Kimi K2 Thinking: benchmark scores, price and ranking
Kimi K2 Thinking from Moonshot ranks #401 of 743 LLMs on the LLMs Tiger aggregate benchmark score (47.8), based on 8 benchmark results captured up to 2026-10-13.
- Organisation
- Moonshot
- Type
- Open-weight / local model
- Open weights
- Yes
- Cheapest API price
- $0.550 in / $2.20 out per million tokens (via vercel_ai_gateway, observed 2026-10-13T19:47Z)
- Aggregate rank
- #401 of 743 (score 47.8)
Relative to other models, Kimi K2 Thinking is strongest on Epoch ECI (#25 of 204) and weakest on CL-bench (#15 of 16).
Across the 8 benchmarks where at least ten models are scored, Kimi K2 Thinking is in the top 10% on 0 and in the top half on 5.
Its cheapest listed input price, $0.550 per million tokens, is lower than that of 47% of the 427 models with a listed price.
Kimi K2 Thinking benchmark results
| Benchmark | Score | Rank | Source | Captured |
|---|---|---|---|---|
| Epoch ECI | 146 | #25 of 204 | Epoch AI Benchmarking Hub | 2026-10-13 |
| GPQA Diamond | 84.5% | #84 of 275 | HuggingFace GPQA Diamond leaderboard | 2026-10-13 |
| SWE-bench Verified | 63.4% | #28 of 73 | SWE-bench Verified leaderboard | 2026-10-13 |
| METR time horizon | 54 min | #12 of 31 | Epoch AI Benchmarking Hub | 2026-10-13 |
| WeirdML | 42.8% | #50 of 109 | Epoch AI Benchmarking Hub | 2026-10-13 |
| Terminal-Bench 2.0 | 35.7% | #16 of 28 | Epoch AI Benchmarking Hub | 2026-10-13 |
| ALE-Bench | 598 | #55 of 79 | Epoch AI Benchmarking Hub | 2026-10-13 |
| CL-bench | 11.9% | #15 of 16 | Epoch AI Benchmarking Hub | 2026-10-13 |
About Kimi K2 Thinking
Kimi K2 Thinking is an open-weight model that can run on your own hardware from Moonshot. The cheapest API listing we track charges $0.550 per million input tokens and $2.20 per million output tokens, through vercel_ai_gateway. We hold 8 benchmark results for it across 3 areas, and none is in the top tenth of its board.
Where Kimi K2 Thinking stands
On our aggregate score, Kimi K2 Thinking is #401 of 743 ranked models, around the middle of the table. Its score is 47.8; the scale is centred on 50, so it sits 2.2 points below the typical model. That beats 46% of the ranked field. Among the 13 ranked Moonshot models it is #10. The next one up is Kimi K2.6 at 48.4. The next one down is Kimi K2.7 Code at 47.3.
Kimi K2 Thinking, benchmark by benchmark
Epoch ECI puts Kimi K2 Thinking at 146, #25 of 204. That beats 88% of the models tested, but 10 points short of Qwen 3.8 Max (156). The results either side are nemotron-3-ultra (146) and MiniMax-M2.7 (146). It covers Epoch Capabilities Index: Epoch's own aggregate of ~40 benchmarks. The board median is 127, so Kimi K2 Thinking is 19 points above it. Source: Epoch AI Benchmarking Hub, captured 2026-10-13.
On GPQA Diamond, Kimi K2 Thinking reaches 84.5%. It ranks #84 of 275, ahead of 69% of the field, and 11.3 points behind GPT-6 Astra (max) at 95.8%. The results either side are Qwen3.5 Plus (84.8%) and Darwin-9B-NEG (84.3%). It covers expert-level science questions. The board median is 71.7%, so Kimi K2 Thinking is 12.8 points above it. Source: HuggingFace GPQA Diamond leaderboard, captured 2026-10-13.
Kimi K2 Thinking manages 63.4% on SWE-bench Verified, good for #28 of 73. Claude Opus 4.7 (max) leads at 83.5%, 20.1 points further up, and 62% of models sit below. The results either side are o1-preview (64.6%) and Claude 3.7 Sonnet w/ Review Heavy (62.4%). It covers real GitHub issues resolved. The board median is 56.4%, so Kimi K2 Thinking is 7.0 points above it. Source: SWE-bench Verified leaderboard, captured 2026-10-13.
