DeepSeek officially launched V4 Pro on August 13, 2026, one day after the production API began identifying the model as DeepSeek-V4-Pro-0813. The full announcement fills in several important blanks: broader agent benchmarks, selectable reasoning effort, native Responses API support, and a new time-based pricing schedule.
The short version is that 0813 is consistently stronger than V4 Flash 0731 and the April V4 Pro Preview across DeepSeek's published agent evaluations. The largest gains show up on repository work, software engineering, cybersecurity, and automation tasks. This looks like more than a small checkpoint refresh.
V4 Pro 0813 versus Flash 0731 and Pro Preview
| Benchmark | V4 Pro 0813 | V4 Flash 0731 | V4 Pro Preview |
|---|---|---|---|
| HLE, without tools | 42.7 | 37.8 | 37.7 |
| HLE, with tools | 60.0 | 51.5 | 48.2 |
| Terminal-Bench 2.1 | 87.9 | 82.7 | 72.1 |
| NL2Repo | 61.5 | 54.2 | 38.5 |
| Cybergym | 83.3 | 76.7 | 52.7 |
| DeepSWE | 62.7 | 54.4 | 12.8 |
| Toolathlon-Verified | 74.1 | 70.3 | 55.9 |
| Agents' Last Exam | 25.7 | 25.2 | 16.5 |
| AutomationBench, public | 31.8 | 25.1 | 12.8 |
| DSBench-FullStack | 71.1 | 68.7 | 41.8 |
| DSBench-Hard | 67.2 | 59.6 | 31.1 |
Source: DeepSeek's official August 13 launch table. Higher is better. HLE reports separate results without and with tools.
0813 leads the other two DeepSeek models in every row. Some gaps are modest. Agents' Last Exam moves only 0.5 points over Flash. Others are hard to dismiss. On DeepSWE, 0813 scores 62.7 against 54.4 for Flash and 12.8 for Pro Preview. On NL2Repo, it reaches 61.5 versus the Preview's 38.5. DSBench-Hard rises from 31.1 to 67.2.
The Terminal-Bench improvement is 15.8 percentage points over Pro Preview, from 72.1 to 87.9. That is about 21.9 percent when measured relative to the Preview score. It also places 0813 5.2 points above Flash 0731.
Why 0813 is more capable than V4 Flash
The April release positioned Flash as the smaller, quicker option. It has 284 billion total parameters with 13 billion active for each token. Pro is much larger at 1.6 trillion total parameters and 49 billion active. Both support a one-million-token context window and thinking or non-thinking modes.
DeepSeek originally said Flash could approach Pro's reasoning ability and match it on simple agent tasks. That still appears plausible for lighter work. Agents' Last Exam differs by only half a point, while DSBench-FullStack differs by 2.4. More demanding repository and automation benchmarks create a wider separation.
Flash remains the efficiency model. From August 16, Pro will cost exactly three times as much as Flash at both peak and off-peak rates. For short, predictable edits, Flash may still produce the better cost-to-result ratio. For longer tasks where a failed attempt means another full run, the stronger Pro model can be cheaper in practice despite its higher token rate.
The bigger story is the jump over Pro Preview
The more revealing comparison is with DeepSeek's April Pro Preview. Architecture remains 1.6 trillion total parameters with 49 billion active, yet performance changes sharply. 0813 gains 49.9 points on DeepSWE, 36.1 on DSBench-Hard, 30.6 on Cybergym, and 23 points on NL2Repo.
DeepSeek describes the release as a major agent upgrade with production gains, but it has not explained which post-training changes produced them. The model size and one-million-token context window are unchanged. Until a new technical report appears, attributing the gain to a specific training method would be guesswork.
Low, high, and max reasoning effort
Both V4 Pro and V4 Flash now expose three reasoning levels. DeepSeek recommends low for simple work, high for normal agent workflows, and max for difficult tasks. This gives an application a more useful control than a basic thinking-mode switch.
DeepSeek ran the public code-agent tests for 0813 at max reasoning with top_p=0.95 and temperature=1.0. It used an upcoming DeepSeek Harness in minimal mode. Results can vary with another harness or reasoning level, so the table should be read as a high-effort result rather than a default-settings guarantee.
Native Responses API and agent support
DeepSeek now supports the OpenAI Responses API format at https://api.deepseek.com. That makes V4 models easier to connect to clients built around Responses, including Codex. DeepSeek also provides a one-command Codex setup in its integration guide.
Compatibility is not complete. DeepSeek's documentation says unsupported Responses parameters are silently ignored. Built-in tools such as file search, code interpreter, computer use, and MCP are ignored by the API. Custom apply_patch is supported for Codex compatibility. Anyone porting an existing agent should check the compatibility table instead of assuming every OpenAI feature works.
How to access V4 Pro 0813
The stable API model name remains deepseek-v4-pro. DeepSeek's pricing table identifies the version behind it as DeepSeek-V4-Pro-0813. The endpoint supports Chat Completions, the Anthropic-compatible interface, and now the Responses API. V4 Pro is also available through Expert Mode in DeepSeek's web and mobile apps.
Existing integrations can keep the same base URL and model name. The update sits behind the stable alias, so there is no migration step for direct API users. That convenience has a tradeoff: an alias can move to a later build. Teams that need reproducible evaluations should record the model version returned by the provider and rerun their test set after an alias changes.
New peak and off-peak API pricing
DeepSeek is raising V4 API prices at 16:00 UTC on August 16, 2026. It is also splitting the day into peak and off-peak periods. Peak hours run from 01:00 to 04:00 UTC and 06:00 to 10:00 UTC. Every other hour receives a 50 percent discount.
| Model and period | Cached input | Uncached input | Output |
|---|---|---|---|
| V4 Pro, off-peak | $0.022 | $0.66 | $1.98 |
| V4 Pro, peak | $0.044 | $1.32 | $3.96 |
| V4 Flash, off-peak | $0.007 | $0.22 | $0.66 |
| V4 Flash, peak | $0.014 | $0.44 | $1.32 |
USD per one million tokens. Effective August 16, 2026 at 16:00 UTC.
The price increase changes yesterday's value calculation. Pro currently costs $0.435 per million uncached input tokens and $0.87 per million output tokens. Its new off-peak rates will be $0.66 and $1.98. Peak rates rise to $1.32 and $3.96. Scheduling batch work outside the seven daily peak hours will cut the new rate in half, but even off-peak output will cost more than twice today's price.
My read on the release
V4 Pro 0813 is a meaningful agent update. It beats Flash 0731 and Pro Preview across every benchmark DeepSeek disclosed, sometimes by a wide margin. Selectable reasoning effort and Responses API support also make the model easier to use in real agent systems.
Two cautions remain. The detailed training changes behind 0813 are still undocumented, and the benchmark results use DeepSeek's forthcoming harness at max reasoning. The new pricing also narrows the cost advantage. Even so, 0813 now has much stronger evidence behind the claim that it is the most capable model in the V4 family.