ChatGPT 6 Astra: What GPT-6 Astra Actually Does, and Who Can Use It

ChatGPT 6 Astra: What It Actually Does, and Who Can Actually Use It

OpenAI shipped GPT-6 Astra on September 3, 2026. The pitch, in OpenAI's own words on X, was blunt: "Anything you can do on a computer, Astra can do for you. Fast."

That is a big claim, and the launch came wrapped in benchmark numbers, a staged rollout, a cybersecurity warning, and a plan-by-plan access chart that confuses almost everyone who reads it. So here is the plain version: what Astra is, what it is genuinely good at, what it costs, where you can actually use it, and why some safety researchers are unhappy about it.

First, the Name

There is no product called "ChatGPT 6." ChatGPT is the app. GPT-6 Astra is the model inside it.

It gets more confusing, because the same model shows up under three different names depending on where you find it:

• In ChatGPT Work and Codex, it is called GPT-6 Astra (and GPT-6 Astra Ultra at the highest reasoning setting).
• In the regular chat model picker, the top reasoning option is labeled GPT-6 Pro.
• In the API, the model id is gpt-6-astra.

Same underlying model, three labels. Keep that in mind when you go looking for it and cannot find it.

The Release Timeline

• September 3, 2026 — Limited release to approved organizations through OpenAI's Daybreak early-access program, which started with companies in its application-based cybersecurity program.
• September 4, 2026 — General availability begins rolling out to paid ChatGPT plans, the OpenAI API, and AWS.
• September 5, 2026 — OpenAI grants a full banked-usage reset to paid subscribers.
• September 6-7, 2026 — Users start reporting sharply tightened usage caps.

That last point matters, and we will come back to it.

What Astra Is Actually Built For

The headline is not "it writes better paragraphs." GPT-6 Astra is an agent model. It is designed to go do things on a computer rather than describe how you might do them.

Computer and Browser Use

This is the flagship capability. Astra clicks, types, navigates, fills forms, and works through multi-step tasks in a browser or a desktop environment. OpenAI reports roughly 2x faster computer automation than its previous model. Wikipedia's summary of the launch lists examples OpenAI demonstrated: preparing taxes, building a video game, ordering food, searching for jobs.

On OSWorld 2.0, the standard computer-use benchmark, Astra scores 72.6%, taking around 40 minutes per task. That is the number to hold on to: this is a model that works for tens of minutes on one job, not one that answers in three seconds.

Professional Documents

Astra is tuned to produce finished documents, spreadsheets, and presentations that follow an existing template rather than inventing a new style each time. It will also stop and ask you a clarifying question when it is missing something, instead of guessing and handing you a confidently wrong deck.

Codex Gets a Real Memory

The most practically useful change for developers is quiet. Codex now keeps searchable notes across a coding session instead of compressing old context into a lossy summary. So a requirement you stated at the start, or a test result from twenty minutes ago, can be retrieved verbatim later. Anyone who has watched a long agent session slowly forget the original brief will understand why this is a bigger deal than most of the benchmark numbers.

Cybersecurity, With a Lock On It

Astra is OpenAI's strongest security model, aimed at defensive work: reviewing code, finding bugs, patching, threat modeling. The offensive half of that capability is deliberately gated behind access controls. More on this below, because it is the most interesting story in the whole launch.

The Benchmark Numbers

OpenAI's reported scores:

Two footnotes worth knowing. The ARC-AGI-3 result requires a stateful harness; plain stateless API calls score lower. And FrontierMath Tier 4 at 97.6% essentially saturates that benchmark, which means it stops being a useful measuring stick from here on.

What People Are Actually Building With It

The benchmark numbers are one thing. The more informative signal in launch week was watching what people pointed Astra at once they had it.

A crowd-sourced tally on GitHub logged 235 reported use cases pulled from 21,435 tweets, and the category ranking is not what OpenAI's own announcement would lead you to expect:

The tally's author is explicit that these are reports in tweets, not verified successes — but the shape is clear. 3D and games dominated, well ahead of the office-document work the launch materials led with.

