[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-openai-gpt-6-sol-luna-api-pricing":3,"topics-all":41,"news-related-a574d32a-c8f7-4d46-ac95-b76406f36b3b":60},{"id":4,"title":5,"summary":6,"content":7,"original_url":8,"source_id":9,"tags":10,"translations":27,"news_slug":34,"published_at":35,"created_at":36,"modified_at":37,"is_published":38,"publish_type":39,"image_url":14,"view_count":40},"a574d32a-c8f7-4d46-ac95-b76406f36b3b","OpenAI 把 GPT-6 拆成两档:Sol\u002FLuna 同款训练法,API 价格砍一半","OpenAI 把 GPT-6 系列拆出 Sol 和 Luna 两档,沿用 Astra 的训练方法,API 价格较上一代直接砍 50%,在 AutomationBench、DeepSWE、OSWorld 等专业工作评测上以显著更低的成本追平或跑赢 Claude Opus 5 \u002F Fable 5。","这次 GPT-6 家族扩列,OpenAI 把\"前沿智能\"按价格拆成两档——Sol 和 Luna。两者都用了与 Astra 相近的训练方法,但 API 价格相比上一代 Sol\u002FLuna 直接砍掉一半。\n\n## 价格直接砍一半\n\n先把最实在的账单摆出来。GPT-6 Sol 输入价从 5.6 代的 4 美元\u002F百万 token 降到 2 美元,输出从 20 美元降到 10 美元。GPT-6 Luna 更激进,输入从 0.20 美元降到 0.10 美元,输出从 1.20 美元降到 0.50 美元。OpenAI 在官方公告里给出的理由是缓存与推理基础设施的改进,把这些节省直接转嫁给用户。\n\n## 五项专业工作评测\n\nbenchmark 层面,官方选了 AutomationBench、Agents' Last Exam、FrontierCode、DeepSWE 1.1、OSWorld 2.0 offline 五个与\"复杂专业工作\"挂钩的评测,核心信息是\"在显著低于对手的价位上达到接近对手的水平\"。\n\nAutomationBench 1.0.6(覆盖销售、市场、运营、客服、财务、HR 等 47 个工具的端到端工作流)上,GPT-6 Sol xhigh effort 拿到 33.2%,每任务成本 0.27 美元;作为对比 Claude Opus 5 max effort 拿 26.9%,成本是 Sol 的 11.1 倍;Claude Fable 5.1 加 Opus 5 fallback 拿到 31.4%,成本是 Sol 的 8.9 倍以上。Agents' Last Exam V1 上,GPT-6 Sol max effort 拿到 56.4%,比 Claude Opus 5 最高分更高,而每任务成本低 60%。\n\n编码方面,FrontierCode 1.1 Main(评估能否产出可合并到真实代码库的改动,不仅看正确性,还看测试质量、范围控制、代码风格等)上,GPT-6 Sol 比 GPT-5.6 Sol 进步明显,并能以更低成本匹配 Claude Fable 5.1 xhigh。DeepSWE 1.1 上,GPT-6 Sol max effort 拿到 68.8%,距离 Claude Fable 5 在该评测中的最高分 69.9% 只差 1.1 个百分点,而每任务成本约为其 20%。GPT-6 Luna max effort 拿到 66.6%,与 Claude Opus 5 和 Fable 5 的 medium effort 表现相当,但成本只有 Opus 5 的 7%、Fable 5 的 4%。OpenAI 在公告里直接点名 DeepSWE 1.1 是\"针对真实代码库中的复杂软件工程任务的原创长程任务\"。\n\ncomputer use 方面,OSWorld 2.0 offline 上,GPT-6 Sol xhigh 拿到 60.5%,Claude Opus 5 medium 60.3%,Sol 成本约为 Opus 5 的 20%。GPT-6 Luna max 跑赢 GPT-5.6 Sol medium,成本只有后者的十分之一。\n\n合作风格上,官方说 Sol 和 Luna 也接住了 Astra 的对话风格改进了——更清晰、更少行话、更少模糊的措辞、更少无关细节,总体稍短但不损失信息量。文中给了一个对照示例:同一段把网站改成 bento grid 的请求,GPT-5.6 Sol 直接拍板\"不必 React\",GPT-6 Sol 则先核对再动手,并明确说明自己查过什么、没查什么。\n\n## 缓存改进做成产品级特性\n\n更值得开发者关注的是缓存层。OpenAI 同时发布了一篇专门的\"GPT-6 缓存改进\"文章,核心改动是默认更高的缓存命中率、对缓存输入 token 提供 90% 折扣,并上线 Prompt Caching Dashboard、诊断工具、显式 breakpoint API,允许在不破坏缓存的前提下调整推理强度和工具可用性。GitHub 反馈过去几个月 OpenAI 模型数十亿次请求中\"需要全新处理的 prompt token 占比下降超过 50%\",Copilot 响应因此更快。\n\n## 对齐与可用性\n\n对齐方面,官方称 Sol 和 Luna 沿用 Astra 的对齐工作,在内部编码欺骗评估上较 GPT-5.6 系列有所改进,误述自身编码工作的比率更低。\n\n可用性上,GPT-6 Sol 和 Luna 当天进入 ChatGPT Work 与 Codex,覆盖 Plus、Pro、Business、Enterprise、Edu 全档,免费和 Go 用户可以在桌面应用里用到 Luna。OpenAI 计划在一天内逐步放量,API 端以 `gpt-6-sol` 和 `gpt-6-luna` 提供。\n\n值得展开的反而是这次发布的\"产品哲学\":Astra 仍然是旗舰,但 OpenAI 显然想把\"接近旗舰的能力\"以更便宜的方式铺到日常任务里。AutomationBench 上 Sol 跑赢 Astra low effort,DeepSWE 上 Luna 用 4% 的成本达到 Opus 5 medium effort——这意味着\"贵模型 + 高 effort\"不再是完成复杂任务的唯一组合,性价比组合已经可以替代不少过去必须上旗舰的负载。