Kagi LLM Benchmark leaderboard
- Claude Fable 5 leads
- 91.4%
- models measured
- 68
- lead over #2
- +2.6 pts
- median
- 55.9%
Leader on Kagi LLM Benchmark · read October 21
On October 21, 2026, 70 of our catalog's models had a Kagi LLM Benchmark result, and Anthropic Claude Fable 5 led with 91.4%, ahead of Anthropic Claude Opus 4.8 at 88.8%.
Kagi runs an unpublished set of reasoning, coding and instruction-following tasks, including non-English languages and noisy context, on models used without tools such as web search. The task set was last revised on February 13, 2026 and changes over time. The number is the accuracy Kagi reports for each model.
Role in the index
Limit: Kagi notes that models that use chain of thought do drastically better in this benchmark (read October 21, 2026).
Kagi LLM Benchmark is shown as published and does not feed any of our indexes. Read the methodology
| # | Model | Result | Cost of the suite | Evidence | Coding index | Output $/Mtok |
|---|---|---|---|---|---|---|
| 1 | Claude Fable 5Anthropic | 91.4% | $10.40 | 79.3 | $50.00 | |
| 2 | Claude Opus 4.8Anthropic | 88.8% | $8.80 | 65.0 | $25.00 | |
| 3 | GPT-5.5OpenAI | 88.8% | $2.20 | 71.3 | $30.00 | |
| 4 | Claude Opus 4.6Anthropic | 83.6% | $4.80 | 57.2 | $25.00 | |
| 5 | Grok 4.5xAI | 83.5% | $1.00 | 58.6 | $6.00 | |
| 6 | Claude Opus 4.7Anthropic | 80.7% | $2.80 | 67.5 | $25.00 | |
| 7 | Claude Opus 4.5Anthropic | 80.2% | $4.50 | 44.0 | $25.00 | |
| 8 | Kimi K2.5Moonshot AI | 78.5% | $0.50 | 38.6 | $2.85 | |
| 9 | GPT-5 ProOpenAI | 76.8% | $31.80 | Not benchmarked | $120.00 | |
| 10 | Grok 4.20xAI | 75.0% | $1.30 | 33.3 | $2.50 | |
| 11 | GLM 5Z.ai | 75.0% | $2.00 | 41.7 | $3.20 | |
| 12 | GPT-5.2OpenAI | 73.3% | $1.40 | 55.9 | $14.00 | |
| 13 | GPT-5OpenAI | 72.7% | $7.10 | 38.8 | $10.00 | |
| 14 | o3 ProOpenAI | 72.1% | $34.20 | Not benchmarked | $80.00 | |
| 15 | Gemini 2.5 ProGoogle | 70.3% | $1.70 | 19.1 | $10.00 | |
| 16 | GPT-5 MiniOpenAI | 70.3% | $4.90 | 34.2 | $2.00 | |
| 17 | R1RetiredDeepSeek | 69.4% | $9.90 | Not benchmarked | $2.50 | |
| 18 | o3OpenAI | 67.6% | $4.80 | Not benchmarked | $8.00 | |
| 19 | o4 MiniOpenAI | 67.6% | $3.10 | Not benchmarked | $4.40 | |
| 20 | Gemini 3.1 Flash LiteGoogle | 67.2% | $0.00 | Not benchmarked | $1.50 | |
| 21 | GPT-5.6 SolOpenAI | 67.0% | $2.70 | 73.5 | $20.00 | |
| 22 | Qwen3 Next 80B A3B InstructQwen | 66.7% | $1.00 | Not benchmarked | $1.10 | |
| 23 | Kimi K2 0711Moonshot AI | 64.4% | $0.80 | Not benchmarked | $2.30 | |
| 24 | GPT-5.4OpenAI | 63.8% | $0.30 | 64.4 | $15.00 | |
| 25 | Gemma 4 31BGoogle | 63.5% | $1.00 | 39.0 | $0.40 | |
| 26 | GLM 5.2Z.ai | 62.6% | $0.70 | 61.1 | $4.40 | |
| 27 | GPT-5 NanoOpenAI | 62.2% | $0.40 | Not benchmarked | $0.40 | |
| 28 | Grok 3RetiredxAI | 61.3% | $2.60 | Not benchmarked | $15.00 | |
| 29 | GLM 4.5VZ.ai | 59.8% | $0.70 | Not benchmarked | $1.80 | |
| 30 | gpt-oss-120bOpenAI | 58.6% | $0.40 | Not benchmarked | $0.50 | |
| 31 | Claude Sonnet 4.5Anthropic | 57.9% | $1.00 | 39.5 | $15.00 | |
| 32 | GLM 4.5Z.ai | 57.9% | $1.00 | Not benchmarked | $2.20 | |
