Model comparisonSo sánh model模型对比
Benchmarks
| Benchmark | Grok 4.6 | Claude Fable 5 | Claude Opus 5 | GPT-5.6 Sol | Kimi K3 | |
|---|---|---|---|---|---|---|
| Params (total / active) | – | – | – | – | – | |
| Context | 2M | 1M | 1M | 1.1M | 1.1M | |
| General | ||||||
| General | AA Intelligence Index | 61.0 | 59.9 | 58.9 | 57.7 | 57.1 |
| SimpleQA (verified) | 49.3 | 68.3 | 56.7 | 71.6 | 42.7 | |
| Coding | ||||||
| Coding | AA Coding Index | – | 76.5 | 76.5 | 78.3 | 76.2 |
| CursorBench | 41.4 | – | – | 41.7 | – | |
| SciCode | 56.5 | 60.2 | 54.3 | 56.0 | 58.7 | |
| Math / STEM | ||||||
| Math / STEM | GPQA Diamond | 94.0 | 92.6 | 93.7 | 93.1 | 93.5 |
| OTIS Mock AIME | 99.2 | 99.7 | 98.9 | 100.0 | 97.2 | |
| Reasoning | ||||||
| Reasoning | Humanity's Last Exam | – | 53.3 | 50.7 | 44.7 | 44.3 |
| ARC-AGI-2 | 67.1 | 89.2 | 88.3 | 92.5 | 60.4 | |
| Agentic / Tool use | ||||||
| Agentic / Tool use | Terminal-Bench Hard | – | 62.9 | – | 61.4 | – |
| Terminal-Bench | – | 84.6 | 87.6 | 89.5 | 85.0 | |
| τ²-bench Telecom | – | 98.5 | – | 84.8 | – | |
| τ-bench Banking | – | 26.8 | 32.8 | 32.6 | 33.4 | |
| Long context | ||||||
| Long context | AA-LCR | – | 70.0 | 67.0 | 71.0 | 74.7 |
| Instruction following | ||||||
| Instruction following | IFEval | – | 63.5 | – | 71.0 | – |
| Cost & speed | ||||||
| Cost & speed | Input $/1M↓ | $2.00 | $10.00 | $5.00 | $2.00 | $3.00 |
| Output $/1M↓ | $6.00 | $50.00 | $25.00 | $10.00 | $15.00 | |
Bold marks the best score in each row. ↓ means lower is better. – means no published score. Scores in percent unless the row is an index.
Benchmark data from Epoch AI (CC-BY 4.0).
x-ai/grok-4.6 Details | anthropic/claude-fable-5 Details | anthropic/claude-opus-5 Details | openai/gpt-5.6-sol Details | moonshotai/kimi-k3 Details | |
|---|---|---|---|---|---|
| Overview | |||||
| Description | SpaceXAI's smartest model with frontier performance on coding, knowledge work, and STEM. Builds on Grok 4.5 for long-running agents and visual work. | Mythos-class model built for autonomous knowledge work and coding with 1M-token context | Anthropic's Claude Opus 5 is the flagship model for demanding reasoning, coding, and long-horizon agentic work — end-to-end software tasks, code review and bug finding, visual analysis of charts and documents, and coordinating parallel subagents. | OpenAI's GPT-5.6 Sol, a reasoning-focused model in the GPT-5.6 family with a 1M+ token context window, image understanding, configurable reasoning effort, and tool-use support via the Responses API. | Kimi K3 is an open-source multimodal reasoning model from Moonshot AI with a 1M-token context window and native visual understanding. It accepts text, image, and video inputs and is built for long-horizon software engineering, knowledge work, and deep reasoning, with particular strength where code meets visual and spatial reasoning — frontend, game development, and CAD workflows. Its architecture uses KDA and Attention Residuals for computational efficiency, and its thinking mode is always on. |
| Category | Text Generation | Text Generation | Text Generation | Multimodal | Multimodal |
| Context length | 2M | 1M | 1M | 1.1M | 1.1M |
| Providers | 8 | 11 | 10 | 7 | 19 |
| Pricing (per 1M tokens) | |||||
| Input | $2.00 | $10.00 | $5.00 | $2.00 | $3.00 |
| Output | $6.00 | $50.00 | $25.00 | $10.00 | $15.00 |
| Capabilities | |||||
| Reasoning | – | – | – | ||
| Tool calling | |||||
| Vision | – | – | |||
| Streaming | |||||
| Token activity (30 days) | |||||
| Usage | |||||
| Artificial Analysis | |||||
| Intelligence | 61.0 | 59.9 | 58.9 | 57.7 | 57.1 |
| Coding | – | 76.5 | 76.5 | 78.3 | 76.2 |
| Benchmarks | |||||
| GPQA Diamond | 0.9 | 0.9 | 0.9 | 0.9 | 0.9 |
| HLE | – | 0.5 | 0.5 | 0.4 | 0.4 |
| ARC-AGI-2 | 0.7 | 0.9 | 0.9 | 0.9 | 0.6 |
| IFEval | – | 0.6 | – | 0.7 | – |
| τ²-bench Telecom | – | 1.0 | – | 0.8 | – |
| SciCode | 0.6 | 0.6 | 0.5 | 0.6 | 0.6 |
| OTIS Mock AIME | 1.0 | 1.0 | 1.0 | 1.0 | 1.0 |
| CursorBench | 0.4 | – | – | 0.4 | – |
| SimpleQA (verified) | 0.5 | 0.7 | 0.6 | 0.7 | 0.4 |
Benchmark data from Epoch AI (CC-BY 4.0).