对企业来说,这次价格调整直接重置了 Claude Opus \u002F Fable vs GPT-6 的 TCO 算式;对个人开发者,Luna 这档几乎是把门槛再往下降一截。\n\n所以呢:OpenAI 这一轮不是单纯发新模型,而是把 GPT-6 系列的价格曲线整体下移 50%,并把缓存改进做成产品级特性。下一步值得盯着两件事——ChatGPT 端用户能否真切感知到风格与事实性改进(公告里点名\"接近 Astra 级事实性但只有一半错误\"),以及 Anthropic 是否会用 Opus 5 系列的定价调整做回应。","https:\u002F\u002Fopenai.com\u002Findex\u002Fintroducing-gpt-6-sol-and-luna\u002F","15975962-b5fe-49e5-ae68-687ba6cb7015",[11,15,18,21,24],{"id":12,"name":13,"slug":13,"description":14,"color":14},"7ac06d8e-b074-4147-abfc-ffaa4c6b8744","ai-efficiency",null,{"id":16,"name":17,"slug":17,"description":14,"color":14},"baf131c1-687a-49f4-87f6-4dd87c1c692f","gpt",{"id":19,"name":20,"slug":20,"description":14,"color":14},"01598627-1ea6-4b27-a5d8-874971571a71","llm",{"id":22,"name":23,"slug":23,"description":14,"color":14},"7e89b5cc-57db-4f37-bc6d-28919a73931c","model-release",{"id":25,"name":26,"slug":26,"description":14,"color":14},"42e59a88-7795-47dc-a334-ef1e72c24347","openai",[28],{"id":29,"lang":30,"title":31,"summary":32,"content":33},"77cd6acb-62ab-4e8b-8b9a-65240d0f0377","en","OpenAI splits GPT-6 into Sol and Luna, halving API prices","GPT-6 Sol and Luna ship with API prices halved vs GPT-5.6 and beat Claude Opus 5 \u002F Fable 5 on AutomationBench DeepSWE OSWorld at lower per-task cost.","OpenAI is no longer asking teams to pay flagship prices for near-flagship capability. With GPT-6 Sol and GPT-6 Luna, the company is shipping two new tiers trained with the same methods as GPT-6 Astra, then cutting API prices in half relative to their GPT-5.6 predecessors.\n\nThe headline numbers are clean. GPT-6 Sol drops from $4 to $2 per million input tokens and from $20 to $10 for output. GPT-6 Luna goes from $0.20 to $0.10 on input and $1.20 to $0.50 on output. OpenAI credits caching and inference infrastructure improvements for letting the savings flow through.\n\nThe benchmark story is \"near-flagship at a fraction of the cost.\" OpenAI picked five evaluations that map to complex professional work: AutomationBench, Agents' Last Exam, FrontierCode, DeepSWE 1.1, and OSWorld 2.0 offline.\n\nOn AutomationBench 1.0.6, which runs AI agents through end-to-end workflows across 47 tools spanning sales, marketing, operations, support, finance, and HR, GPT-6 Sol at xhigh effort scored 33.2% at $0.27 per task. Claude Opus 5 at max effort scored 26.9% at 11.1x Sol's per-task cost. Claude Fable 5.1 with Opus 5 fallback hit 31.4% at 8.9x. On Agents' Last Exam V1, GPT-6 Sol at max effort reached 56.4%, beating Claude Opus 5's top score while costing 60% less per task.\n\nOn the coding side, FrontierCode 1.1 Main grades coding agents on whether their changes can actually be merged into real codebases — not just correctness, but test quality, scope discipline, code style, and adherence to codebase standards. GPT-6 Sol improved substantially over GPT-5.6 Sol and matched Claude Fable 5.1 xhigh at much lower cost. On DeepSWE 1.1, GPT-6 Sol max effort scored 68.8%, within 1.1 percentage points of Claude Fable 5's 69.9% at xhigh, but at roughly 20% of the cost per task. GPT-6 Luna max effort scored 66.6%, comparable to Claude Opus 5 and Fable 5 at medium effort, while costing 93% less than Opus 5 and 96% less than Fable 5 per task.