DeepSeek has launched the manufacturing model of its flagship mannequin, DeepSeek V4 Professional, ending a preview interval that ran almost 4 months. The construct, designated V4 Professional 0813, appeared on August 12, 2026, because the general-availability launch on OpenRouter’s mannequin web page, and DeepSeek’s personal API documentation now lists DeepSeek-V4-Professional-0813 because the mannequin model behind the deepseek-v4-pro endpoint.
The API economics carry over from preview: $0.435 per million enter tokens on a cache miss, $0.003625 per million on a cache hit, and $0.87 per million output tokens, with a one-million-token context window and a most output of 384,000 tokens. DeepSeek’s first-party pricing web page confirms the 0813 model string, the pricing, and a concurrency restrict of 500 for the Professional endpoint in opposition to 2,500 for Flash.
The GA caps a staged rollout DeepSeek has run in public since spring. The corporate previewed the V4 sequence on April 24, 2026, transport open weights for each Professional and Flash underneath the MIT license alongside API entry. On July 31, 2026, it graduated the smaller Flash mannequin to official standing and stated in its change log that the Professional’s official launch “will comply with quickly.” That follow-through is what landed with the 0813 construct.
What the 0813 Construct Sits On
V4 Professional is a mixture-of-experts system with 1.6 trillion complete parameters and 49 billion energetic per token, per the mannequin card on Hugging Face. The structure combines two consideration variants DeepSeek calls Compressed Sparse Consideration and Closely Compressed Consideration, which the corporate says minimize single-token inference compute to 27 p.c and KV cache to 10 p.c of what its V3.2 era wanted on the million-token setting. Each V4 fashions have been pre-trained on greater than 32 trillion tokens, with post-training that grows domain-specific specialists individually after which consolidates them into one mannequin by means of on-policy distillation.
Open weights for the preview builds have been downloadable since April, and the Professional repo logged greater than 1.4 million downloads within the final month on Hugging Face. The cardboard recommends a context window of at the least 384,000 tokens when operating the mannequin’s most reasoning mode domestically.
The Numbers DeepSeek Is Placing Ahead
The headline analysis claims are vendor-reported, from the mannequin card and the corporate’s personal harness runs. At its most reasoning effort, which DeepSeek labels V4-Professional-Max, the cardboard experiences:
- SWE-bench Verified: 80.6 p.c resolved
- Terminal Bench 2.0: 67.9 p.c accuracy
- GPQA Diamond: 90.1 p.c move@1
- Humanity’s Final Examination: 37.7 p.c move@1
- MMLU-Professional: 87.5 p.c
- LiveCodeBench: 93.5 p.c move@1
- Codeforces ranking: 3,206
- MRCR at a million tokens: 83.5 MMR
The cardboard’s personal comparability desk locations these scores in opposition to named frontier methods: V4-Professional-Max trails GPT-5.4 at xHigh effort on Terminal Bench 2.0 (67.9 to 75.1) and Gemini-3.1-Professional on Humanity’s Final Examination (37.7 to 44.4), whereas posting the desk’s high LiveCodeBench and Apex Shortlist scores. On SWE-bench Verified it lands at 80.6, degree with Gemini-3.1-Professional and a fraction behind Claude Opus 4.6 at 80.8. None of those columns has but been replicated by an impartial evaluator for the 0813 construct.
The API exposes three working modes (non-thinking, a excessive reasoning effort, and a max effort that the documentation describes as pushing “the boundary of mannequin reasoning functionality”) and helps the OpenAI ChatCompletions format, the Anthropic Messages format, and DeepSeek’s personal Responses API, with device calling and JSON output on each Professional and Flash.
How DeepSeek Acquired Right here
The Professional GA completes the second half of a launch technique that put the cheaper mannequin in entrance first. When V4-Flash went official on July 31, 2026, DeepSeek printed agent-benchmark outcomes displaying the re-post-trained Flash construct outscoring the V4-Professional-Preview on its inside coding-agent suites, a deliberate transfer that made the small mannequin the default for agent workloads whereas the flagship stayed in preview. The 0813 construct is the flagship’s reply, and it arrives with the Professional endpoint’s context window, output ceiling, and have set unchanged in identify. The change is that the preview label is gone.
The pricing posture bears watching. A discover on DeepSeek’s pricing web page states the corporate plans “a major enhance” in total API pricing within the close to future, with specifics to return by official discover. For now, the listed V4 Professional charges maintain at their preview ranges, and the common worth really paid by means of OpenRouter sits properly beneath the $0.435 checklist worth, which OpenRouter attributes to caching and reductions.
DeepSeek has spent the previous yr constructing outward from the mannequin itself: the V4 sequence is skilled across the agentic workloads (coding assistants, multi-step automation, long-document synthesis) that its April preview announcement stated already drive the corporate’s personal in-house growth. Chinese language open-weight labs have been transport flagship-class fashions on a near-monthly cadence, from Moonshot’s Kimi K3 to MiniMax’s agent-focused M2.7, and V4 Professional’s GA is DeepSeek’s bid to maintain its flagship on the entrance of that pack.
What Ships Subsequent
The open merchandise is new weights. The Hugging Face repositories nonetheless host the April preview builds, and the mannequin card’s obtain desk factors to these artifacts; DeepSeek has not introduced a timeline for publishing 0813 weights or stated whether or not the GA construct differs from preview past post-training. The corporate’s said cadence for the V4 line, per its change log, runs by means of the API first.
On the business aspect, the pricing-page discover commits DeepSeek to an official announcement of its revised worth plan, with the rise making use of throughout API providers. Till that discover lands, deepseek-v4-pro continues to serve the 0813 construct on the charges printed on August 12, 2026 — $0.435 in, $0.87 out, per million tokens, at a million tokens of context.