| 33 | MiniMax M2MiniMax | 57.8% | $0.10 | Not benchmarked | $1.02 | |
| 34 | DeepSeek V3.1 TerminusDeepSeek | 57.4% | $0.20 | Not benchmarked | $1.0142 | |
| 35 | Gemini 2.5 FlashGoogle | 56.8% | $0.50 | Not benchmarked | $2.50 | |
| 36 | Llama 4 MaverickMeta Llama | 55.9% | $0.20 | Not benchmarked | $1.00 | |
| 37 | N Hermes 4 405BNous Research | 55.8% | $2.60 | Not benchmarked | $3.00 | |
| 38 | MiniMax M2.5MiniMax | 55.2% | $1.90 | 37.2 | $1.20 | |
| 39 | DeepSeek V4 Pro 0423DeepSeek | 53.5% | $0.20 | 38.7 | $2.716 | |
| 40 | DeepSeek V3.1DeepSeek | 53.2% | $0.40 | Not benchmarked | $1.50 | |
| 41 | gpt-oss-20bOpenAI | 53.2% | $0.50 | Not benchmarked | $0.15 | |
| 42 | DeepSeek V3DeepSeek | 52.3% | $0.40 | Not benchmarked | $0.9594 | |
| 43 | GPT-4.1OpenAI | 52.3% | $2.10 | Not benchmarked | $8.00 | |
| 44 | DeepSeek V3.2 ExpDeepSeek | 52.2% | $0.20 | Not benchmarked | $0.41 | |
| 45 | DeepSeek V4 Flash 0423DeepSeek | 52.2% | $0.00 | Not benchmarked | $0.28 | |
| 46 | Kimi K2 0905Moonshot AI | 52.2% | $0.30 | Not benchmarked | $2.50 | |
| 47 | GPT-5.6 TerraOpenAI | 51.3% | $2.50 | 63.9 | $12.00 | |
| 48 | Qwen3 Coder 480B A35BQwen | 49.5% | $0.80 | Not benchmarked | $1.50 | |
| 49 | GPT-5.6 LunaOpenAI | 49.1% | $1.00 | 58.0 | $1.20 | |
| 50 | Qwen3 14BQwen | 49.1% | $0.50 | Not benchmarked | $0.23 | |
| 51 | Qwen3 32BQwen | 48.7% | $0.50 | Not benchmarked | $0.40 | |
| 52 | GPT-4.1 MiniOpenAI | 48.6% | $0.40 | Not benchmarked | $1.60 | |
| 53 | GLM 4.6Z.ai | 45.7% | $0.10 | 8.8 | $2.20 | |
| 54 | Nova Premier 1.0Amazon | 44.8% | $0.40 | Not benchmarked | $12.50 | |
| 55 | GLM 4.5 AirZ.ai | 43.0% | $0.20 | Not benchmarked | $1.10 | |
| 56 | MiniMax-01MiniMax | 42.5% | $0.10 | Not benchmarked | $1.10 | |
| 57 | Mistral Medium 3.5Mistral AI | 41.4% | $0.40 | 25.5 | $7.50 | |
| 58 | Gemini 2.5 Flash LiteGoogle | 40.5% | $0.10 | Not benchmarked | $0.40 | |
| 59 | Mistral Small 4Mistral AI | 40.5% | $0.00 | Not benchmarked | $0.60 | |
| 60 | GPT-5.4 NanoOpenAI | 39.7% | $0.00 | 44.8 | $1.25 | |
| 61 | GPT-5.4 MiniOpenAI | 37.9% | $0.10 | 35.8 | $4.50 | |
| 62 | Llama 4 ScoutMeta Llama | 36.9% | $0.10 | Not benchmarked | $0.50 | |
| 63 | Qwen2.5 VL 72B InstructQwen | 36.0% | $0.00 | Not benchmarked | $1.00 | |
| 64 | Gemma 3 27BGoogle | 35.1% | $0.10 | Not benchmarked | $0.30 | |
| 65 | GPT-4.1 NanoOpenAI | 33.3% | $0.10 | Not benchmarked | $0.40 | |
| 66 | Codestral 2508Mistral AI | 32.5% | $0.10 | Not benchmarked | $0.90 | |
| 67 | Command ACohere | 28.8% | $1.90 | Not benchmarked | $10.00 | |
| 68 | GPT-4o-miniOpenAI | 28.8% | $0.10 | Not benchmarked | $0.60 | |
| 69 | Gemma 3 4BGoogle | 25.2% | $0.00 | Not benchmarked | $0.075 | |
| 70 | UI-TARS 7BByteDance | 20.7% | $0.00 | Not benchmarked | $0.20 |
Only models in our catalog are listed, and each appears once with its best published run. 70 rows were read from the source on October 21, 2026, and rows that match no catalog model are left out. Cost of the suite is what Kagi reports it cost to run the entire benchmark suite end to end, in US dollars, not the cost of one task. It is mostly output tokens.