DeepSeek has moved its flagship large language model, DeepSeek V4 Pro, out of preview and into general availability, closing a roughly four-month testing period with a build the Chinese AI lab labeled 0813, after its release date.
Unlike DeepSeek’s earlier launches, the update arrived with little fanfare. According to the South China Morning Post, the company briefly posted a note on its website describing “significantly enhanced agent capabilities” before quietly removing the statement a day later. DeepSeek’s own model card, published on Hugging Face, confirms the build supersedes the preview version and adds three selectable reasoning-effort levels — low, high and max — that control how much the model deliberates before answering.
Sharp gains on coding, mixed on reasoning
The new build’s agentic and coding scores jumped sharply from the preview version, according to the model card: DeepSWE rose from 12.8 to 62.7, CyberGym climbed from 52.7 to 83.3, and Terminal-Bench 2.1 went from 72.1 to 87.9.
On broader reasoning benchmarks, the results were more muted. Artificial Analysis’s Intelligence Index gave V4 Pro 0813 a score of 53 — level with Zhipu AI’s GLM-5.2, but four points behind OpenAI’s GPT-5.6 and seven behind Moonshot AI’s Kimi K3. The Vals Index placed it 12th overall, trailing OpenAI’s GPT-5.5 and Anthropic’s Claude Opus 5. The Post reported the model struggled with tasks in sandboxed terminal environments and with building complex financial spreadsheets, while researchers said it performed comparatively well on niche cybersecurity tasks.
Pricing holds, for now
V4 Pro 0813 keeps preview-era pricing — $0.435 per million input tokens on a cache miss and $0.87 per million output tokens — even as DeepSeek has signaled a “significant” price increase is coming, without saying when. The model remains open-weight under an MIT license, continuing the pattern DeepSeek set with V4 Flash, which outperformed the larger Pro preview on several agent benchmarks when it shipped in July.
The quiet rollout and mixed scorecard drew a subdued developer reaction, with some early testers saying they expected more from a lab known for outsized value relative to price. The release keeps DeepSeek in a crowded field of AI agent-focused model launches this month, alongside SpaceXAI’s Grok 4.6 and Meta’s Muse Glimmer.