News · High impactBack to News

DeepSeek Launches the Official V4-Pro Release: Benchmarks Beat Preview Across the Board

The official DeepSeek V4-Pro release is now live as DeepSeek-V4-Pro-0813, with the existing API model name and calling method unchanged. DeepSeek's results show the official release beating Preview across ten Agent-related evaluations and joining the leading group on tasks such as Cybergym and AutomationBench. Production workloads should still be retested internally.

DeepSeek Quietly Rolls Out the Official V4-Pro Release

As of 2026-08-13, the official release of DeepSeek V4-Pro is live under the version number DeepSeek-V4-Pro-0813.

The calling method is unchanged. Developers still use deepseek-v4-pro, and the base URLs for the OpenAI and Anthropic formats remain the same. V4-Pro defaults to thinking mode and can switch to non-thinking mode.

The official release arrived quietly. DeepSeek has not published a separate news post or change-log entry, but the pricing page now points to 0813 and the accompanying benchmark materials explicitly label it as the official V4-Pro release. API users do not need to change the model name; the service now routes them to the official release.

The Official Release Focuses on Agent Performance

DeepSeek's comparison table shows the official V4-Pro release outperforming Preview across all ten listed Agent-related benchmarks.

The official-release results also include V4-Flash 0731, GLM-5.2, Kimi-K3, Opus-4.8, and Fable 5, making the comparison more representative of the current field than the older lineup used at the Preview launch.

BenchmarkV4-Pro Official 0813V4-Flash 0731V4-Pro PreviewV4-Flash PreviewGLM-5.2Kimi-K3Opus-4.8Fable 5 (with fallback)
HLE (without / with tools)42.7 / 60.037.8 / 51.537.7 / 48.234.8 / 45.140.5 / 54.743.5 / 56.049.8 / 57.953.3 / 63.0
Terminal Bench 2.187.982.772.161.881.088.385.088.0
NL2Repo61.554.238.539.448.969.7
Cybergym83.376.752.738.780.078.383.1
DeepSWE62.754.412.87.346.267.558.070.0
Toolathlon-Verified74.170.355.949.759.976.576.277.9
Agents' Last Exam25.725.216.515.823.827.625.7
AutomationBench (Public)31.825.112.810.812.930.827.229.1
DSBench-FullStack71.168.741.837.061.873.771.677.2
DSBench-Hard67.259.631.125.854.563.071.768.3

Compared with V4-Pro Preview, the official release rises from 72.1 to 87.9 on Terminal Bench 2.1 and from 12.8 to 62.7 on DeepSWE. Cybergym improves from 52.7 to 83.3, while AutomationBench moves from 12.8 to 31.8. The largest gains cluster around software engineering, terminal operation, and automation-agent tasks.

The Official Release Reaches the Leading Group on Several Benchmarks

The official V4-Pro release has the highest reported scores on Cybergym and AutomationBench, while the leading models continue to trade wins elsewhere.

Its 87.9 on Terminal Bench 2.1 is close to Kimi-K3 at 88.3 and Fable 5 at 88.0, while exceeding Opus-4.8 at 85.0. On DeepSWE, 62.7 beats Opus-4.8 and GLM-5.2 but trails Kimi-K3 and Fable 5. Its 67.2 on DSBench-Hard is above Kimi-K3 and GLM-5.2 but below Opus-4.8 at 71.7.

HLE is mixed. The official V4-Pro release scores 42.7 without tools, below Kimi-K3, Opus-4.8, and Fable 5. With tools, it reaches 60.0, ahead of Kimi-K3 and Opus-4.8 and behind only Fable 5 at 63.0.

The Official Release Keeps the Existing Model Name and Interfaces

The official V4-Pro release keeps the existing integration path, with a 1M-token context, a 384K maximum output, and a current output price of $0.87 per million tokens.

The official documentation lists support for JSON output, tool calls, the Responses API, the Anthropic API, Chat Prefix Completion, and FIM Completion, with FIM limited to non-thinking mode. Cache-hit input costs $0.003625 per million tokens, cache-miss input costs $0.435, and the concurrency limit is 500.

DeepSeek warns that overall API pricing may rise soon. Teams already using deepseek-v4-pro do not need to change the model name, but they should rerun codebase tasks, tool-call regressions, long-context stability checks, and cost tests to confirm that the server-side version switch has not altered critical workflows.

Next step

Keep tracking DeepSeek

Continue along the same topic.

Open entity record