GPT-6 Astra Is On Every Plan: What It Costs, What It's Good At, and Which Effort Level to Use

GPT-6 Astra launched on September 3 at 12:32 PM PT to “a limited set of organizations.” By 3:52 PM PT the next day Altman posted “Now out to all Plus and Business users.” Twenty-seven hours from launch to every paid plan. This post is the practical version: where to get it, what it costs, whether it is better than what you have, and which effort level to run it at. I tested the last one myself.

Where You Get It

• ChatGPT: Plus, Pro, Business and Enterprise. In the model picker it is labelled “GPT-6 Pro,” not Astra.
• Codex CLI: -m gpt-6-astra, with model_reasoning_effort from low to max. Uses your ChatGPT plan.
• API: gpt-6-astra, $10 per million input, $50 output. Cache reads $1, four times Anthropic’s new rate. Fast mode doubles the price. Prompts over 272K input tokens pay double input and 1.5x output. Context 1.05 million, output 128K.
• Inside Claude Code: claude-launcher runs the Claude Code harness on other backends, and its Codex backend reuses the codex CLI login, so a ChatGPT plan pays for it. Put Astra in one role slot and keep Fable in another: /model codex:gpt-6-astra mid-session swaps the model and nothing else. The effort sweep below went through the same login, via the Codex CLI directly.
• Elsewhere: GitHub Copilot on Pro+, Max, Business and Enterprise at list price; OpenRouter; Vercel AI Gateway; AWS.

Is It Good?

Artificial Analysis rebuilt its index to version 4.2 with the launch, so the numbers below are all on the new scale.

Two points behind Fable 5.1 at 42% of the cost per task, on 30% of the tokens. On the public Terminal-Bench 4.0 board the two are within error, 58.2% against 57.9%, and Astra’s run cost $3.3k to Fable’s $6.2k. ARC-AGI-3 needs its own line: 99.9% through a “provider adapter” that preserves hidden reasoning state between turns, 62.7% through the standard harness.

The reviews agree with the table. Every: “a big upgrade from 5.6-Sol, with some frustrating habits that keep it from matching Fable at the top end,” and “Anthropic’s Fable still has better instincts for building a product.” latent.space ran it as an autonomous engineer for two days on 20 billion tokens and priced the result at under $6 an hour. The Hacker News launch thread, 1,003 points, was underwhelmed: “a very mundane release compared to GPT-4 and GPT-5… there wasn’t even a livestream.” A Codex user’s complaint that carries over: no auto mode, so it asks more than Claude Code does.

Computer Use

This is the part OpenAI leads with, and in my own use it is very good. The launch post calls Astra “the world’s best computer use model” and lists the jobs: filling forms, updating a CRM, organising a calendar, installing and testing software, running frontend QA on a site it just built. The numbers behind that, from OpenAI’s own table:

The speed claim matters more than the two-point OSWorld lead: about 47% less wall-clock time per task than Sol, and with the updated Codex harness OpenAI claims 1.9x faster completion on Mind2Web. Fable 5.1 has no OSWorld score in the row, and the launch-day footnote explained why with a dig: Claude’s scores use “the official settings, and not the modified tasks and modified grading from the Fable 5.1 System Card.” OpenAI has since reworded that footnote.

The safety numbers are the reason you can leave it alone. In OpenAI’s adversarial work-environment eval, messaging, email, browsing, project management and sales apps with ambiguous instructions and a safe route always available, Astra’s unauthorized-transaction rate was 6.8% against Sol’s 38.0%, data exfiltration 4.3% against 14.1%, destructive actions 0.0% against 2.9%. That is the model that deleted a database in July, measured against its successor, on a generic harness with no confirmation policy. Add Codex’s confirmation policy and the rate drops further.

Which Effort Level

Effort does not change the per-token price. It changes how many tokens the model spends. Artificial Analysis ran every level:

Low to medium buys three points for 53 cents. Xhigh to max buys one point for 72 cents. OpenAI’s own guide puts agentic coding and research at medium, complex debugging at high, and says to use xhigh “only when your evals show a clear benefit.”