METR time horizon puts Kimi K2 Thinking at 54 min, #12 of 31. That beats 61% of the models tested, but 991 minutes short of Claude Mythos Preview (Early) (1,045 min). The results either side are claude-3-7-sonnet-20250219 (60 min) and Gemini 2.5 Pro Preview (Jun 2025) (39 min). It covers length of task (minutes) done at 50% success. The board median is 23 min, so Kimi K2 Thinking is 31 minutes above it. Source: Epoch AI Benchmarking Hub, captured 2026-10-13.
On WeirdML, Kimi K2 Thinking reaches 42.8%. It ranks #50 of 109, ahead of 54% of the field, and 50.5 points behind GPT-6 Astra (max) at 93.3%. The results either side are grok-4-fast (42.9%) and Grok 3 mini (42.6%). It covers unusual machine-learning tasks. The board median is 41.2%, so Kimi K2 Thinking is 1.6 points above it. Source: Epoch AI Benchmarking Hub, captured 2026-10-13.
Kimi K2 Thinking lands below the middle on Terminal-Bench 2.0: 35.7%, #16 of 28, with 54% of the field scoring higher. Gemini 3 Pro Preview leads at 69.4%; the closest result above is MiniMax-M2.1 at 36.6%. It covers real terminal tasks (agent + model). The board median is 37.3%, so Kimi K2 Thinking is 1.6 points below it. Source: Epoch AI Benchmarking Hub, captured 2026-10-13.
Kimi K2 Thinking lands below the middle on ALE-Bench: 598, #55 of 79, with 68% of the field scoring higher. GPT-6 Astra (max) leads at 2,951; the closest result above is MiniMax-M2.7 at 599. It covers AtCoder heuristic-optimisation contests (performance rating). The board median is 764, so Kimi K2 Thinking is 166 points below it. Source: Epoch AI Benchmarking Hub, captured 2026-10-13.
CL-bench is a weak spot. Kimi K2 Thinking scores 11.9%, #15 of 16, behind 88% of the models measured. The leader, GPT-5.4 (xhigh), reaches 27.9%. It covers learning new rules from context. The board median is 17.7%, so Kimi K2 Thinking is 5.8 points below it. Source: Epoch AI Benchmarking Hub, captured 2026-10-13.
Strengths and weaknesses
Software engineering: uneven, ahead of 45% of the field on average across 3 benchmarks. Its best board here is SWE-bench Verified (#28 of 73) and its weakest is ALE-Bench (#55 of 79). General knowledge and chat: solid, ahead of 52% of the field on average across 4 benchmarks. Its best board here is Epoch ECI (#25 of 204) and its weakest is CL-bench (#15 of 16).
Price and value
The median input price across the 427 priced models is $0.500 per million tokens, so Kimi K2 Thinking costs more 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 ernie-5.1 ($0.563) and kimi-k2-0711-preview ($0.570).
Reading these numbers
The 8 scores on this page come from 3 sources: Epoch AI Benchmarking Hub (6), HuggingFace GPQA Diamond leaderboard (1), SWE-bench Verified leaderboard (1). 1 of them come from an official leaderboard and 7 from datasets or third-party runs.
Models ranked nearby
- #398 Opus 4.5 (Thinking, None) (47.9)
- #399 glm-4p7 (47.9)
- #400 gemini-2.0-flash-thinking-exp-01-21 (47.9)
- #402 GPT-5.1 (Thinking, Medium) (47.8)
- #403 Tiny Recursion Model (TRM) (47.8)
- #404 DeepSeek-V2-Chat (47.8)
More from Moonshot
- Kimi K3 (max) (#15, 55.7)
- Kimi K2.5 (Fireworks) (#50, 52.8)
- Kimi K2.5 (#173, 50.7)
- Kimi K2 Thinking Turbo (#187, 50.5)
- Kimi K2 Thinking (Together) (#240, 49.7)
- Kimi-K2-Instruct-0905 (#258, 49.5)
Compare it on the live leaderboard → · How the aggregate score works