Head-to-head正面对比
Claude Haiku 4.5 vs GPT-5.6 Luna
Anthropic meets OpenAI: sourced benchmark scores side by side, dimension by dimension, with prices and specs. Every conclusion below is derived from the published dataset — no opinions, no sponsorship.
Anthropic 对 OpenAI:带来源的 benchmark 成绩逐项并排,附价格与规格。下面每条结论都从公开数据集推导——没有主观意见,没有厂商赞助。
◈
The verdict, derived from data结论(全部由数据推导)
- GPT-5.6 Luna leads overall: 83.3 vs 51 on the Intelligence Score.总分上 GPT-5.6 Luna 领先:83.3 对 51。
- GPT-5.6 Luna is stronger in reasoning, coding, math, preference.GPT-5.6 Luna 在推理、代码、数学、偏好上更强。
- GPT-5.6 Luna is 4.3× cheaper on blended price ($0.700 vs $3.00 per 1M tokens, in/out average).混合价(输入输出均值)上 GPT-5.6 Luna 便宜 4.3 倍:$0.700 对 $3.00/百万 token。
- Long documents: GPT-5.6 Luna has the much larger context window (1.1M vs 200K tokens).长文档选 GPT-5.6 Luna:上下文窗口大得多(1.1M 对 200K token)。
● Claude Haiku 4.5 ● GPT-5.6 Luna · missing dimensions count as 0 in the chart缺数据的维度在图上按 0 计
≡
Spec by spec逐项规格
| Claude Haiku 4.5 | GPT-5.6 Luna | |
|---|---|---|
| Provider厂商 | Anthropic | OpenAI |
| Released发布 | 2025-10-15 | 2026-07-09 |
| Intelligence Score智能评分 | 51 | 83.3 |
| Benchmark coveragebenchmark 覆盖 | 95% | 68% |
| Reasoning推理 | 55.8 | 80.1 |
| Coding代码 | 59.1 | 91.9 |
| Knowledge知识 | 68.4 | — |
| Math数学 | 21.9 | 94.5 |
| Agent智能体 | 39.9 | — |
| Preference偏好 | 28.5 | 38.2 |
| API input $/1M输入价 $/百万 | $1.00 | $0.200 |
| API output $/1M输出价 $/百万 | $5.00 | $1.20 |
| Blended $/1M混合价 $/百万 | $3.00 | $0.700 |
| Value (score per $)性价比(分数/美元) | 17 | 119 |
| Context window上下文 | 200K | 1.1M |
| Max output最大输出 | 64K | 128K |
| Reasoning tiers推理档位 | low / medium / high | none / minimal / low / medium / high / xhigh |
| Open weights开放权重 | No否 | No否 |
| CN-direct国内直连 | No否 | No否 |
| Free tier免费层 | No否 | No否 |
⚔
Benchmark by benchmark逐项 benchmark
| Benchmark基准 | Claude Haiku 4.5 | GPT-5.6 Luna |
|---|---|---|
| Humanity's Last Exam | 25.2 | 37.2 |
| GPQA Diamond | 64.6 | 91.1⚡max |
| SWE-bench Verified | 73.3 | 93 |
| AIME 2025 | 39 | 100 |
| FrontierMath | 2.083 | 78.6 |
| Terminal-Bench | 27.5 | 75.7 |
| LMArena (Chatbot Arena) | 1413 | 1452 |
Claude Haiku 4.5 dossier → 档案 → · GPT-5.6 Luna dossier → 档案 → · How we score评分方法