that several tasks AI agents previously couldn't handle well are now doable. First, long tasks no longer fall apart midway. Previously, using V4 Pro preview to write a complete project often required re-describing the context after each file; V4 Pro 0813 can read the entire repository at once before starting, maintaining coherence across multi-step tasks. Second, tool calls are more reliable. Previously, when asking the AI to run commands, modify files, then run commands again, it would often get stuck at some step; V4 Pro 0813 can complete tool chains of five or more steps without derailing. Third, it can self-correct when things go wrong. Ask the AI to write code, hit a compilation error — previous versions would often stop and ask "should I continue?" V4 Pro 0813 will read the error message itself, fix the code, recompile, and only report back after completing the full loop.

Taken together, these changes mean AI has taken a major step toward being "a colleague who can deliver." The DeepSWE jump from 12.8 to 62.7 reflects a qualitative leap in this category of engineering scenarios.

Pricing is another card

V4 Pro official pricing (per 1M tokens): cache-hit input ¥0.025, cache-miss input ¥3, output ¥6. By comparison, V4-Flash is ¥0.02, ¥1, and ¥2. V4 Pro is 3-6x more expensive than Flash, but compared to Fable 5 it's a different story entirely. Fable 5's blended rate is $30/million tokens; V4 Pro converts to roughly $0.65 — a 45x price gap. 2.8% performance difference, 45x price difference — that's DeepSeek's math.

Architectural support comes from two elements

V4 Pro uses two attention variants — compressed sparse attention (CSA) and heavily compressed attention (HCA) — reducing per-token inference FLOPs to 27% of the V3.2 generation and KV cache usage to 10% of the original. These two mechanisms compress the inference cost of the 1M-token context window, which is what creates room for the pricing.

The API also got compatibility upgrades this time

V4 Pro supports both OpenAI and Anthropic API formats simultaneously, with no change to how you call it — developers can switch to the latest version using deepseek-v4-pro or deepseek-v4-flash. If you're already building in agent frameworks like Claude Code, OpenClaw, OpenCode, or CodeBuddy, you can simply swap the base_url. Three reasoning tiers are also available: standard mode, high reasoning, and highest reasoning (V4 Pro Max) — the latter appearing for the first time.

As for weights, the rhythm remains API-first, weights-later for open source. V4-Flash went official on July 31 with weights released afterward; V4 Pro should follow the same path.

DeepSeek isn't the only news in the AI world today. Minutes after DeepSeek V4 Pro launched, xAI released Grok 4.6 on its own API at the same $2/$6 pricing, scoring 61 on Artificial Analysis's Intelligence Index — tied with GPT-5.6 Sol Max. Cursor and Grok Build integrated it the same day. Qwen 3.8 Max also went open source today. Three companies releasing new models almost simultaneously — that doesn't happen often.