\n\nOn OSWorld 2.0 offline, GPT-6 Sol xhigh scored 60.5% — essentially matching Claude Opus 5 medium at 60.3% — at about 20% of the cost. GPT-6 Luna max beat GPT-5.6 Sol medium at one tenth the cost.\n\n## Caching as a product feature\n\nThe more interesting shift may be in the caching layer. OpenAI published a separate piece on the same day detailing prompt caching improvements for GPT-6: higher default cache hit rates, a 90% discount on cached input token reads, a Prompt Caching Dashboard, a diagnostics tool, and explicit breakpoint APIs that let developers adjust reasoning effort and tool availability without breaking cache. GitHub reported that across billions of requests to OpenAI models over the past few months, the share of prompt tokens requiring fresh processing fell by more than 50%, helping Copilot respond faster.\n\n## Alignment and availability\n\nOn alignment, OpenAI says Sol and Luna inherit the alignment work shipped with Astra, including lower rates of misleading claims about coding work in their internal coding deception evaluation.\n\nAvailability lands the same day in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise, and Edu tiers, with Free and Go users getting Luna on desktop. On the API side, both ship as `gpt-6-sol` and `gpt-6-luna`, rolled out gradually throughout the day to keep service stable.\n\n## Why this matters\n\nThe philosophical move is more important than any single number: Astra is still the flagship, but OpenAI is now explicitly distributing \"near-flagship capability at lower cost\" as the design center of the GPT-6 family. Sol beats Astra at low effort on AutomationBench. Luna matches Opus 5 at medium effort on DeepSWE at 7% of the cost. The expensive-model-plus-high-effort combo is no longer the only way to clear complex workloads — and that resets the TCO math for anyone running Claude Opus or Fable in production today.\n\nFor enterprise buyers, the practical question is whether the cached-input economics, the new breakpoint APIs, and the 50% price cut combine into a workload they can actually migrate. For individual developers, Luna drops the entry price one more notch. Watch two things next: how visibly the style and factuality improvements show up for ChatGPT users (OpenAI claims \"approaching Astra-level reliability\" with \"about half as many mistakes\"), and whether Anthropic responds with pricing moves on the Opus 5 family.","openai-gpt-6-sol-luna-api-pricing","2026-09-23T00:00:00Z","2026-09-23T09:05:56.872987Z","2026-09-23T09:05:56.872997Z",true,"agent",1,[42,51],{"slug":43,"tag_slug":43,"title_zh":44,"title_en":45,"intro_zh":46,"intro_en":47,"id":48,"is_active":38,"created_at":49,"modified_at":50},"ai-for-science","AI for Science 2026：从 UniPert 到 GPT-Rosalind 的硬核进化","AI for Science 2026: from UniPert to GPT-Rosalind","生命科学、化学材料、物理世界模型——AI 正在从\"语言工具\"变成\"实验伙伴\"。