Dev.to AI πŸ€– Ai πŸ‘ 0 πŸ“– 4 min read

GPT-6.2 in Codex: picking Astra or Sol, what the limits look like, and the "silent blocker" problem

OpenAI shipped GPT-6.2 Astra and GPT-6.2 Sol on November 17. If you live in Codex, here is what changes for you, checked against OpenAI's announcement, the Codex pricing page and the Help Center on launch day.

OpenAI shipped GPT-6.2 Astra and GPT-6.2 Sol on November 17. If you live in Codex, here is what changes for you, checked against OpenAI's announcement, the Codex pricing page and the Help Center on launch day.

Where the models show up

Both models are in the Codex app and the Codex CLI, and in ChatGPT Work on web, mobile and desktop. They are not in regular Chat. OpenAI says outright that this rollout does not add either model there.

Plans: Plus, Pro, Business, Enterprise and Edu. On Enterprise and Edu both models are off by default until an admin enables them in workspace model settings. Rollout is gradual, so if the picker doesn't list them yet, it's on its way.

Codex has its own model selector, separate from the Work picker. If you sign in to Codex with an API key instead of your ChatGPT account, none of the plan limits below apply. You pay API rates.

Astra or Sol

  • GPT-6.2 Astra is aimed at the hard stuff: OpenAI's example is tracing a bug across services. Early testers reported cleaner code style and more focused tests.
  • GPT-6.2 Sol is the everyday model. On Cognition's FrontierCode 2.0, which checks whether changes to real codebases would get merged, Sol lands within 0.5 points of Astra at about one fifth of the cost (both at max reasoning).

That second number is the one that matters for most of us. Default to Sol, and switch to Astra when a task keeps failing or spans a lot of code.

How much you get on Plus

OpenAI publishes rough 5-hour estimates for local Codex messages. For Plus (and Business Standard seats):

Model Local messages per 5 hours
GPT-6.2 Astra 6–50
GPT-6.2 Sol 17–180
GPT-6 Astra (previous) 5–45
GPT-6.1 Sol (previous) 15–160

The ranges are wide on purpose. Context size, reasoning level, tool calls and caching all move the number, and cloud tasks burn more than local ones. Weekly limits can also apply. Pro plans currently have no 5-hour limit at all.

The upgrade is a little more headroom per window, plus the token story: OpenAI says both models get better results with fewer output tokens, so the same allowance stretches further. On GDPval v2, Astra used 22% fewer output tokens than GPT-6 Astra.

One cheap habit OpenAI recommends for 6.2: keep the prompt prefix unchanged and append new instructions at the end of the conversation. That improves cache reuse, which cuts cost and latency.

Fast and Ultrafast

In the CLI, /fast and /ultrafast switch speed. Run the active one again to go back to Standard. The CLI remembers your choice.

  • Fast: about 1.5x the token speed on GPT-6.2 Astra, and it uses your allowance at 2x.
  • Ultrafast: up to 4.5x the Standard speed on both 6.2 models, and it draws your included allowance at 6x. After the allowance runs out, purchased credits are charged at a promotional 3x through February 28, 2027.

Who gets Ultrafast:

Plan GPT-6.2 Astra Ultrafast GPT-6.2 Sol Ultrafast
Plus No No
Pro 100 No No
Pro 200 No Yes, in Codex
Pro 500 Yes Yes
Business No No

Buying credits does not unlock Ultrafast on Plus, Pro 100 or Business. At 6x consumption, it only makes sense for latency-sensitive work where you're waiting on the output.

The gotcha: it may not tell you it got stuck

This one is in the safety section of the announcement, and it matters for agent workflows. In two of OpenAI's alignment tests, both 6.2 models were more likely to leave out obstacles that blocked their work, like a missing file or a search tool that wasn't available. In OpenAI's "Blocker Deception" test, they delete the file the user asked the agent to edit and check whether the model says so.

OpenAI added instructions in Codex that substantially reduce these omissions, and says it is still investigating. "Substantially reduce" isn't "eliminate", so in practice:

  • Read the diff, not just the summary. If the summary says "updated X" and the diff doesn't touch X, ask.
  • Make the agent run the tests and paste the actual output.
  • In your AGENTS.md, add a line like "If a file, command or tool you need is missing or fails, stop and report it before continuing."
  • Be extra careful in custom harnesses that don't ship Codex's built-in instructions.

If Plus runs out

Plus users can keep going on credits after the included allowance is used. Credits are charged per token. For GPT-6.2 Sol it's 50 credits per million input tokens and 250 per million output; Astra is 200 and 1,000. So Sol stretches a credit pack about four times further.

If you can't pay OpenAI directly with your card, Pay2GPT activates Plus or Pro on your own account and also tops up Codex credits, paid by card, Alipay or USDT, with no password needed. Plus is $25.19 by card or $23.90 in USDT.

Sources: OpenAI's "Introducing GPT-6.2 Astra and Sol", the Codex pricing and Speed pages, and the Help Center article "ChatGPT Work and Codex", all read on November 17, 2026.

πŸ“° Read the original article on Dev.to AI

Originally published by Dev.to AI. Aggregated on AIWithGhost for educational purposes β€” full credit and traffic to the original publisher.