DeepSeek V4-Pro Is GA: New Agent Benchmarks and Pricing

DeepSeek V4-Pro is now generally available. The GA release rolled out on August 13 across the DeepSeek APP, Web, and API — nine days after we tracked the model’s confirmation — and it ships with a new pricing structure that takes effect this weekend: peak/off-peak API rates start August 16 at 16:00 UTC.

The official changelog confirms the model name stays deepseek-v4-pro (version DeepSeek-V4-Pro-0813), so the API calling method is unchanged. What is new: a big jump in agent benchmarks, native support for the OpenAI Responses API format, three thinking-effort levels, and the first-ever peak/off-peak billing model in DeepSeek’s pricing history.

Key facts

• GA on August 13, 2026 — DeepSeek-V4-Pro rolled out on APP, Web, and API as deepseek-v4-pro; API calling method unchanged.
• Agent capability is the headline. Official numbers: Terminal-Bench 2.1 at 87.9, DeepSWE at 62.7, Cybergym at 83.3, Toolathlon-Verified at 74.1, HLE with tools at 60.0, and DSBench-FullStack at 71.1.
• 1M context, 384K max output, and a concurrency limit of 500 on the API.
• Native Responses API support — the DeepSeek API now speaks the OpenAI Responses format directly and is specifically adapted for Codex, with a one-click configuration script.
• Thinking effort levels: V4-Pro and V4-Flash now support low / high / max thinking modes.
• Peak/off-peak pricing effective August 16, 16:00 UTC — off-peak rates are half of peak rates; peak hours are 01:00–04:00 and 06:00–10:00 UTC.

What changed

The August 13 changelog marks the moment the V4 family stopped being a preview story. V4-Flash reached public beta on July 31; V4-Pro now goes GA with benchmark results that improve on the Flash retrain across the board — Terminal-Bench 2.1 moves from 82.7 on V4-Flash-0731 to 87.9 on V4-Pro-0813.

Two developer-facing changes matter more than the scores. First, the Responses API: instead of the Chat Completions workaround, DeepSeek now natively supports the format OpenAI built for agent loops, and documents a one-click Codex configuration. Second, thinking effort: low, high, and max give developers a direct lever on cost and latency per task, the same pattern Google ships with Gemini.

The pricing change is the most consequential. When the new rates take effect at 16:00 UTC on August 16, current per-million-token prices ($0.435 input / $0.87 output for V4-Pro) are replaced by a dynamic schedule: off-peak $0.66 / $1.98, peak $1.32 / $3.96. V4-Flash moves to off-peak $0.22 / $0.66 and peak $0.44 / $1.32. Cache-hit input stays dramatically cheap at $0.022–$0.044 per million.

Why it matters

• DeepSeek is monetizing time, not just tokens. Peak/off-peak pricing is an infrastructure play — it spreads load and rewards off-hours batch work. No major Western API has adopted this pattern at scale, and it lets DeepSeek raise headline prices without raising the effective cost for disciplined users.
• The open frontier keeps its roadmap. The GA follows the V4-Pro confirmation we covered on August 7, and it lands three days after OpenAI previewed GPT-5.6 Sol Ultrafast at 750 tokens per second. Both camps shipped speed and agent capability in the same week — the frontier race is now an agent-benchmark race.
• Even at peak prices, V4-Pro is an order of magnitude below closed flagships — see our open-weight API pricing tracker for the full matrix. DeepSeek remains the anchor that closed labs price against, exactly as it was when V4-Flash made $0.14-per-million input the baseline.

What to watch

• The August 16 price switch. Peak/off-peak billing begins 16:00 UTC Saturday; the first bill cycle will show whether developers shift workloads to off-peak windows or just pay more.
• Independent retests. DeepSeek’s benchmark table is self-reported; Artificial Analysis and LMArena runs will tell us whether Terminal-Bench 87.9 holds under third-party harnesses.
• Codex adoption. The one-click configuration plus Responses API support makes DeepSeek the cheapest serious option for agentic coding workflows — watch for enterprise routing decisions.
• The open-weight release. The GA changelog is API-focused; DeepSeek has not yet attached a date to the open weights of the 1.6T-class model family.

Official source

• DeepSeek: Change Log — DeepSeek-V4-Pro Update (August 13, 2026)
• DeepSeek: Models & Pricing — current and peak/off-peak rates

Updated August 16, 2026.