DeepSeek released V4-Pro to full production availability on August 13, 2026, at 11 p.m. Beijing time, after four months in preview. The company announced the move via its API documentation and web platform without a separate launch event or press release—a quiet transition that mirrors DeepSeek's operational style: staged rollouts, public preview, then GA with no ceremony.
V4-Pro is a 1.6-trillion-parameter mixture-of-experts model with approximately 49 billion active parameters per token, a 1-million-token context window, and support for output up to 384,000 tokens. DeepSeek shipped the model first as an open-weight release on Hugging Face (MIT-licensed, full commercial rights) on April 24, 2026, then gradually moved to private-weight cloud access. The GA today makes both the open weights and the cloud API fully stable production targets.
The strategic focus for V4-Pro is agent capabilities—tasks where AI systems use tools, execute code, complete multi-step workflows, and handle reasoning across tool outputs without human intervention. This positions V4-Pro not as a general-purpose upgrade to V3.2 but as DeepSeek's answer to a specific question: can an open-weight model deliver agent performance that matters to the companies now racing to deploy agentic systems in production?
Model specifications and access
- Total parameters
- 1.6 trillion
- Active per token
- ~49 billion
- Context window
- 1 million tokens (input); up to 384,000 output
- Availability
- Open weights (MIT, Hugging Face); Cloud API (DeepSeek platform and OpenRouter)
- Training license
- MIT — full commercial rights, no revenue cap, no usage restrictions
DeepSeek released benchmarks specific to agent tasks. V4-Pro scored 87.9 on Terminal Bench 2.1 (a test of whether the model can autonomously resolve real GitHub issues), 62.7 on DeepSWE (a variant of SWE-bench targeting agentic software engineering), and 61.5 on NL2Repo (natural-language-to-repository-action benchmarks measuring whether the model can understand a codebase and execute multi-file changes). These are not general-reasoning benchmarks like MMLU or ARC; they are scaffolded evaluations of whether the model can complete tool-grounded tasks.
V4-Pro's three released agent-task scores
On those agent-focused tests, DeepSeek reports that V4-Pro lands within a rounding error of Claude Opus 4.6: V4-Pro at 87.9 Terminal Bench versus Opus 4.6 at approximately 87.8-88.0 (Claude's own SWE-bench Verified score of 80.6 is a different test, and Anthropic has published that Opus 4.6 scores in a similar range on SWE-bench Verified at 80.8). The convergence on agent-specific tasks does not mean parity on all dimensions—long-horizon reasoning, multi-turn reliability under production load, and tool-use error recovery are still open questions—but it does narrow the claimed gap between closed and open agentic models from "orders of magnitude" to "within measurement noise on these specific benchmarks."
The API pricing carries over from preview without change: $0.003625 per million cached input tokens (a compressed-context cache hit, using DeepSeek's reported 27% compute savings on cached inference), $0.435 per million uncached input tokens, and $0.87 per million output tokens. DeepSeek has publicly flagged that it plans to raise API prices in the near future and expects the increase to be significant—a signal that the model's performance-to-cost ratio is attracting enough traffic to strain margins.
The quiet launch contrasts with the scale of the decision underneath it. Four months into preview, DeepSeek had the data it needed to bet that V4-Pro's agent performance was solid enough for production. That decision reframes what "open-weight AI" can now deliver: not a prototype or a research artifact, but a 1.6-trillion-parameter model shipping with 1-million-token context, agent benchmarks within rounding error of Anthropic's current flagship, and full commercial rights, released under the MIT license.
What remains unresolved is whether agent benchmarks predict production reliability. Terminal Bench and similar tests measure isolated instances where an agent is given a task, tool access, and a reasonable deadline, and asked whether it can complete it. They do not measure what happens when agents run for hours across hundreds of tool calls, hallucinate about tool capabilities, fail to recover from errors, or attempt tasks that were never meant to be autonomous. DeepSeek has not published long-horizon reliability data or failure-mode analyses. OpenAI and Anthropic have been public about running agent systems against real workloads (OpenAI through ChatGPT's browser agent, Anthropic through Claude's early deployments), and both have discovered that benchmarks dramatically underestimate the gap between "works on this test" and "works in production." V4-Pro's GA means the benchmarks are now comparable. Production reliability—whether the model handles tool-use errors, multi-step cascades, and recovery patterns as well as the closed models—is still empirical, and waiting on real deployments.
- DeepSeek moved V4-Pro from preview to production on August 13 2026, 11 p.m. Beijing time, across all access tiers.
- The model is a 1.6T-parameter MoE with 49B active per token, focused on agent tasks: code, tool use, multi-step workflows.
- Benchmarks: 87.9 on Terminal Bench 2.1, 62.7 on DeepSWE, 61.5 on NL2Repo—agent-specific tests, not general reasoning.
- API pricing: $0.003625 per million cached input tokens, $0.435 uncached, $0.87 per million output—DeepSeek has flagged future price increases.
- The GA closes a gap between open and closed agentic models but does not resolve questions of end-to-end reliability and multi-turn tool-use patterns under production load.