本专题收录 AI 在三大科学方向的关键节点：UniPert 统一基因与化学扰动空间、GPT-Rosalind 端到端生命科学推理、达摩院 AI 智能体 28 小时找到 4 种超导新材料、Anthropic Claude Science 把工作台做成标准品。","From language tool to lab partner — AI is reshaping life sciences, chemistry\u002Fmaterials, and physical world models. This topic covers the key milestones: UniPert unifying genetic-chemical perturbation spaces, GPT-Rosalind's end-to-end life-sciences reasoning, DAMO's AI agent discovering 4 superconducting materials in 28 hours, and Anthropic's Claude Science workbench going mainstream.","988a4300-5fab-41c4-b5d8-63711a2dc757","2026-09-10T01:34:15.296649Z","2026-09-10T01:34:15.296663Z",{"slug":52,"tag_slug":52,"title_zh":53,"title_en":54,"intro_zh":55,"intro_en":56,"id":57,"is_active":38,"created_at":58,"modified_at":59},"h3-series","MiniMax H3 系列：从开源权重到 35 倍吞吐","MiniMax H3 Series: from open weights to 35x throughput","MiniMax H3 自 2026 年 8 月开源以来节奏密集：官方把生成、参考与编辑收回一个模型；ComfyUI 当天压进 RTX 3060；摩尔线程 3 小时完成国产 GPU 适配；fal 后训练版把吞吐拉到 35 倍；FastH3 蒸馏再砍推理成本。本专题持续追踪 H3 的发布—开源—蒸馏—部署全链路。","Since MiniMax open-sourced H3 in August 2026 the pace has been relentless: one unified omni-modal model, same-day ComfyUI support down to an RTX 3060, a 3-hour Day-0 port to Moore Threads GPUs, fal's post-trained H3 Max at 35x throughput, and FastH3 distillation cutting inference cost further. This topic tracks the full H3 chain — release, open weights, distillation, deployment.","83ef0daa-3c31-4cb3-86ed-e5ee58654d5f","2026-09-08T07:33:19.942193Z","2026-09-08T07:33:19.942209Z",{"items":61},[62,67,72,77,82,87],{"id":63,"title":64,"news_slug":65,"published_at":66},"5bf8fa2d-258e-41d3-bfb6-5c2053433cfd","GPT-6 Astra 正式上线:8 月因安全被暂停的旗舰回来了","gpt-6-astra-launch","2026-09-04T03:12:38+00:00",{"id":68,"title":69,"news_slug":70,"published_at":71},"6c5f1bc4-d877-483c-a0e9-70aff0e30dbe","微软内部 Ramp 账单:一名工程师 28 天烧掉 2.8 万美元 AI 费","microsoft-internal-ramp-ai-spending-28000-28-days","2026-08-31T03:00:00+00:00",{"id":73,"title":74,"news_slug":75,"published_at":76},"d95940eb-69c1-467e-9d60-5886ab71d985","GPT-5.6-Cyber 上线、Daybreak 分层、Astra 推迟:OpenAI 把\"网络安全模型\"做成一个独立产品线","openai-gpt-5-6-cyber-daybreak-astra-2026","2026-08-11T04:00:00+00:00",{"id":78,"title":79,"news_slug":80,"published_at":81},"c2ee2a09-d001-4740-9820-21fb672eee8b","Copilot 默认模型切到 GPT-5.6 Sol：tokenmaxxing 终结","microsoft-gpt5-6-default-token-budget","2026-08-08T08:00:00+00:00",{"id":83,"title":84,"news_slug":85,"published_at":86},"88944bec-d33f-4383-aece-0d5207a06eab","GPT-5.6 全面开放:Ultra 把 4 agent 并行写进 API,程序化工具调用把 token 效率再压一档","gpt-5-6-launch","2026-07-10T06:03:00+00:00",{"id":88,"title":89,"news_slug":90,"published_at":91},"69613959-04c5-43d0-97ec-9311473d8d93","GPT-5.6 三档齐发:用 1\u002F3 token 追平 Mythos,OpenAI 把效率-能力前沿压到新位置","gpt-5-6-three-tiers-1-3-tokens-mythos","2026-06-27T04:00:00+00:00"]