BeamvsInkling
Beam 与 Inkling 有 9 个共同 benchmark，但两者公开成绩的测试模式不同，下方分数仅供参考，不判定胜负。
测试模式不同的成绩
共 9 项。两款模型的公开成绩来自不同测试模式，分数不直接可比，不计入胜负和平均分差。
| 评测项 | Beam | Inkling |
|---|---|---|
| Terminal-Bench 2.1 | 80.10Unspecified Reasoning (With Tools) | 63.80Thinking (With Tools) |
| BrowseComp | 77.40Unspecified Reasoning (With Tools) | 77.10Thinking (With Tools + Internet) |
| AIME 2026 | 97.80Reported best (effort unspecified) | 97.10Thinking (No Tools) |
| SWE-Bench Pro V2 Hard | 77.20Unspecified Reasoning (With Tools) | 56.90Extra-High (With Tools) |
| SWE-bench Verified | 80.90Unspecified Reasoning (With Tools) | 77.60Thinking (With Tools) |
| CritPt | 16.30Reported best (effort unspecified) | 5.40Extra-High (No Tools) |
| GPQA Diamond | 90.50Reported best (effort unspecified) | 88.26Extra-High (No Tools) |
| τ³-Banking | 38Unspecified Reasoning (With Tools) | 29.10Extra-High (With Tools) |
| MCP-Atlas | 78.70Unspecified Reasoning (With Tools) | 76Extra-High (With Tools) |
规格对比
| 字段 | Beam | Inkling |
|---|---|---|
| 发布机构 | Reflection AI | Thinking Machines Lab |
| 发布时间 | 2026-10-05 | 2026-07-15 |
| 模型类型 | 推理大模型 | 多模态大模型 |
| 架构 | MoE 架构 | MoE 架构 |
| 参数规模 | 5010亿 | 9750亿 |
| 上下文长度 | 256K | 1M |
| 最大输出 | 128K | 暂无数据 |
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 | Beam | Inkling |
|---|---|---|
| 文本输入 | 暂无公开价格 | $3.74 / 1M tokens |
| 文本输出 | 暂无公开价格 | $9.36 / 1M tokens |
| 缓存读取 | 暂无公开价格 | $0.748 / 1M tokens |
部分模型公开价格不完整，缺失字段按"暂无公开价格"展示。
小结
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。