DeepSeek shipped the general-availability build of V4-Pro-0813 on August 13, an MIT-licensed 1.6-trillion-parameter Mixture-of-Experts model with 49 billion active parameters per token and a 1.0M-token context window, per Artificial Analysis. The same release note announced that starting 16:00 UTC on Sunday, August 16, flat API pricing goes away in favor of peak and off-peak rates that Nikkei Asia characterizes as "nearly five times higher" than the numbers DeepSeek had previously advertised.

Two things happened at once, in other words: an open-weights flagship graduated out of its April preview, and the pricing page moved.

The company's own benchmarks, reproduced by VentureBeat from DeepSeek's release table, are what you'd expect from a vendor claiming agent-tier maturity: 87.9 on Terminal-Bench 2.1, 74.1 on Toolathlon-Verified, 71.1 on DSBench-FullStack, and 67.2 on DSBench-Hard. For the public Code Agent evaluations, DeepSeek tested V4-Pro-0813 using its own harness in "minimal mode," a detail worth remembering when the numbers get quoted downstream. Fable 5 posts higher on the same in-house table (77.9 Toolathlon-Verified, 77.2 DSBench-FullStack).

Independent measurement is less flattering but still substantial. Artificial Analysis puts V4-Pro-0813 at 53 on its Intelligence Index v4.1.1 in reasoning-max-effort configuration, against a median of 27 for similarly sized open-weights peers. SCMP reports the model trails OpenAI's GPT-5.6-series Terra by roughly four points on the same index. Throughput lands at 76.8 tokens per second with a 1.80-second time-to-first-token. Current API pricing, before Sunday's cutover, is $0.43 per million input tokens and $0.87 per million output.

The accompanying release is arguably the more interesting artifact. DeepSeek Harness v0.1, MIT-licensed, is positioned as an open runtime for agentic workflows, a direct answer to Anthropic's Claude Code and OpenAI's Codex. DeepSeek's own note describes V4-Pro-0813 as offering "significantly enhanced agent capabilities and support for the Responses API and Codex integration." The framing matters: DeepSeek isn't just shipping weights, it's shipping the scaffolding around them, and licensing the scaffolding permissively. Alongside the flagship sits V4-Flash, a 284-billion-parameter, 13-billion-active sibling for cheaper inference.

The pricing whiplash tells its own story. A version of DeepSeek's statement was pulled from the company's site Thursday afternoon, per SCMP, which suggests the messaging around the increase is still being tuned. Zhipu's GLM-5.2 shipped in June into the same open-weights lane, and the compression of that lane is now visible in the shape of DeepSeek's decisions: give away the harness, charge for the tokens, and hope nobody notices the two moves are financially connected.

Sources

https://wtaq.com/2026/08/13/deepseek-releases-official-v4-pro-model-as-it-steps-up-expansion/
https://www.scmp.com/tech/big-tech/article/3363895/deepseeks-updated-v4-pro-ai-model-struggles-benchmarks-shines-cybersecurity
https://asia.nikkei.com/business/technology/artificial-intelligence/deepseek-releases-official-v4-pro-model-with-sharply-higher-user-prices
https://venturebeat.com/technology/deepseek-harness-launches-as-open-source-rival-to-claude-code-alongside-v4-pro-on-api-with-higher-prices
https://artificialanalysis.ai/models/deepseek-v4-pro