GPT-6 Astra: Features, Benchmarks, Pricing, and How to Access It

GPT-6 Astra in a Nutshell

• GPT-6 Astra is OpenAI's new frontier flagship, succeeding GPT-5.6 Sol, priced at $10/$50 per million input/output tokens.
• It scores 72.6% on OSWorld 2.0 computer use at ~47% less time per task than Sol, and saturates FrontierMath Tier 4 (97.6%) and ExploitBench (100%).
• The marquee ARC-AGI-3 99.9% depends on a stateful, expensive harness; stateless API calls score far lower.
• It crosses the Critical cybersecurity threshold, so exploit-creation capabilities ship gated behind OpenAI's Daybreak program.
• It trails Claude Fable 5.1 on Humanity's Last Exam with tools (57.2% vs 65.0%), so it is not a clean sweep.

OpenAI has launched GPT-6 Astra, which it calls the world's most intelligent and aligned model.

Astra lands into a crowded frontier tier where Anthropic's Claude Fable 5.1 (released September 1, 2026) and Claude Opus 5 are the reference points for coding and agentic work.

The headline claims are aggressive.

Astra saturates FrontierMath Tier 4 with a 97.6% score, saturates ARC-AGI-3 with a 99.9% score under OpenAI's provider adapter harness, and hits 100% on ExploitBench. It also sets a new frontier on computer and browser use, scoring 72.6% on OSWorld 2.0 at roughly 47% less time per task than its predecessor GPT-5.6 Sol.

Both ARC-AGI-3 and FrontierMath Tier 4 were designed specifically to stay ahead of AI capability, so saturating them signals something qualitatively different from beating a standard leaderboard.

In this article, I'll cover everything new with GPT-6 Astra, looking at the new features, exploring the benchmarks, and putting it through its paces with hands-on examples.

To see how the competition stacks up, check out our comparison of Claude Opus 5 and Claude Sonnet 5 and our guide to learning Claude.

What is GPT-6 Astra?

GPT-6 Astra is OpenAI's new frontier flagship, succeeding GPT-5.6 Sol as the top model for reasoning, computer use, and agentic work.

OpenAI positions it around three areas:

• State-of-the-art computer use
• A step change in professional work
• A jump in cybersecurity capability that crosses the Critical threshold under its Preparedness Framework.

The most useful benchmark headline for practitioners is OSWorld 2.0, where Astra scores 72.6% in about 40 minutes per task, compared with GPT-5.6 Sol's 65.7% at roughly 75 minutes.

That combination of higher accuracy and lower wall-clock time is the difference between an agent you babysit and one you can hand a task to.

Astra ships alongside an updated Codex harness that OpenAI says delivers 1.9x faster task completion versus the current GPT-5.6 Sol experience on the Mind2Web benchmark.

It is available as gpt-6-astra in the OpenAI API and through Amazon Bedrock, with a GPT-6 Astra Pro tier for Pro, Business, and Enterprise plans.

What's New With GPT-6 Astra?

Astra's changes cluster around agentic execution: using a computer, producing finished professional artifacts, maintaining context across long coding sessions, and staying within authorized boundaries.

Here are the capabilities that stand out.

Drive your computer end-to-end

Astra can operate a computer directly, handling the kind of tedious multi-step work that usually eats an afternoon.

Think filling out a batch of expense forms, updating records in a CRM, running frontend QA checks on a site you just built, or installing and troubleshooting software while watching what happens on screen.

The reason this matters is speed under real conditions.

On latency simulations on OSWorld 2.0, Astra hits 72.6% at roughly 40 minutes per task versus GPT-5.6 Sol's 65.7% at roughly 75 minutes, a 47% cut in time per task.

OSWorld 2.0 measures whether a model can actually operate a computer, such as navigating real interfaces, clicking, typing, and completing multi-step tasks the way a human would.

The 47% cut in time per task matters as much as the accuracy gain because agent cost scales with wall-clock time; a model that finishes in 40 minutes instead of 75 is not just faster, it is roughly half the price to run on the same workload.

That being said, independent ARC Prize testing shows Astra's headline computer-use and reasoning numbers are highly harness-dependent.

On ARC-AGI-3, the standard stateless harness produced scores from roughly 17% to 63% depending on reasoning tier, while the provider adapter harness that preserves reasoning state reached the marketed ~99.9%.

If you call the model statelessly, expect lower numbers than the launch charts.

Produce finished documents, slides, and spreadsheets

Astra is trained to output polished professional artifacts that follow your templates instead of generic first drafts.

It creates documents, presentations, spreadsheets, and analyses that match your writing and visual style, and it pulls only the context that matters into the output rather than padding it with everything it knows.

OpenAI's own demo has Astra building a slide deck about a fictional model from a handful of template slides, keeping the tone and layout consistent throughout.

