DeepSeek quietly updates V4 Pro API with major coding gains
A silent rollout replaced the existing deepseek-v4-pro endpoint with a version claiming huge jumps in agentic coding benchmarks, all while keeping prices and token limits steady.
Without a formal press release or a published changelog, DeepSeek has rolled out an updated build of its V4 Pro model. The revision operates under the existing deepseek-v4-pro API identifier, meaning current workflows will automatically route to the updated model without manual intervention.
According to reported benchmark figures, the update focuses heavily on agentic coding workflows. On the Terminal Bench 2.1 benchmark, the model scored 87.9 points, up from 72.1 in the earlier V4 Pro Preview. That puts it in direct competition with scores reported for Kimi K3 (88.3) and Fable 5 (88.0), while edging out Opus 4.8 at 85.0. Gains were even steeper in other evaluations: DeepSWE scores climbed from 12.8 to 62.7, and NL2Repo rose from 38.5 to 61.5.
Core operational parameters remain unchanged. The model keeps its 1-million-token context window, supports standard and "thinking" modes, handles tool calling, and connects with the Responses API as well as developer environments such as Codex, Claude Code, and OpenCode. Pricing also holds flat at $0.435 per million input tokens and $0.87 per million output tokens.
For now, model weights remain unreleased, leaving the build (identified internally as V4 Pro 0813) accessible solely through the API. DeepSeek has yet to provide full technical documentation or an official explanation of what architectural changes drove the reported performance leap.
Source: habr
Comments
Help shape the next version: Add context or suggest a correction. AI review can add points toward a rewrite. Reviews and updates may take time; a full meter does not guarantee a new version.