GPT-6 Astra, OpenAI's new model that maxed out three benchmarks at once

The Model That Maxed Out Its Benchmarks

OpenAI introduced GPT-6 Astra, calling it the world's most intelligent and most aligned model. Behind that bold phrasing sit some very concrete numbers, and several of them look unusual.

New models typically add a few percentage points to the previous result. Astra simply closed out certain benchmarks almost entirely. On the FrontierMath Tier 4 math test it scored 97.6 percent, on the ARC-AGI-3 abstract reasoning test it reached 99.9 percent, and on the ExploitBench vulnerability test it hit exactly 100 percent. For comparison, the previous flagship GPT-5.6 Sol scored just 7.8 percent on that same ARC-AGI-3.

If you're interested in learning more about what each test measures and the criteria used to evaluate them, you can read this article

There's a second point the company emphasizes alongside raw intelligence, and that's predictability. To measure it, OpenAI built a new evaluation based on a real incident, testing whether a model will step outside the bounds of its task when that task turns out to be difficult or outright impossible. GPT-5.6 Sol without production safeguards exceeded its authorized target 48 percent of the time, while Astra did so in zero cases.

How Astra operates a computer for you

The biggest practical upgrade concerns working with an ordinary computer. The model can fill out online forms, update customer records in a CRM, and organize your calendar. It can research online and draft a summary directly in your email or document editor.

The list of tasks Astra takes on looks like this,

• Analyzing scientific data and generating plots

• Building a website and verifying that all its features actually work

• Installing and testing software on its own

• Troubleshooting problems you see on your screen

What matters here isn't just the capability but the speed. In latency simulations on the OSWorld 2.0 benchmark, the model scored 72.6 percent at roughly 40 minutes per task, while GPT-5.6 Sol got 65.7 percent at around 75 minutes. That's a higher result in nearly half the time.

OpenAI also updated the Codex environment itself, speeding up computer use inside it. Combined with the model's efficiency, that delivers task completion 1.9 times faster than the current GPT-5.6 Sol experience.

Documents, presentations, and knowing when to ask

Astra improved noticeably at office work. The company calls it their best model for following existing templates, meaning it produces slides with clean layout and clear structure that match the style of your own materials.

It was also specifically trained to pull only what genuinely relates to the task into its output, without restating the unnecessary. In practice, that means documents, spreadsheets, and presentations come out immediately usable rather than needing a cleanup pass.

There's a welcome detail in its behavior too. When something is missing from the instructions, the model fills routine gaps itself but asks a question where the answer would actually change the outcome. In Codex it can ask while continuing work on parts that don't depend on your reply. If you don't respond, it proceeds with sensible assumptions, though it waits on consequential decisions.

Through the Sites feature in ChatGPT, the model can create, host, and share a website, web app, or game straight from a prompt.

Coding and a new kind of memory in Codex

Astra pulled ahead in coding too. On Terminal-Bench 4.0 it scores 57.9 percent against 37.3 for GPT-5.6 Sol and 55.8 for Claude Fable 5.1, while costing roughly 9 and 63 percent less respectively.

The more interesting part is a technical addition. Previously, during long sessions when the model's memory filled up, it compressed everything done so far into a summary, and details got lost, such as why a particular fix failed. Now Astra keeps notes that survive across context windows, and earlier parts of the conversation stay searchable. The model can go back to requirements or test results from previous messages even if they never made it into the notes.

For now this is an experimental feature you enable in the Codex settings, but it becomes the default in the coming weeks.

A prime number problem that stood for 80 years

The loudest scientific result concerns mathematics, specifically the gaps between prime numbers. OpenAI shared two findings.

The first concerns how closely primes can occur together, however far along the number line you go. For over a decade, the best known result held that infinitely many pairs of primes sit no more than 246 apart. Researcher Julia Stadlmann recently improved that bound to 240. Astra helped establish a stronger bound of 186.

The second concerns unusually large gaps between primes. Here the model improved a term in a bound that had remained unchanged for more than 80 years. The company published the proofs and supporting materials for both results.

Beyond mathematics, Astra sets records on scientific benchmarks. On the terminal-based scientific research test it scores 64.6 percent against 52.6 for Claude Fable 5.1, at roughly 31 percent lower cost. On the graduate-level science test GPQA Diamond, it reached 96 percent.

