GPT-5.6 SolvsClaude Fable 5
在 28 个共同 benchmark 中，Claude Fable 5 整体领先：GPT-5.6 Sol 领先 10 项，Claude Fable 5 领先 17 项，持平 1 项，平均分差 -0.18。
GPT-5.6 Sol
OpenAI · 2026-06-26 · 推理大模型
Claude Fable 5
Anthropic · 2026-06-09 · 推理大模型
GPT-5.6 Sol10 项(36%)持平1(61%)17 项Claude Fable 5
评测分数
按能力类目分组，每组内按分差大小排列；共 28 项。
自主开发与终端任务
胶着 4/4| 评测项 | GPT-5.6 Sol | Claude Fable 5 | 分差 |
|---|---|---|---|
| Terminal-Bench 4.0 | 37.2724 / 98Max (With Tools) | 44.5518 / 98Max (With Tools) | -7.28 |
| CursorBench 3.2 | 67.204 / 5Max (With Tools) | 70.502 / 5Max (With Tools) | -3.30 |
| Terminal Bench Hard | 65.901 / 244Max (With Tools) | 62.902 / 244Max (With Tools) | +3 |
| Terminal-Bench 3.0 | 34.402 / 11Max (With Tools) | 34.103 / 11Max (With Tools) | +0.30 |
法律
Claude Fable 5 领先 3/3| 评测项 | GPT-5.6 Sol | Claude Fable 5 | 分差 |
|---|---|---|---|
| Harvey's Legal Agent Benchmark | 2.5029 / 43Max (With Tools) | 11.258 / 43Max (With Tools) | -8.75 |
| Harvey Lab-AA | 87.1817 / 44Max (With Tools) | 93.563 / 44Max (With Tools) | -6.38 |
| Legal Research Bench | 48.087 / 42Max (With Tools) | 49.525 / 42Max (With Tools) | -1.44 |
跨应用与工具编排
GPT-5.6 Sol 领先 2/3| 评测项 | GPT-5.6 Sol | Claude Fable 5 | 分差 |
|---|---|---|---|
| PinchBench v2 | 84.197 / 45Reported best (effort unspecified) | 54.8435 / 45Reported best (effort unspecified) | +29.35 |
| MCPMark-Verified | 92.912 / 8Max (With Tools) | 86.614 / 8Max (With Tools) | +6.30 |
| APEX-Agents | 56.703 / 7Max (With Tools) | 59.201 / 7Max (With Tools) | -2.50 |
临床工作与医疗决策
Claude Fable 5 领先 2/2| 评测项 | GPT-5.6 Sol | Claude Fable 5 | 分差 |
|---|---|---|---|
| MedCode | 43.9732 / 64Max (With Tools) | 56.073 / 64Max (With Tools) | -12.10 |
| MedScribe | 85.2320 / 66Max (With Tools) | 88.529 / 66Max (With Tools) | -3.29 |
客服与业务流程
胶着 2/2| 评测项 | GPT-5.6 Sol | Claude Fable 5 | 分差 |
|---|---|---|---|
| Public Benefits Bench v1.1 | 66.5110 / 26Max (With Tools) | 70.434 / 26Max (With Tools) | -3.92 |
| SAGE | 52.564 / 64Max (With Tools) | 51.897 / 64Max (With Tools) | +0.67 |
数学
胶着 2/2| 评测项 | GPT-5.6 Sol | Claude Fable 5 | 分差 |
|---|---|---|---|
| ProofBench v1.1 (Lean 4) | 837 / 28Max (With Tools) | 955 / 28Max (With Tools) | -12 |
| FrontierMath v2 | 89.121 / 58Max (No Tools) | 87.023 / 58Max (No Tools) | +2.11 |
机器学习工程
GPT-5.6 Sol 领先 1/2| 评测项 | GPT-5.6 Sol | Claude Fable 5 | 分差 |
|---|---|---|---|
| WeirdML v2 | 88.762 / 52Thinking High (With Tools) | 87.853 / 52Thinking High (With Tools) | +0.91 |
| WeirdML v3 | 0.157 / 13Extra-High (With Tools) | 0.166 / 13Extra-High (With Tools) | 持平 |
金融
Claude Fable 5 领先 2/2| 评测项 | GPT-5.6 Sol | Claude Fable 5 | 分差 |
|---|---|---|---|