I ran my own eval. One task, read-only, through Codex on my plan: review this blog’s post-scheduling code and report anything that could leak a future-dated post or hide a published one. I know that code, so I could grade the answers.

All seven findings were real. Low found five in a minute: scheduled pages ship with index, follow and no noindex; the OG image route builds cards for unpublished posts; the Brisbane date is computed by reparsing a localised string in the host timezone, which only works on a UTC builder; a +10:00 offset in frontmatter shifts a post’s day; and a Docker layer cache that could make the nightly rebuild ship yesterday’s listings. It wrote a small script to reproduce the timezone bug across three zones before reporting it.

Max found the same five, two more, and changed the ranking. The two new ones: the RSS feed stamps date-only posts at midnight UTC, and the daily cron is a single trigger with no retry. That one is not theoretical. Last night’s scheduled run fired at 17:19 UTC against a 14:01 schedule, three hours and eighteen minutes late. Max also demoted the Docker cache finding that low and high had ranked highest, and it was right to. I checked the deploy logs: the dependency layers come from cache, the build step runs fresh every night.

Date failures and the midnight disagreement reproduced in memory. Pagefind’s data-pagefind-ignore exclusion passed; homepage, index, and sitemap have no separate filtering bypass. Building future post URLs is intentional under this repository’s rules. No edits.

— GPT-6 Astra at max effort, final line of the review

So: low is the daily driver for review and lookup. Max is the audit, when a wrong ranking costs more than seven minutes. High sat between them on every measure and found nothing low had missed. On my task, medium is the level I would not bother with.

What It Refuses

Astra is OpenAI’s first model rated Critical for cybersecurity, and the default tier is built around that. Vulnerability discovery in source is allowed. Writing a proof-of-concept exploit is refused. The system card measures the gap the vetted tier opens:

• Vulnerability discovery and analysis: 66.7% completed by default, 100% with Daybreak Blue access.
• Proof-of-concept exploit creation: 2.4% by default, 92% with.
• Cyber red-teaming: 7.4% by default, 76.9% with.

OpenAI says Daybreak will “roll out less restrictive safeguards in the coming weeks.” Until then, if your work is authorised offensive testing, the default Astra is the wrong tool and so is default Fable. If your work is finding and patching, both will do it. OpenAI’s launch table makes the comparison itself: footnote 12 drops Fable from three biology benchmarks “because they refuse the majority of questions,” and footnote 17 says the Fable cyber scores it shows “come from Mythos, which is Fable with fewer safeguards.”

What You Cannot See

Astra reasons in a way that does not produce a readable chain of thought. TechCrunch calls the technique “opaque recurrence.” The practical effects:

• No thinking to read. In Codex you get actions and a summary, not the working.
• The ARC gap. The adapter that scores 99.9% works by carrying that hidden state between requests. Your harness will not, unless the provider exposes it.
• The monitor. OpenAI’s safeguard for a Critical model was chain-of-thought monitoring. Its card says Astra is “more capable of controlling its own CoT” than Sol, “can sometimes evade our internal monitors” under adversarial instruction, and at max effort some successful attacks “do not contain any chain of thought tokens at all.” OpenAI’s answer is a full-context monitor that watches actions instead. The UK AI Security Institute agrees the capability to evade exists and did not test whether it is used.

Should You Switch?

If you pay per token and run agents, probably. Astra does Fable-class work for less than half the cost per task, because it spends a third of the tokens, and low effort is a real setting rather than a degraded one. If you pay per month, the calculation is the plan meter, not the price, and Fable 5.1 is still first on the board with better product instincts by the reviews I trust. I will keep using both, in one Claude Code session through the launcher. The harness is the product. The model is a slot.

The wolf from August is on the shelf at the same price as the other wolf. It is a good wolf. Buy the effort level, not the tier.