---
title: "GPT-6 Astra lands in Codex: what changes for vibe coders"
description: "OpenAI launched GPT-6 Astra on September 3, 2026 in ChatGPT, Codex and the API. Pricing, rollout and the new context memory explained."
url: "https://vibe-code.blog/en/gpt-6-astra-codex"
lang: "en"
author: "Thierry Gustin"
published: "2026-09-12"
updated: "2026-11-21"
---

# GPT-6 Astra lands in Codex: what changes for vibe coders

> **Key takeaways** — OpenAI launched GPT-6 Astra on September 3, 2026, calling it its best model for software engineering, available in ChatGPT, Codex and the API at $10 / $50 per million tokens. In Codex, Astra brings an experimental way to keep notes across context windows, which helps on long sessions. For subscribers, usage is included in the plan; for an app that calls the API, the price forces you to decide where Astra is really worth it.

**GPT-6 Astra**, launched by OpenAI on September 3, 2026, is the company's new flagship model, and Codex is one of the first products to get it. For vibe coders, two things matter: Astra is included in paid ChatGPT plans, and it comes with a new way to keep track of long coding sessions. On the API side, though, it is an expensive model.

## What did OpenAI launch on September 3, 2026?

OpenAI released GPT-6 Astra, which it calls "the world's most intelligent and aligned model" and, per 9to5Mac, "the best model for software engineering to date." It is available in ChatGPT (including the Work space), in Codex and in the API, under the model ID `gpt-6-astra`.

The rollout came in waves. According to OpenAI's announcement page, Astra first went to a limited set of organizations, then was due to reach all ChatGPT Plus, Pro, Business and Enterprise users "over the coming days," along with the API, Microsoft Azure and AWS Bedrock. 9to5Mac reports that Business and Pro subscribers ($100 and $200 a month) got access a day after the announcement, and Plus subscribers ($20 a month) a few hours after that.

> **Enterprise accounts: off by default** — For Enterprise accounts, Astra is off at launch: an admin has to turn it on for the workspace. If you cannot see it in Codex at work, that is probably why.

### The launch numbers

| Item | GPT-6 Astra (announced September 3, 2026) |
| --- | --- |
| API price (standard) | $10 input / $50 output per million tokens |
| Cached input | $1 per million tokens |
| Context window | 1.05 million tokens |
| Max output | 128,000 tokens |
| Fast mode | up to 2x faster, at 2x the standard price |
| Terminal-Bench 4.0 | 57.9% (GPT-5.6 Sol: 37.3%) |
| DeepSWE v1.1 | 74.1% (GPT-5.6 Sol: 72.7%) |

Pricing, fast mode and benchmarks come from OpenAI's announcement page. Context window and max output are reported by Forklog and by the model card quoted on OpenAI's developer forum. Forklog adds that above 272,000 input tokens, input rates double and output is billed at 1.5x for the whole request.

## What changes in Codex?

The most useful change for vibe coders is context handling. With Astra, Codex can keep notes across context windows instead of repeatedly compressing everything into a summary, and earlier windows stay searchable. In practice, the agent can find a requirement, a test result or a tool output it saw an hour ago.

Three details from the announcement:

- the feature is **experimental**;
- you turn it on in Codex's `config.toml` file (the Codex docs walk you through it);
- OpenAI plans to make it the default for Astra "in the coming weeks."

Why it matters: losing context is the number one reason an agent "forgets" an instruction and breaks something that worked. We cover that failure mode in [12 mistakes that sink vibe coding projects](https://vibe-code.blog/en/vibe-coding-mistakes), and the lasting fix in our guide to CLAUDE.md, AGENTS.md and rules files.

## Is Astra worth the price?

It depends entirely on how you use it.

**If you have a ChatGPT subscription**, 9to5Mac says Astra usage is included in existing allowances, with extra credits available to buy. You have nothing to lose by trying it in [Codex](https://openai.com/codex/).

**If your app calls the API**, the comparison is stark. Here are the vendors' own numbers, published a day apart:

| Model | API price (input / output, per million tokens) | DeepSWE v1.1 (vendor figure) |
| --- | --- | --- |
| GPT-6 Astra (OpenAI, Sept 3, 2026) | $10 / $50 | 74.1% |
| Gemini 3.8 Flash (Google, Sept 2, 2026, intro price) | $0.75 / $3.75 | 73.7% (reported by Android Authority) |

One benchmark does not sum up a model, and OpenAI mostly highlights Terminal-Bench 4.0 and computer use. But for an app feature that generates or fixes code in the background, the price gap is roughly 13x. We covered Google's model in [our Gemini 3.8 Flash article](https://vibe-code.blog/en/gemini-3-8-flash-vibe-coding), and overall budgeting in [what a vibe-coded app really costs](https://vibe-code.blog/en/vibe-coding-cost).

