DeepSeek V4 Pro
DeepSeek · Released 24 Apr 2026 · deepseek-v4-pro
$1.32
Input price / 1M tokens
#31 of 61, cheapest first
$3.96
Output price / 1M tokens
#24 of 61, cheapest first
1.05M
Context window
#11 of 71
104
Output tokens / second
#28 of 61
87.9%
Terminal-Bench 2.1
#6 of 17
42.7%
Humanity’s Last Exam (no tools)
#7 of 23
60%
Humanity’s Last Exam (with tools)
#8 of 16
62.7%
DeepSWE v1.1
#15 of 17
Summary
DeepSeek V4 Pro is an open-weight model from DeepSeek, released on 24 Apr 2026.
At $1.32 input and $3.96 output per million tokens, it is #31 of 61 on input price, cheapest first.
Its 1.05M-token context window ranks #11 of 71.
Measured output speed is 104 tokens per second, #28 of 61.
DeepSeek's flagship MoE model, 1.6T total and 49B active parameters, open weights under MIT (commercial use allowed, no extra conditions). The API now serves the GA build DeepSeek-V4-Pro-0813 (13 Aug 2026), which DeepSeek says greatly improves agent performance and adds low/high/max reasoning effort.
Benchmarks · 6 reported
| Benchmark | Score | Bar | Rank | Source |
|---|---|---|---|---|
| Terminal-Bench 2.1V4-Pro-0813, DeepSeek Harness minimal mode, max effort | 87.9% | #6 of 17 | DeepSeek-V4-Pro-0813 model cardSelf-reported | |
| Humanity’s Last Exam (no tools)no tools, text-only subset | 42.7% | #7 of 23 | DeepSeek-V4-Pro-0813 model cardSelf-reported | |
| Humanity’s Last Exam (with tools)with tools | 60% | #8 of 16 | DeepSeek-V4-Pro-0813 model cardSelf-reported | |
| DeepSWE v1.1max effort | 62.7% | #15 of 17 | DeepSeek-V4-Pro-0813 model cardSelf-reported | |
| Toolathlon-Verified | 74.1% | – | DeepSeek-V4-Pro-0813 model cardSelf-reported | |
| GPQA DiamondPass@1, max effort | 92.4% | #9 of 27 | DeepSeek-V4.1-Flash model card comparison tableSelf-reported |
Self-reported means the lab ran the test itself. A rank appears only where several labs report the same version of a benchmark.
Similar models
Models with the most benchmarks in common with DeepSeek V4 Pro, and the closest scores.
Terminal-Bench 2.1
#6 of 17Terminal tasks, earlier set
- DeepSeek V4 Pro87.9
- Kimi K388.3
- DeepSeek V4.1 Flash90.6
- Qwen3.8-Max86.6
- GLM-5.281
- GLM-5.388.2
Humanity’s Last Exam (no tools)
#7 of 23Expert questions, no search or code
- DeepSeek V4 Pro42.7
- Kimi K343.5
- DeepSeek V4.1 Flash36.8
- Qwen3.8-Max43.6
- GLM-5.240.5
Humanity’s Last Exam (with tools)
#8 of 16Expert questions, with search and code
- DeepSeek V4 Pro60
- Kimi K356
- DeepSeek V4.1 Flash63.9
- Qwen3.8-Max56.2
- GLM-5.254.7
- GLM-5.362.5
DeepSWE v1.1
#15 of 17Long coding tasks in real repos
- DeepSeek V4 Pro62.7
- Kimi K367.5
- DeepSeek V4.1 Flash74.2
- GLM-5.366.9
Compare DeepSeek V4 Pro side by side
Pre-filled with the two closest models. Swap any of them.
Sources
- price, peak/off-peak rules, context, max output, model version: api-docs.deepseek.com/quick_start/pricing, checked 2 Oct 2026
- price history (75% promo made permanent 22 May 2026; peak/off-peak from 16 Aug 2026): web.archive.org/web/20260522172917/https://api-docs.deepseek.com/quick_start/pri, checked 2 Oct 2026
- GA release, pricing change date, continued service after 14 Sep 2026: api-docs.deepseek.com/updates, checked 2 Oct 2026
- launch date, 1.6T/49B parameters: api-docs.deepseek.com/news/news260424, checked 2 Oct 2026
- license, benchmarks: huggingface.co/deepseek-ai/DeepSeek-V4-Pro-0813, checked 2 Oct 2026
- max_tokens limit 393216: api-docs.deepseek.com/api/create-chat-completion, checked 2 Oct 2026
- output speed: artificialanalysis.ai/models/deepseek-v4-pro, checked 2 Oct 2026
Building on DeepSeek V4 Pro?
Get the architecture right before the bill arrives.
We size caching, routing and fallbacks for your workload on a free call, and tell you where a cheaper model is good enough.