“This time, it’s official.” That is how SiliconFlow opened its launch note for DeepSeek-V4-Pro-0813, and the production release largely backs up the hype: DeepSeek’s flagship model is out of preview with dramatically stronger agent benchmarks, native OpenAI Responses API support, and a shift to peak and off-peak pricing that takes effect August 17.
What DeepSeek Shipped
DeepSeek published the production build of V4 Pro on August 12, with the official announcement following on August 13. The model, versioned DeepSeek-V4-Pro-0813, is live in the DeepSeek app and web client through a new “Expert Mode” and on the API under the unchanged model name deepseek-v4-pro. The release completes a promise from July 31, when DeepSeek shipped the V4 Flash production build and said the Pro model would follow soon. In between, the company also released DeepSeek Harness v0.1, an open developer preview of its agent harness, which the official account announced the same week.
The framing this time is agent capability. DeepSeek says the production V4 Pro shows especially strong gains in real production environments, and the published numbers back that up. Terminal Bench 2.1 jumped from 72.1 to 87.9, and DeepSWE, a benchmark for software engineering agents, went from 12.8 to 62.7. The full set of reported scores:
| Benchmark | V4 Pro (production) |
|---|---|
| HLE (without / with tools) | 42.7 / 60.0 |
| Terminal Bench 2.1 | 87.9 |
| NL2Repo | 61.5 |
| Cybergym | 83.3 |
| DeepSWE | 62.7 |
| Toolathlon-Verified | 74.1 |
| Agents’ Last Exam | 25.7 |
| AutomationBench (Public) | 31.8 |
| DSBench-FullStack | 71.1 |
| DSBench-Hard | 67.2 |
Two of those numbers, Terminal Bench and DeepSWE, are the sort of improvements that change purchasing decisions, and they are consistent with what independent observers report. Rohan Paul’s benchmark tracking puts V4 Pro (Max) at 1607 on Code Arena’s WebDev AutoEval, just 15 points behind GPT-5.6 Sol xHigh, while the blended token price is roughly one-thirty-first of that model. SemiAnalysis, which rarely hands out praise freely, called the 1.5T-parameter model a large step ahead of Nemotron 3 Ultra on agent tasks and noted that the V4 Flash 0731 build outperforms Nemotron 3 Ultra too, with 4.2x fewer active parameters.
What Changed on the API
Three API changes matter more than the model swap itself. First, DeepSeek now natively supports the OpenAI Responses API format, with specific tuning for Codex, and ships a one-command configuration script. That is a meaningful compatibility move: teams already wired to OpenAI-style agent tooling can point at DeepSeek without rewriting their agent loop, and Codex users get a cheaper open-weight backend for the same interface.
Second, both V4 Pro and V4 Flash now support three reasoning-effort levels: low, high, and max. DeepSeek’s guidance is practical: low for simple tasks, high for everyday agent workflows, and max for the hardest problems. This matters because it turns the model into a dial rather than a single price point. The same deployment can burn tokens only when the task warrants it, which is exactly how cost-conscious agent teams want to operate.
Third, pricing is changing structurally. DeepSeek published per-million-token rates of 0.025 yuan for cache hits, 3 yuan for cache misses, and 6 yuan for output. From August 17, the API moves to peak and off-peak pricing, with off-peak rates at half the peak price. SiliconFlow, which listed V4 Pro with day-zero support, quotes $1.32 per million input tokens, $3.96 output, and $0.44 for cache hits, with a 1M context window and an MIT license on the weights. The Decoder’s headline captured the trade-off bluntly: DeepSeek shipped a much stronger model and, at the same time, raised API prices. The direction of travel is up from the preview, even if the capability per dollar still leads the market.
Where V4 Pro Still Trails
The honest picture includes real gaps. The Decoder notes that despite the jump, V4 Pro’s overall ranking still trails Claude Opus 5, so the “frontier-class” framing depends on the benchmark and the task category. Several of the strongest scores, including DSBench-FullStack and DSBench-Hard, come from DeepSeek’s own internal test suites, which means they cannot be independently reproduced. The peak and off-peak pricing scheme also adds forecasting complexity: cost models that were simple before now need to know when a job runs, and teams with bursty workloads will pay full price on the peaks they cannot schedule around.
The reasoning-effort dial is not free either. Low effort cuts token spend but also cuts quality, and choosing the wrong level for an agent loop can silently degrade results. DeepSeek’s own examples are honest about this: the production benchmark figures were run at max effort, so quoting them while running at high or low is apples to oranges. For a sense of how the Flash-tier model performed on agent benchmarks before this release, our earlier coverage of DeepSeek V4 Flash agent benchmarks gives the baseline those numbers improved on.
Who Should Care
Teams building coding agents or tool-using workflows on a budget get the most out of this release, particularly if they already speak OpenAI’s Responses API. The compatibility layer removes the biggest switching cost, and the price, even after the increase, stays far below Western frontier models. Codex users looking for an open-weight fallback have a credible option for the first time. Teams with latency-sensitive or always-on workloads should treat the reasoning-effort settings as a required engineering task, not an optional tweak, because the default choice will materially change both cost and behavior. Regulated or enterprise teams that need auditable, stable API contracts should wait a beat and watch whether the pricing scheme settles, since peak and off-peak structures are new territory for DeepSeek’s API.
The Bottom Line
DeepSeek V4 Pro’s production release is the strongest open-weight agent model the company has shipped. The agent benchmark jumps are large and directionally confirmed by third parties, the Responses API support lowers switching costs, and the reasoning-effort dial gives teams control they did not have before. The caveats, a price increase, self-reported internal benchmarks, and a ranking still below Claude Opus 5, are real but do not change the core story: the cheapest serious agent model just got considerably more serious.