Some representative results:

• Blender, everywhere. Tom Krcha handed Astra an old steam-train drawing and got 3,295 fully editable objects back in a few minutes. Thomas Ricouard turned a Blender scene into a walkable Unreal Engine 5 house. Someone rebuilt a home bar from photos, including the wires. This was the breakout use case of the week, and it has its own full workflow guide here.
• An airplane landing simulator in 30 minutes from one prompt, with accurate dimensions and real physics.
• Playable browser games in Three.js — one builder started in Astra Pro's browser build, then pulled the project local and kept polishing with Astra High.
• GitHub Copilot and VS Code, where Astra went generally available for long-horizon agentic coding on September 4.
• Analytics dashboards — one user pointed it at a SaaS app's events, had it log into PostHog, and got 33 insights across 3 live dashboards.
• LEGO models through BrickLink Studio, with a Golden Gate Bridge first pass in about ten minutes.
• Historical simulations — Ethan Mollick asked for a grounded model of the Library of Alexandria and got the place, not an essay.

One caution before you calibrate on any of this. A widely shared post credited Astra with a 3D website that disassembles the human body into 2,234 interactive pieces; a Community Note clarified that the 2,234 pieces come from BodyParts3D, a pre-existing expert-built MRI dataset, and Astra built the React and Three.js viewer on top of it. Still impressive. Not the same claim. Launch-week demos are worth reading twice.

How It Compares to Claude and Gemini

This is where the story gets more honest, because independent evaluators do not fully agree with OpenAI's framing.

On the Artificial Analysis Intelligence Index, Claude Fable 5.1 leads at 66 (max effort), with GPT-6 Astra at 61 and Gemini 3.8 Flash at 59. On the Coding Agent Index, Fable 5.1 again leads at about 70.4, Astra at 67.0, Gemini 3.8 Flash at 61.1. And on DeepSWE, Astra's ~73-74% sits within noise of Gemini 3 Flash and slightly below Meta's Muse Spark 1.3 in some rankings.

Meanwhile OpenAI reports that Astra beats both its own GPT-5.6 Sol and Anthropic's Fable on cybersecurity benchmarks, and Astra's math and computer-use results are clearly out in front.

The disagreement is the useful information. Different evaluators measure different things under different settings, and no single index tells you which model will be better at your work. The rough read: Astra leads on computer use, math, and security; Fable 5.1 leads on general coding-agent work; Gemini 3.8 Flash leads on price by an enormous margin.

Where You Can Actually Use It

This is the part that generates the most confusion, so here it is directly.

ChatGPT Plus ($20/mo): You get GPT-6 Astra in ChatGPT Work and Codex. You do not get it in the main chat model picker. The chat picker's reasoning tiers run Instant, Medium, High, Extra High, Pro, and Astra lives under Pro, which Plus does not include. This is a plan limitation, not a rollout queue. Waiting will not change it.

ChatGPT Pro ($100 and $200), Business, Enterprise: You get GPT-6 Pro in the chat picker as well as Astra in Work and Codex.

Enterprise: Off by default at launch. An admin has to enable it per workspace.

API and AWS: Available as gpt-6-astra, and through Amazon Bedrock.

Usage Limits, and the Thing That Happened Two Days After Launch

Reported five-hour message estimates at launch:

• Plus and Business Standard: 5-45 messages
• Pro $100: 25-225
• Pro $200: 100-900

These are estimates, not fixed caps, and weekly limits may also apply. Note how low the Plus number is. Astra runs long agent tasks, and each one is expensive to serve.

Then, on September 6-7, users reported that OpenAI had cut Astra usage limits by up to 4x for heavy Plus, Pro, and Business users, roughly 48 hours after granting everyone a full usage reset. OpenAI has not officially confirmed those specific numbers, and separately acknowledged that compute is under real strain around Astra.

This is not a scandal so much as a predictable pattern: a much larger, more expensive model gets launched into subscription pricing that was set around cheaper models, and the caps get adjusted once real traffic shows what serving it actually costs. The practical lesson is simply not to build a workflow that assumes today's quota is permanent.

API Pricing

Per one million tokens:

• Input: $10
• Cached input: $1
• Cache write: $12.50
• Output: $50

Batch and Flex processing get a 50% discount. Fast mode runs up to 2.5x faster at 2x the cost. The context window is 1 million tokens, with 96.3% retrieval accuracy in the 512K-1M band. API rate limits start at 500 requests per minute and 500,000 tokens per minute at Tier 1 and rise with usage tier.

For scale: Astra's $10/$50 matches Claude Fable 5.1 exactly. Gemini 3.8 Flash is $0.75 input and $3.75 output through the end of 2026, roughly 13x cheaper. If your workload is high-volume classification or extraction, Astra is the wrong tool and the pricing sheet says so clearly.

Worth doing the arithmetic before you wire Astra into a loop. A long agent run that reads 200K tokens of context and writes 20K tokens of output costs about $3 per run. Prompt caching at $1 per million input tokens is the difference between a viable production agent and a surprise invoice.