For anyone who has spent an hour reformatting an LLM's markdown dump into a corporate template, that template adherence is the part I'd want to test first.

With Sites in ChatGPT, Astra can create, host, and share websites, web apps, and games directly from a prompt, applying stronger visual judgment to renderings than earlier models.

Ask focused questions instead of guessing

Astra decides when to ask you a question and when to proceed on sensible assumptions.

When instructions leave room for interpretation, it fills routine gaps on its own and asks focused questions only when the answer could change the outcome.

In OpenAI's side-by-side, GPT-5.6 Sol built a personal career website autonomously in 13 minutes 15 seconds, while Astra paused after 20 seconds to ask what career the user was moving into.

In Codex, Astra can ask asynchronously, continuing work that does not depend on your reply and waiting only on consequential decisions.

Astra also stays oriented as a task changes.

Earlier models sometimes treated a mid-task steering message as a brand new goal and lost the original constraints; Astra incorporates new requirements and answers side questions without dropping the broader job.

Keep notes across context windows in Codex

GPT-6 Astra introduces a new way for Codex to preserve and retrieve context when the context window fills, replacing repeated summarization with searchable notes.

Historically, models use compaction to compress long debugging sessions or large refactors into a single summary, which can drop details about why a fix failed or how a component behaves.

With Astra, Codex keeps notes across windows and leaves earlier windows searchable, so it can find a requirement or test result from an earlier message even if the note did not capture it.

You enable this experimental feature in your Codex config.toml, and OpenAI says it will become the default for Astra in the coming weeks.

Run defensive cybersecurity work

Astra reaches the Critical threshold in cybersecurity under OpenAI's Preparedness Framework, which is both its strongest new capability and its most constrained.

At launch, it will help with secure code review and patching, but it refuses more advanced tasks such as creating proof-of-concept exploits.

OpenAI plans to expand access through its Daybreak program with less restrictive safeguards, enabling vulnerability and proof-of-concept validation, malware analysis, and detection engineering.

Because of the raised risk, extra safety checks can pause or stop legitimate work: in ChatGPT or Codex you may be asked to review an action, and in the API the task will stop outright.

GPT-6 Astra Benchmarks

Astra sets new highs across computer use, math, coding, and cybersecurity in OpenAI's reported evaluations, often while using fewer output tokens than GPT-5.6 Sol or Claude models.

The scores below come from OpenAI's launch tables, so treat them as vendor-reported and note the harness caveats where they apply.

OSWorld 2.0 and computer use

Astra scores 72.6% on the OSWorld 2.0 offline set, versus 65.7% for GPT-5.6 Sol and 70.2% for Claude Opus 5.

OSWorld measures whether an agent can complete real desktop tasks like navigating apps and manipulating files, so it is the closest proxy to "can this thing actually use a computer for me."

On ScreenSpot-Pro, which tests grounding UI elements on a screen without tools, Astra hits 92.7% against GPT-5.6 Sol's 76.9% and Claude Fable 5's 87.3%.

On Agents' Last Exam, it reaches 59.3%, ahead of Claude Opus 5's 55.5% and GPT-5.6 Sol's 53.6%, while using roughly 65% fewer output tokens than Opus 5.

FrontierMath Tier 4 and GPQA Diamond

Astra scores 97.6% on FrontierMath Tier 4 v2, the toughest tier of a research-grade math benchmark, versus 87.8% for both Claude Fable 5.1 and Fable 5, and 73.2% for Claude Opus 5.

OpenAI describes this as saturation, and given the ceiling, that is a fair reading.

On GPQA Diamond, a set of graduate-level questions across biology, chemistry, and physics, Astra reaches 96.0% against Gemini 3.8 Flash's 95.3% and GPT-5.6 Sol's 94.6%.

It doesn't top all the metrics, however. On Humanity's Last Exam with tools, Astra scores 57.2%, behind Claude Fable 5.1's 65.0% and Opus 5's 63.6%, so it is not sweeping every reasoning eval.

Coding: Terminal-Bench and FrontierCode

On Terminal-Bench 4.0, which tests agents on software engineering, system configuration, and data analysis in a terminal, Astra scores 57.7% versus GPT-5.6 Sol's 37.3%, Claude Fable 5.1's 55.8%, and Gemini 3.8 Flash's 19.1%.

That is a wide lead over Gemini Flash and a modest one over Fable 5.1.

The coding picture is closer elsewhere.

On FrontierCode 1.1 Main, Astra scores 53.3% against Fable 5's 53.5% and Opus 5's 53.4%, effectively a tie, and on DeepSWE v1.1 it reaches 74.1% versus Gemini 3.8 Flash's 73.8% and Fable 5's 69.9%.

When we reviewed GPT-5.6 Terra in our Terra vs Claude Sonnet 5 comparison, Terra hit 87.4% on Terminal-Bench 2.1, a different version, so cross-version comparisons here are not apples to apples.

