1. Home
2. Guides
3. OpenAI

Codex models and plans explained: GPT-6 Astra, Sol and Luna, and which plan gets which

Pick any GPT-6 model you like on your laptop, but a Codex cloud task on a ChatGPT plan still runs on last generation's Sol.

By Think Facility · September 23, 2026 · Codex OpenAI models usage limits

Codex runs three new models as of September 23, 2026: GPT-6 Astra, GPT-6 Sol and GPT-6 Luna. Astra arrived in the Codex CLI's model picker on September 9, and Sol and Luna started rolling out on September 22. Which ones you get depends on your ChatGPT plan, and less obviously, on whether the task runs on your machine or in OpenAI's cloud.

I read OpenAI's Codex models page, its pricing page and the Codex changelog to put this together. Codex now lives inside ChatGPT, and OpenAI's own docs sit at learn.chatgpt.com.

Astra, Sol and Luna in one line each

OpenAI describes Astra as "our most capable model for complex work across code, apps, and research". Sol is "built for complex coding and agentic workflows". Luna is "our most efficient model for focused, high-volume tasks".

The models page turns that into advice. Use Astra for complete jobs across code, apps and research that need sustained judgment, Sol for everyday work and hard code changes, and Luna when you already know what a good result looks like (extraction, classification, structured summaries). Sol and Luna don't work in Chat, the ordinary ChatGPT conversation window; in ChatGPT they're limited to Work and Codex.

The GPT-5.6 generation hasn't gone anywhere yet. GPT-5.6 Sol, Terra and Luna "remain available during the rollout", and OpenAI adds that picking a new model "doesn't change workspace permissions or grant access to it".

Which plan gets which model

Even Free and Go get a slice of Codex. Both get GPT-6 Luna at standard speed in the ChatGPT desktop app, "subject to rollout", and nothing else. Go costs $8 a month. It's a real way to try an agent on a small job, but the terminal program and the editor extension aren't included.

Plus, at $20 a month, is where Codex opens up: the web, the CLI, the IDE extension and iOS, with Astra and Sol added to Luna. Pro starts at $100 and gives five or twenty times the Plus allowance. Business costs $20 per user per month billed annually, or $25 monthly, and the $100 Business tier uses the same estimates as Pro 5x. On Enterprise and Edu, an administrator has to enable the new models before anyone sees them.

OpenAI publishes its limits as estimated local messages per five-hour window. These are the GPT-6 rows.

Local messages per five hours, from OpenAI's Codex pricing page, read September 23, 2026. OpenAI calls these estimates and says weekly limits may also apply. Standard Business matches Plus.

The ranges are wide on purpose. OpenAI says a small script might use "only a fraction" of your allowance while a long task that holds a lot of context uses "significantly more per message". If you're on Astra, our Astra limits page has what happens when you run out.

Codex cloud still runs on GPT-5.6 Sol

This one took me two passes to believe. On the models page, the cards for Astra and Luna both mark "Codex cloud" as unavailable. The pricing page then says "cloud chats on ChatGPT plans use GPT-5.6 Sol", and the models page adds that you can't currently change the default model for them.

So a task you hand to Codex in the cloud runs on the previous generation's Sol whatever you've set locally. Cloud work also needs a ChatGPT sign-in, according to the authentication page, so an API key can't reach it at all. Local messages and cloud chats draw on the same allowance.

Picking an effort level

Reasoning effort matters about as much as the model. OpenAI's starting points are Medium for Sol, High for Luna and Light for Astra (the CLI calls Astra's Light setting low). In the desktop app these are bundled into six presets, from Luna High up to Astra Extra High, and Sol Light is the one you start on. Some paid plans don't get Astra Extra High.

Two modes sit above the ordinary levels. Max gives one model more time on a single task. Ultra splits the work across subagents running in parallel. Luna goes as far as Max but can't use Ultra. OpenAI says "most tasks do not need Max or Ultra."

What GPT-6 costs in credits

Past your included usage, Plus and Pro users can keep going on credits, and the rate card shows why OpenAI's changelog calls the new models cheaper. GPT-6 Sol costs 50 credits per million input tokens and 250 per million output, exactly half of GPT-5.6 Sol's 100 and 500. GPT-6 Luna is half of GPT-5.6 Luna on input too. Astra sits far above both at 250 and 1,250, and Fast mode multiplies any of the three by 2.5. There's more on Astra's rates on our credits page.

When you run out, and the API key route

Hitting the limit mid-task doesn't kill the task. OpenAI's pricing page says that if you reach it "during an active turn, the agent will be able to continue working on that turn", within fair use. After that you can wait for the window to reset, buy credits, or sign in with an API key and pay standard API rates for extra local work. In the CLI, /status shows what's left.

Signing in with an API key instead of ChatGPT is a different product in practice. You get Codex in the CLI, the SDK and the IDE extension, with no cloud features at all, so no GitHub code review and no Slack. The models you can pick are whatever the API allows your key, which is why none of the retirements below apply to it.

Plus and Pro users also get an opt-in experiment, off by default, where Astra keeps notes across context windows and can search earlier messages from the same task. OpenAI says it isn't available on Business, Enterprise or API-key sign-ins at launch.

What's leaving, and the advice that changed

GPT-5.4 and GPT-5.4 mini retired from Codex with a ChatGPT sign-in on August 31, 2026. GPT-5.2 and GPT-5.3-Codex were already deprecated there, and GPT-5.3-Codex-Spark went on September 14. GPT-5.5 is next, on October 14, across every ChatGPT plan.

The replacement advice moved within a week, so a config file you set up in mid-September may point at the wrong model. The changelog entry from September 14 says to switch from GPT-5.5 to GPT-5.6 Sol. The models page, updated after Sol and Luna arrived, now says GPT-6 Sol on paid plans and GPT-6 Luna on Free and Go. None of these retirements touch Codex run with your own API key, where the models follow whatever the API offers your key.

If you're weighing Codex against Anthropic's or Google's agents, the three-way comparison covers plans and limits side by side, and OpenAI's model release dates puts every name in order.

The Codex CLI's 0.156.1 release, on September 23, 2026, added Sol and Luna to its model picker and changed the prompt you see at a rate limit to recommend Luna.