OpenAI 发布了 GPT-6 Sol 和 GPT-6 Luna，这是其 GPT-6 家族中的 2 款新模型。它们的定位低于本月早些时候推出的 GPT-6 Astra。OpenAI 使用与 Astra 类似的方法训练了这两款模型。目标是将 Astra 的进步带到更快、更实惠的模型上。
今天即可部署？ 是的。两款模型均已在 OpenAI API 中上线，分别为 gpt-6-sol 和 gpt-6-luna。它们是仅限 API 的模型，因此没有可供自行部署的权重。
三个层级，同一配方
GPT-6 家族现在有 3 个层级。Astra 是处理最困难工作的顶级模型。Sol 以更低的成本面向复杂编程和专业任务。Luna 面向快速、大批量的日常工作。
OpenAI 团队表示，更好的缓存和推理使其能够以更低的成本提供这些模型。它将 Sol 和 Luna 的 API 价格相较其 GPT-5.6 促销定价下调了 50%。
| 模型 | 输入（每 1M tokens） | 输出（每 1M tokens） |
|---|---|---|
| GPT-6 Astra | $10.00 | $50.00 |
| GPT-6 Sol | $2.00（原价 $4） | $10.00（原价 $20） |
| GPT-6 Luna | $0.10（原价 $0.20） | $0.50（原价 $1.20） |
有一个细节值得注意。Luna 的输出价格从 $1.20 降至 $0.50，降幅约为 58%，而非 50%。
基准测试：OpenAI 报告了什么
专业工作：在 AutomationBench 1.0.6 上，Sol 在 xhigh 推理强度下得分为 33.2%，每任务成本 $0.27。Claude Opus 5 在 max 推理强度下得分为 26.9%，成本是前者的 11.1 倍。低推理强度的 Astra 得分为 30.3%，成本是 Sol 的 3.9 倍。Luna 在高推理强度下比其前代提升 5.4 个百分点，每任务成本降低 58%。
在 Agents’ Last Exam 上，Sol 在 max 推理强度下得分为 56.4%。这超过了 Claude Opus 5 的最佳成绩，且每任务成本低 60%。
编程：在 DeepSWE v1.1 上，Sol 在最大努力下得分 68.8%。这比 xhigh 下的 Claude Fable 5 落后 1.1 个百分点，而每任务成本约低 80%。Luna 在最大努力下得分 66.6%，与 medium 努力下的 Opus 5 和 Fable 5 相当。在这些对比中，Luna 每任务成本比 Opus 5 低 93%，比 Fable 5 低 96%。
在 FrontierCode 1.1 Main 上，该评测衡量代码是否已准备好合并，Sol 在 xhigh 下与 Claude Fable 5.1 持平，而成本低得多。
计算机使用：在离线 OSWorld 2.0 上，xhigh 下的 Sol 得分 60.5%，而 medium 下的 Opus 5 为 60.3%。Sol 的每任务成本约低 80%。最大努力下的 Luna 击败了 medium 下的 GPT-5.6 Sol，而成本仅为后者的 1/10。
事实性：OpenAI 的内部测试使用经过脱敏的 ChatGPT 对话，其中用户标记了模型错误。Sol 犯的错误约为其前代的一半。更高努力下的 Luna 与 GPT-5.6 Sol 相当，而成本约为后者的 1/100。
OpenAI 还沿用了 Astra 的沟通风格。可以预期回答会更清晰、略短，术语更少，尤其是在编程对话中。
面向长时间运行智能体的提示词缓存
智能体每一轮都会重新发送相同的指令、工具和历史记录。GPT-6 搭载了改进的提示词缓存系统，默认缓存命中率更高。缓存输入读取可享受最高 90% 的折扣。在 30 分钟窗口内复用的符合条件的共享前缀现在也可享受此优惠。
面向开发者的新控制项：
- 提示词缓存仪表盘可随时间追踪命中率。
- 诊断工具可解释未命中的原因，例如
"reason": "tools_changed"。 - 显式断点让你可以选择缓存前缀的结束位置。
- 推理强度可通过
configuration_update在对话中途更改，而不会破坏缓存。 allowed_tools限制可调用的工具，同时保持定义稳定。- 预热会在首个用户请求之前准备好已知上下文。
完整的提示词缓存指南涵盖了每种模式。GitHub 报告称，这些改动将需要全新处理的提示词 token 占比降低了 50% 以上，帮助 Copilot 更快响应。
可用性
- API：
gpt-6-sol和gpt-6-luna。 - ChatGPT Work 和 Codex：Plus、Pro、Business、Enterprise 和 Edu 用户。
- Free 和 Go：ChatGPT 桌面应用中的 Luna。
- 尚未在 Chat 中推出。ChatGPT 的推送将在发布日逐步进行。
交互式讲解
核心要点
- Sol 的价格为每 1M tokens $2/$10，Luna 为 $0.10/$0.50。
- Sol 在 xhigh 下于 AutomationBench 上击败 Opus 5 max，成本仅为后者的 9%。
- Luna 在 DeepSWE v1.1 上得分 66.6%，每项任务成本比 Opus 5 低 93%。
- 缓存输入读取最高可享 90% 折扣，并新增缓存控制功能。
- 两者均已在 API、ChatGPT Work 和 Codex 中上线；尚未在 Chat 中推出。
OpenAI has released GPT-6 Sol and GPT-6 Luna, 2 new models in its GPT-6 family. They sit below GPT-6 Astra, which launched earlier this month. OpenAI trained both with methods similar to Astra’s. The aim is to bring Astra’s advances to faster, more affordable models.
