DeepSeek-V3: benchmark scores, price and ranking

DeepSeek-V3 from DeepSeek ranks #607 of 743 LLMs on the LLMs Tiger aggregate benchmark score (45.1), based on 10 benchmark results captured up to 2026-10-13.

Organisation: DeepSeek
Type: Open-weight / local model
Open weights: Yes
Released: 2024-12-26 (source: curated)
Cheapest API price: $0.200 in / $0.200 out per million tokens (via hyperbolic, observed 2026-10-13T19:47Z)
Aggregate rank: #607 of 743 (score 45.1)

Relative to other models, DeepSeek-V3 is strongest on SWE-bench Lite (#6 of 20) and weakest on SimpleBench (#57 of 61).

Across the 10 benchmarks where at least ten models are scored, DeepSeek-V3 is in the top 10% on 0 and in the top half on 4.

Its cheapest listed input price, $0.200 per million tokens, is lower than that of 67% of the 427 models with a listed price.

DeepSeek-V3 benchmark results

DeepSeek-V3 benchmark scores, ranked against all models measured on each benchmark
Benchmark	Score	Rank	Source	Captured
SWE-bench Lite	30.7%	#6 of 20	SWE-bench Lite leaderboard	2026-10-13
MMLU-Pro	64.4%	#53 of 136	HuggingFace MMLU-Pro leaderboard	2026-10-13
Epoch ECI	132	#81 of 204	Epoch AI Benchmarking Hub	2026-10-13
GSM8K	89.3%	#8 of 17	HuggingFace GSM8K leaderboard	2026-10-13
GPQA Diamond	58.2%	#168 of 275	HuggingFace GPQA Diamond leaderboard	2026-10-13
METR time horizon	18 min	#19 of 31	Epoch AI Benchmarking Hub	2026-10-13
OTIS Mock AIME	15.8%	#135 of 189	Epoch AI Benchmarking Hub	2026-10-13
CritPt	0.0%	#87 of 115	Epoch AI Benchmarking Hub	2026-10-13
SciCode	35.4%	#81 of 106	Epoch AI Benchmarking Hub	2026-10-13
SimpleBench	18.9%	#57 of 61	Epoch AI Benchmarking Hub	2026-10-13

About DeepSeek-V3

DeepSeek-V3 is an open-weight model that can run on your own hardware from DeepSeek. It was released on 2024-12-26, according to the developer's announcement. The cheapest API listing we track charges $0.200 per million input tokens and $0.200 per million output tokens, through hyperbolic. We hold 10 benchmark results for it across 3 areas, and none is in the top tenth of its board.

Where DeepSeek-V3 stands

On our aggregate score, DeepSeek-V3 is #607 of 743 ranked models, in the lower half. Its score is 45.1; the scale is centred on 50, so it sits 4.9 points below the typical model. That beats 18% of the ranked field. Among the 33 ranked DeepSeek models it is #28. The next one up is DeepSeek Chat V2.5 at 46.3. The next one down is DeepSeek-R1-Distill-Qwen-1.5B at 44.6.

DeepSeek-V3, benchmark by benchmark

SWE-bench Lite puts DeepSeek-V3 at 30.7%, #6 of 20. That beats 70% of the models tested, but 26.0 points short of Claude 4 Sonnet (56.7%). The results either side are Claude 3.5 Sonnet (39.0%) and o3-mini_1.0 (30.3%). It covers real GitHub issues resolved. The board median is 24.7%, so DeepSeek-V3 is 6.0 points above it. Source: SWE-bench Lite leaderboard, captured 2026-10-13.

On MMLU-Pro, DeepSeek-V3 reaches 64.4%. It ranks #53 of 136, ahead of 61% of the field, and 23.7 points behind Darwin-180B-RSI at 88.1%. The results either side are Qwen2.5-VL-72B-Instruct (64.9%) and granite-4.1-30b (64.1%). It covers 57-subject knowledge + reasoning. The board median is 55.3%, so DeepSeek-V3 is 9.1 points above it. Source: HuggingFace MMLU-Pro leaderboard, captured 2026-10-13.

DeepSeek-V3 manages 132 on Epoch ECI, good for #81 of 204. Qwen 3.8 Max leads at 156, 24 points further up, and 60% of models sit below. The results either side are Qwen2.5-Max (133) and Llama 4 Maverick (132). It covers Epoch Capabilities Index: Epoch's own aggregate of ~40 benchmarks. The board median is 127, so DeepSeek-V3 is 5 points above it. Source: Epoch AI Benchmarking Hub, captured 2026-10-13.

GSM8K puts DeepSeek-V3 at 89.3%, #8 of 17. That beats 53% of the models tested, but 10.3 points short of mimo-v2.5-pro (99.6%). The results either side are Qwen2-72B (89.5%) and granite-4.1-3b (86.9%). It covers grade-school math word problems. The board median is 86.9%, so DeepSeek-V3 is 2.4 points above it. Source: HuggingFace GSM8K leaderboard, captured 2026-10-13.

