DeepSeek V4 Pro 0813 Arrives With Strong Agent Scores, Beating Opus 4.8 on Several Benchmarks
DeepSeek V4 Pro 0813 delivers large agent and coding gains, beats Opus 4.8 on several tests, and keeps the same API name, 1M context window, and 384K maximum output.
DeepSeek quietly rolled out the official DeepSeek-V4-Pro-0813 model at about 11 p.m. Beijing time on August 13, 2026. There was no long launch event or elaborate countdown. The clearest confirmation appeared in DeepSeek's own API documentation, where the model version changed to DeepSeek-V4-Pro-0813.
Existing API users do not need a new model ID. Requests sent to deepseek-v4-pro now reach the 0813 build, so current integrations can pick up the release without changing the model parameter.
The biggest gains are in agent work
The benchmark table shared with the release shows V4 Pro 0813 ahead of V4 Flash 0731 and V4 Pro Preview on every listed agent benchmark. The jump over Pro Preview is especially large on coding tasks. Terminal Bench 2.1 rises from 72.1 to 87.9, DeepSWE moves from 12.8 to 62.7, and DSBench-FullStack improves from 41.8 to 71.1.
The comparisons with closed models are strong, but they need a careful reading. V4 Pro 0813 beats Opus 4.8 on Terminal Bench 2.1, Cybergym, DeepSWE, and AutomationBench. It ties Opus on Agents' Last Exam. Opus remains ahead on HLE, NL2Repo, Toolathlon-Verified, DSBench-FullStack, and DSBench-Hard.
Fable 5 also keeps the lead on most shared tests. DeepSeek edges it on Cybergym and AutomationBench, while Terminal Bench 2.1 is almost even at 87.9 versus 88.0. That is more useful than a blanket claim that one model simply beats another. The result depends on the task.
These are vendor-reported scores. Agent benchmarks are sensitive to the surrounding harness, prompts, effort settings, and tool environment. The table is good evidence that 0813 is a substantial coding update, but it is not a controlled independent ranking.
The API contract stays familiar
The official model page lists the same 1-million-token context window and a maximum output length of 384K tokens. V4 Pro supports thinking and non-thinking modes, JSON output, tool calls, the Responses API, the Anthropic API, chat prefix completion, and FIM completion in non-thinking mode.
That makes the model usable in both OpenAI-style and Anthropic-style developer stacks. DeepSeek's current integration guides cover Codex through the Responses API and tools such as Claude Code, OpenCode, and OpenClaw. For teams already using one of those routes, the 0813 update is mostly a backend swap rather than a migration project.
Pricing has not changed yet
At launch, DeepSeek's pricing page still lists V4 Pro at $0.003625 per million cached input tokens, $0.435 per million uncached input tokens, and $0.87 per million output tokens.
There is an important footnote: DeepSeek says it plans a significant API price increase in the near future, with the final schedule to follow in a separate notice. The current price is real, but developers should not treat it as permanent.
What did not ship
The long-rumored first-party DeepSeek Harness did not arrive with the model. DeepSeek now documents more coding-agent integrations and both V4 models support the Responses API, but there is still no public DeepSeek Harness download or official repository.
For now, V4 Pro 0813 is the release that matters. It brings a clear step up in agent and coding performance, preserves the existing API call pattern, and gives developers a much stronger Pro model without forcing an integration rewrite.
Sources
Frequently asked
Does DeepSeek V4 Pro 0813 beat Opus 4.8?
It depends on the test. In the vendor-reported table, V4 Pro 0813 leads Opus 4.8 on Terminal Bench 2.1, Cybergym, DeepSWE, and AutomationBench, ties on Agents' Last Exam, and trails on five other listed benchmarks.
Does DeepSeek V4 Pro 0813 beat Fable 5?
Not overall in the shared table. DeepSeek leads Fable 5 on Cybergym and AutomationBench, comes within 0.1 point on Terminal Bench 2.1, and trails on the other benchmarks where both have scores.
Do existing DeepSeek V4 Pro API integrations need a new model name?
No. DeepSeek keeps the API model name deepseek-v4-pro. Its official pricing page now identifies the backend version as DeepSeek-V4-Pro-0813.