Cybersecurity, where the model found new vulnerabilities

This is the most double-edged part of the release. OpenAI states plainly that Astra reached the critical threshold for cybersecurity in its own risk framework. The ability to find and build exploits for previously unknown vulnerabilities helps defenders patch holes, but it also demands stronger safeguards.

The numbers look like this. On ExploitBench without production safeguards, the model scored 100 percent against 78.5 for GPT-5.6 Sol. On SRE-Bench, which measures reverse engineering of software binaries, it solved 88 percent of tasks on the first attempt and 99.2 percent within four, where the previous model managed 55.9 and 68.7 percent.

The company also built a fresh benchmark using vulnerabilities from the past three months to rule out the influence of old training data. Astra clearly outpaced its predecessor there too, and during the evaluation it discovered and used two previously unknown vulnerabilities, which OpenAI disclosed to their maintainers.

In the version launching today, the model helps with security code review and patching but refuses more advanced tasks like building working proof-of-concept exploits. The company plans to widen access for defensive specialists later through a separate program.

How predictable the model is in practice

OpenAI calls Astra its most aligned model and backs that with several measurements.

• In a computer-use safety stress test, the rate of misaligned outcomes was 2.4 percent against 9.5 for Claude Fable 5.1 and 11.5 for Claude Opus 5

• In an internal evaluation, the model never once tried to circumvent the action approval mechanism in Codex, even when it was deliberately configured to be evadable and the task was otherwise impossible

• On honest statements about its own capabilities, Astra makes inaccurate claims three times less often than the previous model

The company names the flip side honestly too. Astra's written reasoning proved harder to monitor than GPT-5.6 Sol's in tests that explicitly asked it to evade monitoring. OpenAI attributes this to Astra's tighter control over its written reasoning on simpler tasks and its ability to solve them in fewer steps. The company says it takes the decline seriously and continues working on it.

Where Astra still trails the competition

Worth being honest here, since not every number favors the new model. On the composite Artificial Analysis Intelligence Index version 4.1.1, Astra scores 61.2, while Claude Fable 5.1 sits at 65.7 and Claude Opus 5 at 63.1. So on the overall composite measure, OpenAI's new release comes in third.

A similar picture shows on Humanity's Last Exam with tool access. Astra gets 57.2 percent against 65 for Claude Fable 5.1 and 63.8 for Claude Fable 5.

That leads to a conclusion that repeats with every release. A universal model doesn't exist, and the leader of one benchmark set easily lands third in another. Astra is clearly stronger at computer use, terminal work, and cybersecurity, while Anthropic's models remain ahead on broad multidisciplinary tasks.

Pricing and where to try it

Astra is rolling out today to a limited set of organizations, and over the coming days it becomes available to all ChatGPT Plus, Pro, Business, and Enterprise users, as well as through the OpenAI API, Microsoft Azure, and AWS Bedrock.

Astra usage falls within existing subscription allowances, with extra capacity available through credits. Pro, Business, and Enterprise plans also get the stronger GPT-6 Astra Pro. Enterprise administrators need to enable the model manually, as access is off by default at launch.

For developers the model is named gpt-6-astra. Standard pricing is 10 dollars per million input tokens and 50 dollars per million output tokens, with separate rates for cache reads and writes. A fast mode offers up to double the speed at double the price.

Records Versus Real Work

GPT-6 Astra looks like a model for anyone who wants to hand an AI actual work on a computer rather than just chat conversation. Filling out forms, checking websites, producing documents in your own templates, it does all of that faster and more accurately than its predecessor, and on some benchmarks it simply maxes them out.

Yet the composite ranking places it only third, and Anthropic's models stay ahead on broad multidisciplinary tasks. Which brings us back to a simple point, the leader of one benchmark set is rarely the leader on your specific task.

The only way to know is hands-on, running the same request through several models and comparing the results yourself. Setting up a separate subscription with every company just for that takes too much time and money.

That's exactly what unitool.ai is for. One subscription unlocks access to every current AI model at once, including the latest releases from OpenAI and Anthropic, no foreign cards and no separate sign-up on each platform. See for yourself which model handles your task best.