DeepSeek-V3 lands below the middle on GPQA Diamond: 58.2%, #168 of 275, with 61% of the field scoring higher. GPT-6 Astra (max) leads at 95.8%; the closest result above is mistral-medium-2505 at 59.5%. It covers expert-level science questions. The board median is 71.7%, so DeepSeek-V3 is 13.5 points below it. Source: HuggingFace GPQA Diamond leaderboard, captured 2026-10-13.

DeepSeek-V3 lands below the middle on METR time horizon: 18 min, #19 of 31, with 58% of the field scoring higher. Claude Mythos Preview (Early) leads at 1,045 min; the closest result above is o1-preview at 20 min. It covers length of task (minutes) done at 50% success. The board median is 23 min, so DeepSeek-V3 is 5 minutes below it. Source: Epoch AI Benchmarking Hub, captured 2026-10-13.

DeepSeek-V3 lands below the middle on OTIS Mock AIME: 15.8%, #135 of 189, with 71% of the field scoring higher. Claude Fable 5.1 (max) leads at 100.0%; the closest result above is qwen-max-2025-01-25 at 16.1%. It covers competition maths (AIME 2024-25 style). The board median is 57.8%, so DeepSeek-V3 is 41.9 points below it. Source: Epoch AI Benchmarking Hub, captured 2026-10-13.

DeepSeek-V3 lands below the middle on CritPt: 0.0%, #87 of 115, with 75% of the field scoring higher. GPT-5.6 Sol (max) leads at 32.3%; the closest result above is Cogito v2.1 (671B) at 0.0%. It covers unpublished research-level physics problems. The board median is 2.9%, so DeepSeek-V3 is 2.9 points below it. Source: Epoch AI Benchmarking Hub, captured 2026-10-13.

DeepSeek-V3 lands below the middle on SciCode: 35.4%, #81 of 106, with 75% of the field scoring higher. Claude Opus 5.5 (max) leads at 66.9%; the closest result above is DeepSeek-R1 at 35.7%. It covers scientific research code. The board median is 44.3%, so DeepSeek-V3 is 8.9 points below it. Source: Epoch AI Benchmarking Hub, captured 2026-10-13.

SimpleBench is a weak spot. DeepSeek-V3 scores 18.9%, #57 of 61, behind 92% of the models measured. The leader, Claude Fable 5 (max), reaches 81.9%. It covers trick questions that test common sense. The board median is 41.4%, so DeepSeek-V3 is 22.5 points below it. Source: Epoch AI Benchmarking Hub, captured 2026-10-13.

Strengths and weaknesses

Software engineering: uneven, ahead of 47% of the field on average across 2 benchmarks. Its best board here is SWE-bench Lite (#6 of 20) and its weakest is SciCode (#81 of 106). Reasoning, maths and science: uneven, ahead of 32% of the field on average across 5 benchmarks. Its best board here is MMLU-Pro (#53 of 136) and its weakest is SimpleBench (#57 of 61). General knowledge and chat: solid, ahead of 51% of the field on average across 3 benchmarks. Its best board here is Epoch ECI (#81 of 204) and its weakest is METR time horizon (#19 of 31).

Price and value

The median input price across the 427 priced models is $0.500 per million tokens, so DeepSeek-V3 costs less than the typical listing. Models that score within 1.5 points or better and cost less: Qwen2.5-Coder-7B-Instruct ($0.010, 49.5), DeepSeek V4.1 Flash (high) ($0.010, 48.9), Qwen2.5-Coder-3B-Instruct ($0.010, 46.5). Similar scores cost more on MiniMax-M3 ($0.230) and MiniMax-M2.7 ($0.250).

Release history

The DeepSeek model released before it was DeepSeek Chat V3 (prev) on 2024-12-26, which scores 50.5. The next one was DeepSeek-R1 on 2025-01-20, at 48.0.

Reading these numbers

The 10 scores on this page come from 5 sources: Epoch AI Benchmarking Hub (6), HuggingFace GPQA Diamond leaderboard (1), HuggingFace GSM8K leaderboard (1), HuggingFace MMLU-Pro leaderboard (1). 1 of them come from an official leaderboard and 9 from datasets or third-party runs.

Models ranked nearby

#604 deepseek-math-7b-instruct (45.2)
#605 ollama/hermes3:8b-llama3.1-fp16 (45.1)
#606 ollama/wojtek/opencodeinterpreter:6.7b (45.1)
#608 DeepSeek-Coder-V2-Lite-Base (45.0)
#609 mistral-small-3-2501 (45.0)
#610 GLM-4.6 (45.0)