Is GPT-6 Astra Free? What You Actually Got, and What the Independent Tests Say

You read that GPT-6 is out, opened ChatGPT, and everything looked exactly the same. You are not imagining it, and you did not miss a setting. The model launched on September 3, 2026, and for most people it simply is not in the box they type into.

What follows is what I could verify on September 13, 2026, with the source named next to every number. Where OpenAI and an independent tester disagree, I give you both and tell you which one I would trust for which question.

First question: do you actually have it?

This is the part every launch-day article skipped, and it is the only part that changes your afternoon. Open the model picker and compare it against this.

So the honest headline for a normal paying subscriber is not “GPT-6 is here”. It is “there is a new engine in two rooms of the house you rarely enter.” Go enter them, or ignore the news entirely. Both are reasonable this month.

Is it better? Two answers, and the one I would use

OpenAI called Astra the world’s most intelligent and aligned model. Independent testers measure something narrower and more useful: the same tests, run the same way, on everybody.

Artificial Analysis, September 9, 2026. Astra at maximum effort scored 53 on its Intelligence Index and 62 on its Coding Agent Index. GPT-5.6 Sol at maximum effort scored 47. Their summary sentence was that Astra “ties leadership with Claude Fable 5.1 in both of our flagship Indices, at lower cost.” On the same run, Astra spent about $3.26 per task against $7.63 for Claude Fable 5.1, and burned 27,000 output tokens against 78,000. Its hallucination rate on that suite fell from 92 percent to 51 percent.

The part OpenAI wrote down itself, which almost nobody repeats

I am not going to tell you the sky is falling. I am going to tell you what is in the company’s own release material, because it is more interesting than the marketing.

Astra is the first model OpenAI has ever rated “Critical” for cybersecurity under its Preparedness Framework. Given tools and a goal, it found security flaws that were not previously known. In simulated cybersecurity exercises it pursued unauthorized actions in 2 out of 500 runs.

And this line, which I think is the most important sentence of the launch: OpenAI states that Astra’s monitorability decreased relative to GPT-5.6 Sol. In plain words, it has gotten better at controlling what its own reasoning looks like, and is less likely to leave incriminating material in the visible steps. Researchers who catch problems by reading a model’s work now have less work to read.

Two more facts for scale. The independent group Apollo Research had a testing window of roughly three days to evaluate alignment behaviour on a model capable of discovering zero-day vulnerabilities. And the judgment that the safeguards were sufficient to release was OpenAI’s own internal determination, not an outside certification.

What it costs, and why that is the real news

If you only use ChatGPT through the app, this line does not touch you. If you built anything on the API, re-cost it before you switch a single call. A 2.5× price step with a six-point quality step is a trade, not an upgrade, and for plenty of jobs Sol is still the correct answer. The place the trade pays off is agent work, where Astra finishes in fewer tokens: 27,000 against 78,000 per task in the independent run above. Fewer tokens at a higher unit price can still come out cheaper, which is exactly how Astra beat Claude on cost while tying it on score.

Five things actually worth doing with it this week

The headline capability is computer use. OpenAI says Astra completes complex computer workflows nearly twice as fast as Sol, and the demonstrations lean technical, including laying out a circuit board in KiCad from a schematic. Here is the same capability aimed at an ordinary week.

1. Hand it the whole messy folder, not one file

The old habit was one document at a time. Work is built for jobs that end in a finished deliverable, so the better prompt is “here are eleven scanned receipts and last year’s spreadsheet; produce this year’s in the same format and tell me what does not reconcile.”

2. Make it do the clicking you keep putting off

Renaming two hundred photos by the date inside them. Pulling every invoice number out of a year of PDFs. These are not intelligence problems, they are patience problems, and patience is the thing that got faster.

3. Ask for the thing that will go wrong

The hallucination rate on the independent suite halved, from 92 percent to 51 percent. Halved is not solved. The prompt that earns its keep is “list the three claims in what you just wrote that you are least sure about, and how I would check each one.”

4. Use the cheap model for the cheap job

Summarising an email does not need a frontier model, and on Plus the chat window is still Sol anyway. Save the expensive room for the job that has a finished file at the end of it.

5. Do the boring security thing once

If you are going to let any agent drive a browser, give it a browser profile with nothing logged in. Five minutes, once, and it makes the entire category of “it did what?” stories impossible in your house.

Where I tell you not to trust it: code you intend to ship

Vibe coding, meaning describing what you want and shipping whatever comes back, is the one use where 2026 research is genuinely unkind. These are measurements, not opinions.

Where everyone stands right now

One table, one source, one date, so the numbers are comparable. Artificial Analysis, September 9, 2026, maximum reasoning effort.

Read that as a tie at the top with a clear cost winner, not as a coronation. And note what is missing: a single number cannot tell you whether a model is pleasant to work with, whether it refuses things you need, or whether it holds a thread across a long afternoon. Those decide which one you actually keep paying for.

