DeepSeek Ships V4-Pro Out of Preview — a 1.6T-Parameter MoE With Native Anthropic and OpenAI API Compatibility
DeepSeek-V4-Pro 0813 has left preview after nearly four months, landing 80.6% on SWE-bench Verified with a mixture-of-experts architecture that cuts inference compute 73% at million-token context.
DeepSeek-V4-Pro 0813 is out of preview. The model went general-availability on August 12, closing out a nearly four-month preview window that started with the April 24 debut, and DeepSeek made it official in an API changelog posted August 13.
The architecture is a mixture-of-experts system with 1.6 trillion total parameters and 49 billion active per token — smaller active-parameter count than most flagship rivals despite the enormous total. DeepSeek’s headline efficiency claim is its “Compressed Sparse Attention” and “Heavily Compressed Attention” variants, which cut inference compute to 27% and KV cache to just 10% of what V3.2 required at million-token context. That’s the number that matters for anyone running long-document or multi-step agent workloads: cheaper long-context inference without giving up the 1-million-token window.
Benchmarks for the top-tier V4-Pro-Max configuration: 80.6% on SWE-bench Verified, 93.5% pass@1 on LiveCodeBench, 90.1% on GPQA Diamond, and 87.5% on MMLU-Pro. Those numbers put it in the same conversation as this month’s other flagship releases — Gemini 3.7 Flash, Grok 4.6 — on coding and reasoning benchmarks, at DeepSeek’s characteristically aggressive pricing.
Current API pricing sits at $0.435 per million input tokens on a cache miss (dropping to $0.003625 on a cache hit) and $0.87 per million output tokens, with a 384,000-token maximum output and support for three reasoning modes: non-thinking, high reasoning, and max effort. The endpoint speaks OpenAI’s ChatCompletions format, Anthropic’s Messages format, and DeepSeek’s own Responses API natively — a deliberate move to make swapping providers a config change rather than a rewrite. That’s notable timing, landing the same week Stripe agreed to buy AI-gateway startup OpenRouter for the exact same reason: nobody wants to be locked into one model’s API shape anymore.
DeepSeek has also flagged “a significant increase” to overall API pricing coming soon, without specifics — a signal that the aggressive rates that made V3 and R1 famous for undercutting Western labs may not hold much longer as usage scales.
The rollout strategy mirrors what DeepSeek did with V3: ship the smaller, cheaper Flash variant first (July 31) as the default for agent workloads, then bring the full Pro model to GA once it’s proven out. Pro carries a 500-request concurrency limit versus 2,500 for Flash, reinforcing that Flash — not Pro — is meant to be the workhorse for high-volume agentic pipelines, with Pro reserved for tasks that need the extra reasoning headroom.
Weights from the April preview builds remain on Hugging Face; DeepSeek hasn’t yet published the 0813 GA weights separately, so self-hosters are for now running an earlier snapshot than what’s live in the API.