DeepSeek V4 Pro (final version)

2026-08-13 · Publisher : DeepSeek · Domain : Artificial Intelligence

DeepSeek releases the final version of V4-Pro (V4-Pro-0813), with markedly strengthened agent capabilities, three levels of thinking effort and a pricing grid with peak and off-peak rates.

### Context The April preview had placed V4-Pro close to closed models on reasoning, but behind on long agent tasks. In the meantime, V4-Flash had received a post-trained version on July 31 (V4-Flash-0731). V4-Pro-0813 replaces the preview and targets production uses: code agents, tool calls, automation. ### What the model brings - Sharply increased agent capabilities, with marked gains in production environments according to the vendor. - Three levels of thinking effort (low, high, max), applicable to V4-Pro and V4-Flash. - Native support for OpenAI's Responses API, adapted to Codex with a one-click configuration script. - A built-in DSpark speculative decoding module, with vLLM and SGLang support, to speed up generation. - Availability in the app and on the website ("Expert Mode") and through the API, under the same identifier `deepseek-v4-pro`. ### Architecture and specifications - Type: Mixture-of-Experts, same structure as V4-Pro (1.6 trillion parameters in total, 49 billion active, context of one million tokens); Hugging Face shows 1.7 T with the DSpark module. - Context: 1 M; maximum output of 384K tokens (pricing page). - License: MIT (code and weights), commercial use allowed. - Weights: huggingface.co/deepseek-ai/DeepSeek-V4-Pro-0813. - Deployment: the card cites a single node of 4 GB300 GPUs with vLLM. Settings: temperature 1.0, top_p 0.95 for agents, maximum output of 384K for the high and max levels. ### Cost and access | API (USD per million tokens) | Input, cache hit | Input, cache miss | 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 | Pricing effective August 16, 2026 at 4 p.m. UTC; peak hours from 01:00 to 04:00 and from 06:00 to 10:00 UTC on weekdays, the rest being off-peak. The post announces off-peak prices half those of peak. ### Benchmarks | Benchmark | V4-Pro-0813 | V4-Pro preview | GLM-5.2 | Kimi K3 | Opus-4.8 | Fable-5 (with fallback) | |---|---|---|---|---|---|---| | HLE without / with tools | 42.7 / 60.0 | 37.7 / 48.2 | 40.5 / 54.7 | 43.5 / 56.0 | 49.8 / 57.9 | 53.3 / 63.0 | | Terminal Bench 2.1 | 87.9 | 72.1 | 81.0 | 88.3 | 85.0 | 88.0 | | NL2Repo | 61.5 | 38.5 | 48.9 | not published | 69.7 | not published | | CyberGym | 83.3 | 52.7 | not published | 80.0 | 78.3 | 83.1 | | DeepSWE | 62.7 | 12.8 | 46.2 | 67.5 | 58.0 | 70.0 | | Toolathlon-Verified | 74.1 | 55.9 | 59.9 | 76.5 | 76.2 | 77.9 | Source: model card for DeepSeek-V4-Pro-0813 on Hugging Face, scores measured by the vendor (DeepSeek Harness framework, max effort). ### Versus closed models The closed models compared by the vendor are Opus 4.8 and Fable 5 (with fallback). V4-Pro-0813 outperforms Opus 4.8 on Terminal Bench 2.1 (87.9 vs 85.0), CyberGym (83.3 vs 78.3) and DeepSWE (62.7 vs 58.0). It remains behind Opus 4.8 on Humanity's Last Exam without tools (42.7 vs 49.8) and on NL2Repo (61.5 vs 69.7), and behind Fable 5 on DeepSWE (62.7 vs 70.0). The progress relative to the preview is considerable on the agent side: DeepSWE goes from 12.8 to 62.7 and Terminal Bench from 72.1 to 87.9. Since the closed models' prices are not in the source, the price gap is not quantified; the vendor lists 1.98 dollars per million output tokens off-peak. ### Key takeaways This final version fixes V4-Pro's main weakness: the agent in real-world conditions. It also introduces time-of-day pricing, which changes how workloads are planned.

Tags : LLM, Open Source, DeepSeek, MIT, MoE, Agents, Contexte 1M