GPT-6 Sol / Luna 定价与 AutomationBench 数据曝光
截图帖列出 Sol 与 Luna 的 API 价格、缓存折扣和 AutomationBench 成绩，称 Sol 以约 9% 单任务成本胜过 Opus 5，Astra 仍是旗舰。
GPT-6 Sol beat Claude Opus 5 on AutomationBench at 9% of the cost per task. This is not a new flagship. Astra stays on top. OpenAI shipped two working floors of the same line: Sol and Luna. Both trained with Astra’s methods. Both 50% cheaper than their 5.6 twins. API price per 1M tokens: Sol: $2 in, $10 out. Was $4 / $20. Luna: $0.10 in, $0.50 out. Was $0.20 / $1.20. Cached reads get a 90% discount: $0.20 for Sol, $0.01 for Luna. Cache writes: $2.50 and $0.125. Changing reasoning effort or the tool set no longer busts the cache. If an agent reuses one system prompt across thousands of tasks, the bill does not fall by 2x. It falls more. The cache now survives mode switches. OpenAI’s numbers, not my runs. AutomationBench, business workflows across 47 tools: Sol at xhigh: 33.2% at $0.27 per task. Opus 5 at max: 26.9% at 11.1x that cost. Internal factual-error check on hard prompts: Sol makes about half as many mistakes as GPT-5.6 Sol. Luna at high effort matches GPT-5.6 Sol at roughly 1% of its cost. FrontierCode: Sol sits next to Claude Fable 5.1 and costs less. Luna on the cheap end of the chart is already where mid-tier used to live. Rollout today: ChatGPT Work and Codex for Plus, Pro, Business, Enterprise, Edu. API IDs: `gpt-6-sol` and `gpt-6-luna`. Luna is in the ChatGPT desktop app for Free and Go. The models page still has nothing. No date given. Astra for the rare heavy jobs. Production default for the next few weeks: Luna for volume, Sol for code and agents. If you do not retune cache and effort to the new prices, you just pay double for the same throughput.