## What about security?

OpenAI says Astra is its first model to reach the "Critical" cybersecurity threshold in its Preparedness Framework. The practical result: the deployed version refuses advanced offensive tasks, such as writing a proof-of-concept exploit, while remaining usable for secure code review and patching. If Codex declines a security request, that is not a bug.

For you, it is a good moment to ask it for a security review of your own project. Here is a prompt to copy:

```text
Review this project as a security auditor. Without changing anything, list:
1) keys or secrets exposed client-side or committed to the repo,
2) routes or tables reachable without authentication,
3) user inputs that are not validated.
For each item: file, line, severity (high/medium/low) and a proposed fix.
```

Keep the list, fix the "high" items first, then run through the [pre-launch security checklist](https://vibe-code.blog/en/vibe-coding-security).

## What to do now

1. **Update Codex** (CLI, extension or app) and check that Astra shows up in the model picker. If not, wait for the staged rollout or ask your admin if it is an Enterprise account.
2. **Test on a copy of your project**, with a long task you know well: a migration, a refactor, a feature spanning several files.
3. **Turn on the experimental context memory** for that test only, and compare: did the agent stick to the instructions from the start?
4. **Do not move your production app to the Astra API** before you have costed it on real requests.
5. **Commit before every session**: a more autonomous model also makes bigger changes. [Git for vibe coders](https://vibe-code.blog/en/git-for-vibe-coders) shows how to roll back in one command.

To see how Codex stacks up, read [our full OpenAI Codex review](https://vibe-code.blog/en/openai-codex-review) and [the Cursor vs Claude Code vs Codex comparison](https://vibe-code.blog/en/cursor-vs-claude-code-vs-codex).

## Frequently asked questions

## Frequently asked questions

### Is GPT-6 Astra available on ChatGPT Plus?

Yes. Per OpenAI's announcement, Astra was due to reach Plus, Pro, Business and Enterprise subscribers in the days after September 3, 2026. 9to5Mac reports that Plus users got it a few hours after Pro and Business users, about a day after launch. Usage counts against your existing plan allowance.

### How much does GPT-6 Astra cost on the API?

OpenAI lists $10 per million input tokens and $50 per million output tokens, with cached input at $1. According to Forklog, requests above 272,000 input tokens cost more. A fast mode runs up to twice as fast at twice the standard price. At launch, it is one of the most expensive models on the market.

### How do I turn on the new context memory in Codex?

The feature is experimental and is enabled in Codex's `config.toml` configuration file, according to OpenAI. It lets Astra keep notes across context windows and search earlier ones. OpenAI plans to make it the default for Astra in the weeks after launch. Try it on a side project first.

### Why does Codex refuse some security requests with Astra?

OpenAI rates Astra at the "Critical" cybersecurity level in its Preparedness Framework, so the public version refuses advanced offensive tasks such as writing a working exploit. It can still review code, find flaws in your own project and propose fixes. Frame your request around defending your app.

### Should I switch from Claude Code or Cursor to Codex for Astra?

Not necessarily. A new model is not a reason to change tools and habits overnight. If you already pay for ChatGPT, try Codex with Astra on a real task and compare it with your usual tool. Keep whichever makes you start over least often, not whichever tops a benchmark.

## Sources

- [GPT-6 Astra — OpenAI (official announcement)](https://openai.com/index/gpt-6-astra/)
- [OpenAI releasing major upgrade to ChatGPT and Codex with GPT-6 Astra — 9to5Mac](https://9to5mac.com/2026/09/04/openai-releasing-major-upgrade-to-chatgpt-and-codex-with-gpt-6-astra-details-here/)
- [OpenAI opens limited access to GPT-6 Astra — Forklog](https://forklog.com/en/openai-opens-limited-access-to-gpt-6-astra/)
- [Introducing GPT-6-Astra — OpenAI Developer Community](https://community.openai.com/t/introducing-gpt-6-astra-the-most-intelligent-and-aligned-model-in-the-world/1394703)
- [Introducing Gemini 3.8 Flash and 3.8 Flash Cyber — Google](https://blog.google/innovation-and-ai/models-and-research/gemini-models/3-8-flash-and-3-8-flash-cyber/)
- [Google's Gemini 3.8 Flash is built to "work harder" — Android Authority](https://www.androidauthority.com/gemini-3-8-flash-google-ai-model-3706483/)

— Thierry Gustin, Formateur en intelligence artificielle (GustinAI) · November 21, 2026
