DeepSeek V4-Pro closes the gap with Fable 5, Opus 4.8, and Kimi K3

The jump between this version and its preview is the real story here. DeepSWE, a benchmark built around real software engineering tasks, went from 12.8 to 62.7 — nearly a fivefold increase in a category DeepSeek had previously struggled with. Terminal Bench 2.1 climbed from 72.1 to 87.9, and two other benchmarks improved by similar margins: NL2Repo rose from 38.5 to 61.5, and Cybergym from 52.7 to 83.3, all inside a single release cycle.

Alongside those gains, DeepSeek added a reasoning-depth control with three settings: low for simple tasks, high as the default for everyday agent work, and max for the hardest multi-step problems that need more time and tokens to work through. V4-Pro is built on a 1.6-trillion-parameter mixture-of-experts architecture, the same general design DeepSeek has used across its flagship releases going back to earlier versions.

On Terminal Bench 2.1, V4-Pro’s 87.9 sits ahead of Claude Opus 4.8 (85.0) and GLM-5.2 (81.0), trails Kimi K3 (88.3) by a fraction, and lands just 0.1 point behind Fable 5’s 88.0. DeepSWE follows a similar pattern: V4-Pro’s 62.7 beats Opus 4.8 (58.0) and GLM-5.2 (46.2), falls behind Kimi K3 (67.5), and sits 7.3 points behind Fable 5’s 70.0.

Cybergym is where V4-Pro actually edges ahead of Fable 5, 83.3 to 83.1, while also beating Opus 4.8 (78.3) and Kimi K3 (80.0). On NL2Repo, V4-Pro’s 61.5 clears GLM-5.2 (48.9) but falls short of Opus 4.8’s 69.7; neither Kimi K3 nor Fable 5 have a published score for that particular test.

The release also adds native support for OpenAI’s Responses API, though not the full feature set: function tools and server-side web search both work, but image and file inputs aren’t supported, and the endpoint is stateless, so it doesn’t retain conversation history or support background mode. V4-Pro is live now in DeepSeek’s app, on its website in Expert Mode, and through the API under the model ID deepseek-v4-pro.