Cybersecurity: ExploitBench and SRE-Bench

Astra scores 100% on ExploitBench versus 78.5% for GPT-5.6 Sol and 70% for Claude Opus 5, and 42.4% on ExploitGym versus 30.3% for Sol.

ExploitBench measures whether a model can turn a known vulnerability into a working exploit, so a perfect score is exactly why OpenAI is gating this capability at launch.

On the contamination-controlled ExploitBench (June-August 2026), built from vulnerabilities in the prior three months, Astra scores 39.0% against Sol's 5.5%.

On SRE-Bench, which measures reverse-engineering binaries without source, Astra solves 88.0% in one attempt versus Sol's 55.9%.

Independent lab Irregular reported Astra solving 86 of 226 FrontierCyber challenges against 34 for Sol, including zero-day findings in browsers and a cloud database.

ARC-AGI-3 and long context

Astra scores 99.9% on ARC-AGI-3 under OpenAI's provider adapter harness, against 7.8% for GPT-5.6 Sol and 30.2% for Claude Opus 5.

This is the number to read carefully: ARC Prize's independent runs show the standard stateless harness scoring roughly 17% to 63% depending on reasoning tier, with the ~99.9% figure requiring the stateful adapter harness and a comprehensive run costing tens of thousands of dollars.

On long context, Astra hits 100% on OpenAI MRCR v2 8-needle in the 256K-512K band and 96.3% in the 512K-1M band, versus 91.5% and 73.8% for GPT-5.6 Sol.

That retrieval reliability at 1M tokens is a genuine step up over Sol for document-heavy pipelines.

Alignment benchmarks

Astra posts 2.4% on OpenAI's internal computer-use safety benchmark (lower is better) versus 22.0% for GPT-5.6 Sol, 9.5% for Fable 5.1, and 11.5% for Opus 5.

In realistic work environments, its overall misaligned-outcome rate without a confirmation policy is 3.4%, against 18.8% for Sol and 13.5% for GPT-5.6 Terra.

The one regression OpenAI flags is chain-of-thought monitorability.

Astra produces shorter, less verbose reasoning and controls its CoT far better than Sol; on CoT-Control between 750 and 1,250 tokens it followed constraints 60.9% of the time versus Sol's 16.1%.

UK AISI found Astra could evade monitoring under adversarial prompting, and OpenAI takes the decline seriously enough to name it a research priority.

GPT-6 Astra vs the Competition at a Glance

Here is how Astra's reported scores line up against the rivals it's aimed at across a few key benchmarks.

GPT-6 Astra Pricing and Availability

GPT-6 Astra is rolling out to a limited set of organizations first, then to all ChatGPT Plus, Pro, Business, and Enterprise users over the coming days, plus the OpenAI API and AWS.

Enterprise admins can enable it per workspace, and it is off by default at launch. Pro, Business, and Enterprise plans also get GPT-6 Astra Pro.

For developers, the model is available as gpt-6-astra in the OpenAI API and in Amazon Bedrock.

Standard API pricing is:

• Input: $10 per million tokens
• Output: $50 per million tokens
• Fast mode: up to 2.5x Standard speed at 2x Standard price (roughly $20 input, $100 output per million tokens)
• Separate rates apply to cache reads and writes

For context, that is well above GPT-5.6 Terra's $2/$12 rates from our earlier coverage, and above Claude Opus 5's $5/$25.

Astra is priced as a frontier reasoning and automation model, not a bulk-text workhorse.

It supports Zero Data Retention for eligible API customers, and usage is included within existing subscription allowances with the option to buy additional credits.

Final Thoughts

GPT-6 Astra is OpenAI making a statement that the next competitive front is agentic execution and computer use, not raw chat quality. 

The saturated math and abstract-reasoning scores are impressive, but the number I'd act on is OSWorld 2.0 at 72.6% in 40 minutes: an agent that finishes real desktop work faster and more accurately than GPT-5.6 Sol is the practical difference for most teams.

I'd temper the 'AGI' excitement with two caveats.

The marquee ARC-AGI-3 figure depends on a stateful, expensive harness, so stateless API callers should not expect ~99% out of the box, and on Humanity's Last Exam with tools, Astra actually trails Claude Fable 5.1 and Opus 5.

This is a strong model with a clear specialization, not a clean sweep across every metric.

The cybersecurity story is the one to watch.

Crossing the Critical threshold means Astra ships heavily gated, refusing proof-of-concept exploit creation until Daybreak access expands, and its safeguards can pause legitimate defensive work.

If your workflow touches security, budget for interruptions and read the system card before committing.

If you want to get hands-on with the models Astra is competing against, our guide to learning Claude is a good place to build practical intuition before you pick a frontier model for your stack.