DeepSeek V4 Pro
DeepSeek's larger V4 model (August 13 snapshot), released with open weights under the MIT license and sold at low prices with a further 50% off-peak discount. On most tests here it now trails DeepSeek's own newer V4.1 Flash.
DeepSeekLanguage modelReleased August 2026Data checked October 18, 2026
This model is not in the current people's rankings: it is either new and still collecting votes, or ranked below the top 12. See the AI leaderboard for the models people rate highest.
On paper
Fixed-test rankings from the benchmarks tab: Speed #13, Reasoning #17.
Strengths and weaknesses
- Open weights under the MIT license
- 61.3% on ARC-AGI-2 and 92.8% on GPQA Diamond
- Off-peak API rates are half the standard $1.32 / $3.96 per million tokens
- Lower Artificial Analysis Intelligence Index (36.0) than the cheaper V4.1 Flash (39.5)
- Not on the LMArena Vision leaderboard; DeepSeek offers image understanding through a separate V4 Flash Vision model
Best for
- Low-cost reasoning and writing
- Self-hosting a large open model
- Off-peak batch jobs
Every score we have
Each bar compares this model with the best result any model on this page has on the same test.
- LMArena text arena (human votes)1,464.6Arena score 1464.6 (95% CI 1458.0 to 1471.2); 9,800 votes; entry "deepseek-v4-pro-high-20260813"; position 56 of 416 on the full board
- LMArena creative writing1,446.2Arena score 1446.2 (95% CI 1432.3 to 1460.1); 1,985 votes; entry "deepseek-v4-pro-high-20260813"; position 46 of 414 on the full board
- LMArena WebDev (coding)1,582Arena score 1582.0 (95% CI 1572.5 to 1591.5); 4,728 votes; entry "DeepSeek-V4-Pro (High, 20260813)"; position 33 of 144 on the full board
- LMArena Agent Arena0.002IPS score 0.0016 (95% CI -0.0161 to 0.0193); 7,070 sessions; entry "DeepSeek-V4-Pro (High, 20260813)"; position 30 of 52 on the full board
- Artificial Analysis Intelligence Index3636.0 on the Artificial Analysis Intelligence Index v4.3.2 (AA shows 36), setting "DeepSeek V4 Pro 0813 (max)"
- Humanity's Last Exam (Artificial Analysis)41.0%41.0% on Humanity's Last Exam (AA, 2,158 text-only questions), setting "DeepSeek V4 Pro 0813 (max)"
- GPQA Diamond, graduate-level science92.8%92.8% on GPQA Diamond (AA; no longer part of the AA index since v4.2), setting "DeepSeek V4 Pro 0813 (max)"
- Terminal-Bench 4.0 (Artificial Analysis run)14.1%14.1% on Terminal-Bench 4.0 as run by AA (mini-swe-agent harness), setting "DeepSeek V4 Pro 0813 (max)"
- Long-context reasoning (AA-LCR)80.3%80.3% on AA-LCR (long-context reasoning), setting "DeepSeek V4 Pro 0813 (max)"
- AA-Omniscience, accuracy minus hallucination0.80.8 on the AA-Omniscience Index (knowledge accuracy minus hallucination, range -100 to 100), setting "DeepSeek V4 Pro 0813 (max)"
- Output speed, tokens per second79 t/s79 output tokens per second (AA median), row "DeepSeek V4 Pro 0813 (max)"
- Cost per Artificial Analysis test task$0.67$0.67 to run one AA Intelligence Index task, row "DeepSeek V4 Pro 0813 (max)"
- Artificial Analysis Coding Agent Index43.143.1 on the AA Coding Agent Index v1.5 (DeepSWE v1.1, Terminal-Bench 4.0, SWE-Atlas-QnA), best listed setup "Codex - DeepSeek V4 Pro 0813 (max)"
- ARC-AGI-2 puzzles61.3%61.25% on ARC-AGI-2 (semi-private set), best listed setting "DeepSeek V4 Pro 0813 (Max)", $0.60 per task
Compare with
- Claude Opus 5.5Anthropic's flagship for long-running agentic coding and knowledge work, and the model Anthropic suggests starting with for most jobs.
- Claude Sonnet 5.5Anthropic's mid-priced model, which Anthropic describes as its best mix of speed and intelligence.
- Claude Fable 5.1Anthropic's top tier for the hardest reasoning and long-horizon agent work, which Anthropic recommends when Opus 5.5 still falls short.
- Claude Haiku 5.5Anthropic's smallest and fastest current model, built for high-volume, latency-sensitive jobs like classification, extraction and routing.