| Finance Agent v2 | 53.7621 / 43Max (With Tools) | 56.3113 / 43Max (With Tools) | -2.55 |
| EMB (Excel Modeling) | 72.345 / 42Max (With Tools) | 73.673 / 42Max (With Tools) | -1.33 |
专业数据分析
Claude Fable 5 领先 1/1| 评测项 | GPT-5.6 Sol | Claude Fable 5 | 分差 |
|---|---|---|---|
| AA-AnalystAgent | 47.507 / 29Max (With Tools + Internet) | 48.756 / 29Max (With Tools + Internet) | -1.25 |
写作与创意表达
GPT-5.6 Sol 领先 1/1| 评测项 | GPT-5.6 Sol | Claude Fable 5 | 分差 |
|---|---|---|---|
| Creative Writing | 1,9729 / 110Normal (No Tools) | 1,94310 / 110Normal (No Tools) | +28.50 |
多轮记忆与持续上下文
GPT-5.6 Sol 领先 1/1| 评测项 | GPT-5.6 Sol | Claude Fable 5 | 分差 |
|---|---|---|---|
| EBR-bench | 44.766 / 23Max (With Tools) | 39.527 / 23Max (With Tools) | +5.24 |
抽象归纳与泛化
Claude Fable 5 领先 1/1| 评测项 | GPT-5.6 Sol | Claude Fable 5 | 分差 |
|---|---|---|---|
| ARC-AGI-1 | 97.5011 / 193Extra-High (No Tools) | 98.501 / 193Extra-High (No Tools) | -1 |
科学计算与科研编程
GPT-5.6 Sol 领先 1/1| 评测项 | GPT-5.6 Sol | Claude Fable 5 | 分差 |
|---|---|---|---|
| Terminal-Bench-Science 0.1 | 22.406 / 18Max (With Tools) | 21.407 / 18Max (With Tools) | +1 |
程序生成与编辑
Claude Fable 5 领先 1/1| 评测项 | GPT-5.6 Sol | Claude Fable 5 | 分差 |
|---|---|---|---|
| Vibe Code Bench v1.1 | 80.5017 / 61Max (With Tools) | 90.352 / 61Max (With Tools) | -9.85 |
维护、优化与工程管理
Claude Fable 5 领先 1/1| 评测项 | GPT-5.6 Sol | Claude Fable 5 | 分差 |
|---|---|---|---|
| Code Migration | 52.927 / 43Max (With Tools) | 55.065 / 43Max (With Tools) | -2.14 |
跨能力聚合指数
Claude Fable 5 领先 1/1| 评测项 | GPT-5.6 Sol | Claude Fable 5 | 分差 |
|---|---|---|---|
| Vals Index v2.1 | 58.019 / 36Max (task-specific tools) | 61.397 / 36Max (task-specific tools) | -3.38 |
规格对比
| 字段 | GPT-5.6 Sol | Claude Fable 5 |
|---|---|---|
| 发布机构 | OpenAI | Anthropic |
| 发布时间 | 2026-06-26 | 2026-06-09 |
| 模型类型 | 推理大模型 | 推理大模型 |
| 架构 | 稠密模型 | 稠密模型 |
| 参数规模 | 暂无数据 | 暂无数据 |
| 上下文长度 | 1.05M | 1M |
| 最大输出 | 128K | 128K |
API 调用价格
价格优先使用 DataLearner 配置的 API 记录；缺失项不做推测。
| 价格项 | GPT-5.6 Sol | Claude Fable 5 |
|---|---|---|
| 文本输入 | $4 / 1M tokens | $10 / 1M tokens |
| 文本输出 | $20 / 1M tokens | $50 / 1M tokens |
| 缓存读取 | $0.4 / 1M tokens | $1 / 1M tokens |
| 缓存写入 | $5 / 1M tokens | $12.5 / 1M tokens |
小结
- GPT-5.6 Sol在以下类目领先:跨应用与工具编排 (2/3)、机器学习工程 (1/2)、写作与创意表达 (1/1)、多轮记忆与持续上下文 (1/1)、科学计算与科研编程 (1/1)
- Claude Fable 5在以下类目领先:法律 (3/3)、临床工作与医疗决策 (2/2)、金融 (2/2)、专业数据分析 (1/1)、抽象归纳与泛化 (1/1)、程序生成与编辑 (1/1)、维护、优化与工程管理 (1/1)、跨能力聚合指数 (1/1)
- 胶着类目:自主开发与终端任务、客服与业务流程、数学
28 个共同 benchmark 上，Claude Fable 5 平均高出 0.18 分。
单项差距最大的 benchmark：PinchBench v2 — GPT-5.6 Sol 84.19，Claude Fable 5 54.84（分差 +29.35）。
本页正文由结构化模型、价格与 benchmark 数据生成，不使用实时 LLM 撰写。