GPT-6.1 Sol（OpenAI） GPT-6.1 Sol (OpenAI)

来源Sources

为什么推荐Why we recommend it 页内写 Astra 在最难科学任务上仍最高（Terminal-Bench Science 0.1 68.1%），故不把 “Near-Astra” 写成全面第一。同板互链 gpt-6-astra、astra-openai-showcase、claude-sonnet-5-5 只表示同板相关（Sonnet 官方表对照过 GPT-6 Sol），不表示评测对齐。整理质量 4 星（来源齐、边界清），不代表基准分。 The page says Astra is still highest on the hardest science (Terminal-Bench Science 0.1 at 68.1%), so “Near-Astra” is not framed as first everywhere. Cross-links to gpt-6-astra, astra-openai-showcase, and claude-sonnet-5-5 are same-board only (Sonnet’s vendor table already contrasts GPT-6 Sol), not an aligned-eval claim.

要点（官方称 / 社区称）Highlights (vendor / community claims)

• 官方称：GPT-6 Sol 的升级款，在 agentic coding、电脑操作与专业工作上几乎对齐 GPT-6 Astra，标准输入与输出 token 价格约为 Astra 的五分之一。标准 API：$2 / 百万输入 token、$0.10 / 百万缓存输入 token、$10 / 百万输出 token。缓存输入官方称比标准输入少 95%，比 GPT-6 Sol 的缓存输入少 50%。 Vendor claim: an upgrade to GPT-6 Sol that nearly matches GPT-6 Astra on agentic coding, computer use, and professional work at about one-fifth of Astra’s standard input and output token prices. Standard API: $2 per million input tokens, $0.10 per million cached input tokens, $10 per million output tokens. Cached input is claimed 95% less than standard input and 50% less than GPT-6 Sol’s cached input.
• 官方称（评测节选，本库未复跑）：DeepSWE v1.1 在约五分之一成本下对齐 Astra，并以更低 effort / 成本超出 GPT-6 Sol 最佳 6.4 个百分点；GDP.pdf 高于带 fallbacks 的 Opus 5.5，成本不到一半；AutomationBench 在 medium effort 比 Opus 5.5 高 2.2 个百分点、约三分之一成本，相对 GPT-6 Sol 高 4.8 个百分点（页内措辞）；OSWorld 2.0 离线 partial 在 max effort 比 GPT-6 Sol 高 7 个百分点、成本不到一半，距 Astra 在 2.1 个百分点以内；Terminal-Bench Science 0.1 在 max effort 超过 GPT-6 Sol 一倍以上、单任务成本不到一半，Astra 在最难科学任务上仍最高（68.1%）；难提示事实性：低 effort 含事实错误的回答占比相对 GPT-6 Sol 从 11.4% 降到 7.7%。 Vendor claim (eval excerpts, not re-run here): on DeepSWE v1.1 it matches Astra at about one-fifth the cost and eclipses GPT-6 Sol’s best by 6.4 percentage points at lower effort and cost; on GDP.pdf it scores higher than Opus 5.5 with fallbacks at less than half the cost; on AutomationBench it is 2.2 percentage points above Opus 5.5 at medium effort and about one-third the cost, and 4.8 percentage points above GPT-6 Sol (page wording); on OSWorld 2.0 offline partial it is 7 percentage points above GPT-6 Sol at max effort at less than half the cost, and within 2.1 percentage points of Astra; on Terminal-Bench Science 0.1 it more than doubles GPT-6 Sol at max effort at less than half the cost per task, while Astra remains highest on the hardest science at 68.1%; factuality on hard prompts: the low-effort share of responses with a factual error falls from 11.4% to 7.7% versus GPT-6 Sol.
• 官方称：即日起向 Plus / Pro / Business / Enterprise / Edu 开放，位于 ChatGPT Work 与 Codex；尚未进入 Chat。开发者可通过 OpenAI API 使用，模型 id 为 `gpt-6.1-sol`，API 已可用。Ultrafast 将在随后数日推出，官方称在 Codex 中相对标准速度最高快 8 倍 token 生成。 Vendor claim: available starting today to Plus / Pro / Business / Enterprise / Edu users in ChatGPT Work and Codex; not yet in Chat. Developers can use it on the OpenAI API as gpt-6.1-sol; the API is available. Ultrafast is coming in the next few days, claimed at up to 8× faster token generation versus standard speed in Codex.
• 第三方榜（Artificial Analysis，非官方、非本库）：GPT-6.1 Sol (max) Intelligence Index 52（#10/222），输出 67.8 tok/s（#107/222），价输入 $2 / 输出 $10、Cache Discount 95%、每 Intelligence Index 任务 $0.72，1M 上下文，发布日 2026-09-29。同站文章称 GPT-6 Sol 在 7 天后换成 6.1 Sol，约比 Astra II 低 1 分、每任务成本不到四分之一。不把 AA 数字与官方页评测合成新指标。 Third-party board (Artificial Analysis — neither vendor nor this library): GPT-6.1 Sol (max) Intelligence Index 52 (#10/222), 67.8 tok/s (#107/222), $2 input / $10 output, Cache Discount 95%, $0.72 per Intelligence Index task, 1M context, released 2026-09-29. The same site’s article says GPT-6 Sol was replaced by 6.1 Sol after 7 days, about 1 point below Astra II at under a quarter of the cost per task. AA figures are not merged with the vendor evals into a new metric.