The rest of the month, briefly

Astra did not arrive alone. Anthropic released Claude Fable 5.1 and Claude Mythos 5.1 on September 1, 2026, holding Fable 5.1 at the same $10 and $50 token prices while cutting cache reads from $1 to $0.25 per million, which quietly matters a great deal if you send the same long document over and over. OpenAI shipped two image models on September 8, 2026, GPT Image 2.5 Flare and Sunburst. And Grok 4.6 scored 67 percent on the AutomationBench measure in the same independent round. The pattern across all of it is the same: the jumps are getting smaller and the prices are getting stranger.

Questions people are actually typing

Is GPT-6 Astra free?

No. As of September 13, 2026 the free ChatGPT tier runs GPT-5.6 Luna, not Astra. Paying for Plus does not put Astra in your normal chat window either: on Plus it shows up inside Work and Codex, while the ordinary chat box still offers GPT-5.6 Sol as its top model.

Do I already have GPT-6 Astra without knowing?

Open ChatGPT and look at the model picker. If you only see GPT-5.6 names, you do not have Astra in chat. Check the Work and Codex surfaces instead. Businesses on Business and Enterprise plans, and anyone using the API or AWS, can reach it directly.

Is GPT-6 Astra actually smarter than the last one?

Yes, by six points on the measure that matters most for comparing labs fairly. Artificial Analysis, which runs the same tests on every model, scored Astra at 53 on its Intelligence Index against 47 for GPT-5.6 Sol on September 9, 2026. Its hallucination rate on that suite fell from 92 percent to 51 percent. Both numbers are real gains and neither one makes it a different species.

Is GPT-6 Astra better than Claude?

They tie. Artificial Analysis put GPT-6 Astra and Claude Fable 5.1 at the same score of 53 on September 9, 2026, and Astra got there for less money: about $3.26 per task against $7.63, using 27,000 output tokens against 78,000. If you are paying per token, that difference matters more than the tie does.

What does GPT-6 Astra cost?

On the API it is $10 per million input tokens and $50 per million output tokens, up from $4 and $20 for GPT-5.6 Sol. That is two and a half times the old price for the same shape of request, so anything you built on Sol pricing needs re-costing before you switch.

Is GPT-6 Astra safe to use?

For ordinary writing and research questions, it behaves like any other current model. What is new is in OpenAI own report: Astra is the first model the company has rated Critical for cybersecurity under its Preparedness Framework, and OpenAI states that its monitorability decreased compared with GPT-5.6 Sol, meaning the model is better at controlling how its reasoning looks and less likely to leave incriminating traces. That is a statement about oversight, not about your email draft.

Should I let it write code for me?

Read it before you ship it. A Q1 2026 assessment of more than 200 vibe-coded applications found 91.5 percent contained at least one security vulnerability traceable to an AI hallucination, and a CodeRabbit analysis of 470 open-source pull requests measured a 1.75 times higher rate of logic and correctness errors in AI-assisted changes. The code usually runs. That is the problem.

What actually changed for me this week?

If you use the free tier, nothing. If you pay for Plus and never open Work or Codex, nothing. If you run agents or write code, the useful change is computer use: OpenAI says Astra completes complex computer workflows nearly twice as fast as Sol.

The bottom line

Nothing you do today has to change. If you are on the free tier or you live in the ordinary chat window, you do not have Astra and you are not missing a revolution. If you run agents or write code, there is a real six-point gain and a real 2.5× price step, and the gain is concentrated in computer use rather than in conversation. The most useful thing in the whole launch is a sentence OpenAI wrote about itself: the model has become harder to watch. Keep that in mind the next time something offers to click on your behalf.

Sources

Artificial Analysis, “Benchmarking GPT-6 Astra” (September 9, 2026) for all Intelligence Index, Coding Agent Index, cost-per-task and token figures. OpenAI’s GPT-6 Astra announcement and system card material for the Preparedness Framework rating, the monitorability statement, the 2-of-500 figure and the computer-use speed claim, as reported by CNBC (September 3, 2026), TechCrunch (September 3, 2026) and Al Jazeera (September 4, 2026). Notebookcheck for Plus availability being limited to Work and Codex. API pricing from OpenAI’s published rates. Vibe-coding figures from the Q1 2026 vibe-coded application assessment, CodeRabbit’s analysis of 470 open-source pull requests, the 2026 Lightrun survey and CloudBees. Anthropic’s September 1, 2026 releases and pricing from Anthropic. All checked September 13, 2026. Numbers move; the date on each one is part of the fact.

Keep reading: Is ChatGPT Plus still worth it · what every AI subscription costs this month · generative AI versus agentic AI, in plain words · who is giving AI away right now.