DeepSeek v4 Pro (max): benchmark scores, price and ranking

DeepSeek v4 Pro (max) from DeepSeek ranks #108 of 743 LLMs on the LLMs Tiger aggregate benchmark score (51.6), based on 13 benchmark results captured up to 2026-10-13.

Organisation

DeepSeek

Type

Open-weight / local model

Open weights

Yes

Released

2026-04-24 (source: epoch)

Cheapest API price

$0.435 in / $0.870 out per million tokens (via tencent, observed 2026-10-13T19:47Z)

Aggregate rank

#108 of 743 (score 51.6)

Relative to other models, DeepSeek v4 Pro (max) is strongest on SWE-bench Verified (#2 of 73) and weakest on FrontierMath T4 (#47 of 50).

Across the 13 benchmarks where at least ten models are scored, DeepSeek v4 Pro (max) is in the top 10% on 1 and in the top half on 10.

Its cheapest listed input price, $0.435 per million tokens, is lower than that of 53% of the 427 models with a listed price.

DeepSeek v4 Pro (max) benchmark results

About DeepSeek v4 Pro (max)

DeepSeek v4 Pro (max) is an open-weight model that can run on your own hardware from DeepSeek. It was released on 2026-04-24, according to Epoch AI. The cheapest API listing we track charges $0.435 per million input tokens and $0.870 per million output tokens, through tencent. We hold 13 benchmark results for it across 4 areas, and 1 is in the top tenth of its board.

Where DeepSeek v4 Pro (max) stands

On our aggregate score, DeepSeek v4 Pro (max) is #108 of 743 ranked models, in the top quarter. Its score is 51.6; the scale is centred on 50, so it sits 1.6 points above the typical model. That beats 85% of the ranked field. Among the 33 ranked DeepSeek models it is #6. The next one up is DeepSeek-V2.5-1210 at 51.7. The next one down is DeepSeek Chat V2 0628 at 51.4.

DeepSeek v4 Pro (max), benchmark by benchmark

On SWE-bench Verified, DeepSeek v4 Pro (max) scores 77.6% and ranks #2 of 73. Only one model scores higher, led by Claude Opus 4.7 (max) at 83.5%, which is 5.8 points ahead. It covers real GitHub issues resolved. The board median is 56.4%, so DeepSeek v4 Pro (max) is 21.2 points above it. Source: Epoch AI Benchmarking Hub, captured 2026-10-13.

On OTIS Mock AIME, DeepSeek v4 Pro (max) reaches 96.7%. It ranks #23 of 189, ahead of 88% of the field, and 3.3 points behind Claude Fable 5.1 (max) at 100.0%. The results either side are Kimi K3 (max) (97.2%) and Kimi K2.6 (96.1%). It covers competition maths (AIME 2024-25 style). The board median is 57.8%, so DeepSeek v4 Pro (max) is 38.9 points above it. Source: Epoch AI Benchmarking Hub, captured 2026-10-13.

DeepSeek v4 Pro (max) manages 89.6% on GPQA Diamond, good for #39 of 275. GPT-6 Astra (max) leads at 95.8%, 6.1 points further up, and 86% of models sit below. The results either side are Claude Haiku 5.5 (max) (89.6%) and grok-4.20-0309-reasoning (89.3%). It covers expert-level science questions. The board median is 71.7%, so DeepSeek v4 Pro (max) is 17.9 points above it. Source: Epoch AI Benchmarking Hub, captured 2026-10-13.

DTBench puts DeepSeek v4 Pro (max) at 90.7%, #23 of 146. That beats 84% of the models tested, but 8.3 points short of Claude Opus 5.5 (max) (98.9%). The results either side are Kimi K2.6 (90.9%) and GPT-5 (high) (90.7%). It covers dynamic tool-use benchmark. The board median is 68.8%, so DeepSeek v4 Pro (max) is 21.9 points above it. Source: Epoch AI Benchmarking Hub, captured 2026-10-13.

On LMCA, DeepSeek v4 Pro (max) reaches 41.2%. It ranks #27 of 108, ahead of 75% of the field, and 27.0 points behind Claude Opus 5.5 (max) at 68.2%. The results either side are DeepSeek V4 Flash 0731 (max) (41.7%) and GPT-5.4 mini (xhigh) (40.8%). It covers language-model capability assessment. The board median is 29.2%, so DeepSeek v4 Pro (max) is 12.0 points above it. Source: Epoch AI Benchmarking Hub, captured 2026-10-13.

DeepSeek v4 Pro (max) manages 48.9% on WeirdML, good for #30 of 109. GPT-6 Astra (max) leads at 93.3%, 44.4 points further up, and 72% of models sit below. The results either side are grok-4-20 (52.3%) and GLM-5 (48.2%). It covers unusual machine-learning tasks. The board median is 41.2%, so DeepSeek v4 Pro (max) is 7.7 points above it. Source: Epoch AI Benchmarking Hub, captured 2026-10-13.