Deployable today? Yes. Both models are live in the OpenAI API as gpt-6-sol and gpt-6-luna. They are API-only models, so there are no weights to self-host.
Three tiers, one recipe
The GPT-6 family now has 3 tiers. Astra is the top model for the hardest work. Sol targets complex coding and professional tasks at lower cost. Luna targets fast, high-volume everyday work.
OpenAI team states better caching and inference let it serve these models more cheaply. It is cutting Sol and Luna API prices by 50% against their GPT-5.6 promotional pricing.
| Model | Input (per 1M tokens) | Output (per 1M tokens) |
|---|---|---|
| GPT-6 Astra | $10.00 | $50.00 |
| GPT-6 Sol | $2.00 (was $4) | $10.00 (was $20) |
| GPT-6 Luna | $0.10 (was $0.20) | $0.50 (was $1.20) |
One detail is worth noting. Luna’s output price falls from $1.20 to $0.50, a cut of about 58%, not 50%.
Benchmarks: what OpenAI reports
Professional work: On AutomationBench 1.0.6, Sol at xhigh effort scores 33.2% at $0.27 per task. Claude Opus 5 at max effort scores 26.9% at 11.1x that cost. Low-effort Astra scores 30.3% at 3.9x Sol’s cost. Luna at high effort gains 5.4 points over its predecessor at 58% lower cost per task.
On Agents’ Last Exam, Sol at max effort scores 56.4%. That beats Claude Opus 5’s best score at 60% lower cost per task.
Coding: On DeepSWE v1.1, Sol at max effort scores 68.8%. That is 1.1 points behind Claude Fable 5 at xhigh, at about 80% lower cost per task. Luna at max effort scores 66.6%, comparable to Opus 5 and Fable 5 at medium effort. In those comparisons, Luna costs 93% less per task than Opus 5 and 96% less than Fable 5.
On FrontierCode 1.1 Main, which grades whether code is ready to merge, Sol matches Claude Fable 5.1 at xhigh at much lower cost.
Computer use: On OSWorld 2.0 offline, Sol at xhigh scores 60.5% versus 60.3% for Opus 5 at medium. Sol’s cost per task is about 80% lower. Luna at max beats GPT-5.6 Sol at medium for 1/10 of the cost.
Factuality: OpenAI’s internal test uses de-identified ChatGPT conversations where users flagged model errors. Sol makes about half as many mistakes as its predecessor. Luna at higher effort matches GPT-5.6 Sol at about 1/100 of its cost.
OpenAI also carried Astra’s communication style over. Expect clearer, slightly shorter answers with less jargon, especially in coding conversations.
Prompt caching for long-running agents
Agents resend the same instructions, tools and history on every turn. GPT-6 ships an improved prompt caching system with higher cache hit rates by default. Cached input reads get discounts of up to 90%. Eligible shared prefixes reused within a 30-minute window now qualify.
New controls for developers:
- A Prompt Caching Dashboard tracks hit rates over time.
- A diagnostics tool explains misses, for example
"reason": "tools_changed". - Explicit breakpoints let you choose where a cached prefix ends.
- Reasoning effort can change mid-conversation via
configuration_updatewithout breaking cache. allowed_toolsrestricts callable tools while keeping definitions stable.- Prewarming prepares known context before the first user request.
The full prompt caching guide covers each pattern. GitHub reports these changes cut the share of prompt tokens needing fresh processing by more than 50%, helping Copilot respond faster.
Availability
- API:
gpt-6-solandgpt-6-luna. - ChatGPT Work and Codex: Plus, Pro, Business, Enterprise and Edu users.
- Free and Go: Luna in the ChatGPT desktop app.
- Not yet in Chat. The ChatGPT rollout is gradual through launch day.
Interactive Explainer
Key Takeaways
- Sol costs $2/$10 and Luna $0.10/$0.50 per 1M tokens.
- Sol at xhigh beats Opus 5 max on AutomationBench at 9% of the cost.
- Luna scores 66.6% on DeepSWE v1.1, costing 93% less per task than Opus 5.
- Cached input reads get up to 90% off, with new cache controls.
- Both are live in the API, ChatGPT Work and Codex; not yet in Chat.