GPT-5.6 Luna 向免费用户开放

2026 年 8 月 7 日，OpenAI 宣布了一件在别家可能只是更新日志的事：免费与 Go 用户的默认模型切换为 GPT-5.6 Luna，下周起无限次文本聊天。文件、图片、语音仍然限额——但文本不再计数。对每天打开 ChatGPT 问几个问题的普通用户，这是免费档历史上第一次用上「当前旗舰」。

更值得注意的，是那个叫 Think 的按钮。过去，深度推理是付费墙后的能力：免费用户只能得到快速回答，想要「多想一步」就得订阅。Think 按钮让免费档第一次可以主动调用更高推理——不是系统偶尔给，而是用户自己按。推理从「订阅特权」变成了「产品参数」，这个转变比任何榜单都更能说明 OpenAI 对免费档的重新定位：免费入口不再只是体验旧模型的窗口，而是当前旗舰的分发渠道。

同一天，Plus/Pro 的 GPT-5.6 Sol 也升级了：官方称事实错误率较 GPT-5.5 Instant 降约 68%，并新增思考强度滑块。68% 是厂商口径，需要带着自述边界读；但「事实错误率」成为可量化的卖点本身，说明 OpenAI 把质量竞争从「谁更聪明」转向「谁更少出错」。思考强度滑块则把推理预算交到用户手里——想要更快的回答还是更稳的答案，自己调。

把时间线拉长看，这次更新是 7 月 9 日 GPT-5.6 家族发布、7 月 31 日 API 降价（Luna -80%、Terra -20%）之后的第三步。模型发布、价格重构、免费开放，三步在一个月内走完。Luna 的 API 价格已经降到接近零，免费档的无限文本聊天在成本上变得可行——OpenAI 在用 Luna 的规模效应，把「最新模型」变成免费产品的默认体验。

对行业，问题变得尖锐：当免费档都能用上当前旗舰和主动推理，付费墙的价值还剩什么？对用户，答案简单得多：那个曾经锁在订阅墙后的按钮，现在可以按了。ChatGPT 免费档的定位，从「体验旧模型」变成了「体验当前旗舰」——而这一步，可能比任何一次模型发布都更改变普通人的使用习惯。

On August 7, 2026, OpenAI announced something that at any other company would be a changelog entry: free and Go users' default model switched to GPT-5.6 Luna, with unlimited text chat starting next week. Files, images, and voice remain capped—but text stops counting. For ordinary users who open ChatGPT a few times a day, this is the first time the free tier has run the current flagship.

The more notable piece is a button called Think. Deep reasoning used to live behind the paywall: free users got fast answers, and "think harder" required a subscription. The Think button gives the free tier its first on-demand higher reasoning—not occasionally granted by the system, but pressed by the user. Reasoning moved from subscription privilege to product parameter, and that shift says more about OpenAI's repositioning of the free tier than any leaderboard: the free door is no longer a window into old models but a distribution channel for the current flagship.

The same day, GPT-5.6 Sol for Plus/Pro also upgraded: the company reported roughly 68% fewer factual errors than GPT-5.5 Instant, plus a new thinking-effort slider. The 68% is a vendor claim and should be read with that boundary; but making "factual-error rate" a quantifiable selling point shows OpenAI moving quality competition from "who is smarter" to "who makes fewer mistakes." The effort slider hands the reasoning budget to the user—faster answers or steadier ones, your call.

Stepping back, this update is the third step after the GPT-5.6 family launch on July 9 and the API price cuts on July 31 (Luna -80%, Terra -20%). Model release, price restructuring, free access—three steps in one month. With Luna's API price near zero, unlimited text chat on the free tier became cost-feasible; OpenAI is using Luna's scale economics to make "latest model" the default free experience.

For the industry, the question turns sharp: when the free tier runs the current flagship with on-demand reasoning, what is left inside the paywall? For users, the answer is simpler: the button that used to sit locked behind the subscription wall can now be pressed. The free tier's positioning moved from "try an old model" to "try the current flagship"—and that step may change ordinary habits more than any single model release.