SimpleQA Verified puts DeepSeek v4 Pro (max) at 47.0%, #26 of 69. That beats 62% of the models tested, but 29.0 points short of gemini-4-argon-high (76.0%). The results either side are Claude Opus 4.6 (max) (47.0%) and Claude Sonnet 5.5 (max) (46.5%). It covers short factual questions. The board median is 41.0%, so DeepSeek v4 Pro (max) is 6.0 points above it. Source: Epoch AI Benchmarking Hub, captured 2026-10-13.

DeepSeek v4 Pro (max) lands below the middle on FrontierMath T1-3: 45.3%, #41 of 68, with 59% of the field scoring higher. GPT-6 Astra (max) leads at 93.7%; the closest result above is Inkling Small (xhigh) at 46.3%. It covers research-level maths, tiers 1-3. The board median is 55.6%, so DeepSeek v4 Pro (max) is 10.4 points below it. Source: Epoch AI Benchmarking Hub, captured 2026-10-13.

DeepSeek v4 Pro (max) lands below the middle on ProofBench: 16.0%, #29 of 46, with 61% of the field scoring higher. Claude Fable 5.1 (max) leads at 100.0%; the closest result above is Muse Spark at 17.0%. It covers writing mathematical proofs. The board median is 21.0%, so DeepSeek v4 Pro (max) is 5.0 points below it. Source: Epoch AI Benchmarking Hub, captured 2026-10-13.

FrontierMath T4 is a weak spot. DeepSeek v4 Pro (max) scores 2.4%, #47 of 50, behind 92% of the models measured. The leader, GPT-6.1 Sol (max), reaches 100.0%. It covers research-level maths, tier 4. The board median is 30.5%, so DeepSeek v4 Pro (max) is 28.0 points below it. Source: Epoch AI Benchmarking Hub, captured 2026-10-13.

The full table above also lists the other 3 results, including boards with fewer than ten models, where a rank says little.

Strengths and weaknesses

Software engineering: strong, ahead of 82% of the field on average across 2 benchmarks. Its best board here is SWE-bench Verified (#2 of 73) and its weakest is SciCode (#36 of 106). Reasoning, maths and science: solid, ahead of 55% of the field on average across 6 benchmarks. Its best board here is OTIS Mock AIME (#23 of 189) and its weakest is FrontierMath T4 (#47 of 50). General knowledge and chat: solid, ahead of 68% of the field on average across 4 benchmarks. Its best board here is LMCA (#27 of 108) and its weakest is SimpleQA Verified (#26 of 69).

Price and value

The median input price across the 427 priced models is $0.500 per million tokens, so DeepSeek v4 Pro (max) costs less than the typical listing. Models that score within 1.5 points or better and cost less: Qwen3.7 Flash ($0.030, 50.5), GLM-5.2 (high) ($0.060, 51.6), DeepSeek V4 Flash 0731 (max) ($0.070, 53.1). Similar scores cost more on Kimi K2.5 (Fireworks) ($0.450) and Kimi K2.5 ($0.450).

Release history

The DeepSeek model released before it was DeepSeek v4 Flash (max) on 2026-04-24, which scores 49.6. The next one was DeepSeek v4 (unknown) on 2026-04-24, at 50.0.

Reading these numbers

The 13 scores on this page come from 1 source: Epoch AI Benchmarking Hub (13). No official leaderboard supplied any of them, so treat them as third-party measurements. The leaderboard also lists 3 other reasoning-effort settings of this model, which redirect here. This page shows the highest.

DeepSeek v4 Pro (max) is also listed at 3 other reasoning-effort settings; this page shows the highest-effort row, which is the one the leaderboard keeps by default.

Models ranked nearby

• #105 GLM-4.7-FP8 (51.6)
• #106 Codex + GPT-6 Sol (51.6)
• #107 o1-2024-12-17 (51.6)
• #109 GLM-5.2 (high) (51.6)
• #110 Hy3-preview (51.6)
• #111 LongCat-Flash-Chat (51.6)

More from DeepSeek

• DeepSeek V4 Pro 0813 (max) (#23, 54.3)
• DeepSeek V4 Flash 0731 (max) (#42, 53.1)
• DeepSeek-V3.2-Exp (high) (#67, 52.4)
• deepseek-v4.1-flash-max (#75, 52.2)
• DeepSeek-V2.5-1210 (#101, 51.7)
• DeepSeek Chat V2 0628 (#127, 51.4)

Compare it on the live leaderboard → · How the aggregate score works