The Controversy Nobody Should Skip

Two things about this launch made safety researchers uncomfortable.

First, the cyber capability. Astra is the first OpenAI model to hit the company's internal Critical threshold for cybersecurity. It scored a perfect 100% on ExploitBench, and on a harder in-house variant, OpenAI engineers reported that the model discovered and exploited two zero-day vulnerabilities. That is why the rollout started with a vetted access program rather than a public launch button, why exploit-creation is gated, and why OpenAI published separate posts about pacing model development around cyber-critical capabilities.

OpenAI also delayed this release after its July 2026 Hugging Face incident in order to add more safeguards.

Second, and less discussed: you can no longer fully read its reasoning. Astra uses a technique variously described as opaque recurrence, recurrent depth, or looped transformers. The practical consequence is that some or all of the model's reasoning no longer happens in readable text. Chain-of-thought monitoring, one of the main tools researchers use to audit why a model did something, gets substantially weaker.

OpenAI chief scientist Jakub Pachocki acknowledged the tradeoff directly, saying that as model capabilities increase, monitorability is getting more challenging. Greg Brockman's framing was that Astra is OpenAI's "most intelligent and, also very importantly, our most aligned model yet."

Both statements can be true. It is still a notable moment: the most capable model yet released is also the hardest one to inspect, and those two properties arrived in the same package.

Should You Use It?

Use Astra if your work looks like a long task rather than a question. Multi-step browser workflows, real codebase work in Codex where cross-session memory matters, hard research math, defensive security review, or documents that have to match a template.

Skip it if you mostly write, summarize, brainstorm, or answer quick questions. It will be slower and more expensive without being noticeably better, and on Plus you will burn a tight quota for no reason.

Do not build cost-sensitive volume on it. At $10/$50, high-throughput classification and extraction belong on a cheap fast model. Gemini 3.8 Flash exists for exactly this.

If you are on Plus, know before you subscribe: Astra lives in Work and Codex only. If you specifically want it in the chat picker, that requires Pro or above.

If you are on Enterprise, check with your admin. It is off until someone turns it on.

Frequently Asked Questions

Is it called ChatGPT 6 or GPT-6 Astra?

GPT-6 Astra. ChatGPT is the app; Astra is the model. There is no product named ChatGPT 6.

When was GPT-6 Astra released?

Limited access on September 3, 2026, with general availability rolling out from September 4.

Can I use GPT-6 Astra on ChatGPT Plus?

Yes, but only inside ChatGPT Work and Codex. It does not appear in the regular chat model picker on Plus, where the equivalent option is GPT-6 Pro and requires Pro, Business, or Enterprise.

Is GPT-6 Astra free?

No. It requires a paid ChatGPT plan or API access.

What does GPT-6 Astra cost through the API?

$10 per million input tokens and $50 per million output tokens, with cached input at $1 and cache writes at $12.50. Batch and Flex get 50% off; Fast mode costs 2x.

How big is the context window?

1 million tokens, with 96.3% retrieval accuracy reported in the 512K-1M range.

Is GPT-6 Astra better than Claude Fable 5.1?

It depends on the task. Astra leads on computer use, advanced math, and cybersecurity benchmarks. Fable 5.1 leads the Artificial Analysis Intelligence Index (66 vs 61) and Coding Agent Index (70.4 vs 67.0). They are priced identically.

Why is GPT-6 Astra considered controversial?

Two reasons: it is OpenAI's first model to reach the company's Critical cybersecurity threshold, scoring 100% on ExploitBench and finding two zero-days in internal testing; and it uses a reasoning technique that obscures its chain of thought, making it harder for researchers to audit.

Did OpenAI reduce Astra's usage limits?

Users reported caps cut by up to 4x on September 6-7, 2026, days after launch. OpenAI has not officially confirmed the specific numbers, but has acknowledged compute strain around the model.

Sources

OpenAI: GPT-6 Astra announcement

OpenAI: GPT-6 Astra system card

OpenAI: Responding to the next frontier of critical cyber capabilities

TechCrunch: OpenAI launches Astra, its powerful (and controversial) new model

CNBC: OpenAI begins rolling out Astra after warning of its advanced cyber capabilities

9to5Mac: OpenAI releasing major upgrade to ChatGPT and Codex with GPT-6 Astra

DataCamp: GPT-6 Astra features, benchmarks, and pricing

Artificial Analysis: Benchmarking GPT-6 Astra

Wikipedia: GPT-6 Astra