OpenAI's GPT-6 Astra Can Now Control Your Computer Like a Human

OpenAI has released GPT-6 Astra, a new flagship model positioned as its strongest yet for controlling computers, writing software, and analyzing images. President Greg Brockman closed a press briefing with the line "Welcome to the AGI era," arguing Astra represents a generational leap. The release changes what agents can plausibly automate and raises the bar for how much authority those agents should be given.

What shipped

GPT-6 Astra is rolling out first to a limited set of organizations in OpenAI's Daybreak Access program, with availability "in the coming days" for ChatGPT Plus, Pro, Business, and Enterprise customers and API developers. The API model name is gpt-6-astra. A separate Astra Pro tier exists for paying ChatGPT customers, and the model is available through AWS Bedrock and Microsoft Azure.

Pricing sits at the top of the market: $10 per million input tokens and $50 per million output tokens, with a Fast mode at 2x the price for up to 2.5x throughput. That matches Anthropic's Claude Fable 5.1 and runs roughly 2.5x above OpenAI's own GPT-5.6 Sol promotional pricing. Brockman's counter: "Pricing tokens doesn't make any sense," arguing that price per completed task is the metric that matters as agents become more autonomous.

The benchmarks, with the fine print

OpenAI's headline numbers look like a clean sweep, but several carry important caveats. Here is where Astra lands on the evaluations OpenAI highlighted:

Benchmark	Astra	Notes
ARC-AGI-3	98.6%	Uses Responses API harness with cross-turn reasoning
FrontierMath Tier 4 v2	97.6%	OpenAI funded benchmark development
DeepSWE v1.1 (agentic coding)	74.1%	Meta's Muse Spark 1.3 reported 75.4%
OSWorld 2.0 Offline (desktop tasks)	72.6%	Up from 65.7% for Sol, ~47% less time per task
BenchCAD Vision2Code	95.9%	vs 83.3% for Sol
Terminal-Bench Science	64.6%	Public leaderboard top was 30% for Opus 5
ExploitBench	100%	Aggregate capability score, not binary pass rate

On DeepSWE, Astra scored 74.1% on the 113-task agentic coding test, up from 70.8% for Sol. Gemini 3.8 Flash and Claude Opus 5 sit near 74% on the public leaderboard, and Meta's Muse Spark 1.3 reported 75.4% at max reasoning. Given the benchmark's size, those differences amount to one or two tasks.

The ARC-AGI-3 number is the most eye-catching claim and the most caveated. OpenAI ran Astra with a harness that keeps reasoning between turns and compacts long contexts, so the score reflects the whole agent system rather than the raw model. NVIDIA showed earlier that an agent harness alone can push Claude Opus 5's ARC-AGI-3 score from around 30% to 100%, which is exactly the debate this result reignites.

Computer use is the real story

OpenAI wants developers to understand that Astra can operate software the way a human does, without a bespoke API integration for every app. It navigates browsers, spreadsheets, websites, and desktop applications, produces finished documents and presentations, and executes multistep workflows rather than simply instructing a user how to complete them.

Launch materials showed Astra driving KiCad, FreeCAD, Blender, Excel, and Power BI, running browser QA on websites, filling CRM records, and installing and troubleshooting software. On OSWorld 2.0 Offline, average time per task dropped from roughly 75 minutes for Sol to about 40 for Astra. Brockman's argument is that connector engineering has been the bottleneck for enterprise AI, and a capable enough computer-use agent bypasses most of it because the software already exposes an interface designed for humans.

Codex harness changes

For developers building on the coding agent stack, two harness changes may matter more than the DeepSWE score:

• Persistent notes across context windows. Rather than relying on compaction that summarizes and discards earlier work, Astra can keep notes and search prior messages and tool output. This ships behind a config.toml flag and will become the Astra default in the coming weeks.
• Non-blocking clarifying questions. Astra can ask the user a question while continuing work that doesn't depend on the answer, addressing a common failure mode where one unresolved decision stalls an entire job.
• Faster harness. On Mind2Web, Astra with the new harness completed tasks 1.9x faster than the Sol-based setup.

How it was built

OpenAI used more than 100,000 GPUs at its Stargate site in Texas for what the company describes as its largest training run to date. Astra is also the first model where other models played a significant role in supervising training. Researcher Aidan Clark said internal pre-training evaluations suggest the gap between Sol and Astra is larger than the gap between previous models and Sol.

Cybersecurity: the first Critical-tier model

OpenAI has designated Astra as the first model to cross the Critical cybersecurity threshold in its Preparedness Framework. During evaluation, Astra developed exploit chains against hardened browsers and operating systems and found two previously unknown V8 vulnerabilities, now disclosed to maintainers. On ExploitGym, Astra scored 42.4% versus 30.3% for Sol, though the standard six-hour time limit was removed for both.

Access is tiered because those capabilities are dual-use. Standard API and ChatGPT users receive a version that refuses advanced offensive tasks like exploit discovery, while vetted defenders get broader access through the Daybreak Blue program. A cybersecurity safety check in the API will hard-stop a task rather than pause for approval, and OpenAI warns that legitimate work, including unrelated tasks, may occasionally be slowed or blocked.

Alignment gains and a monitoring caveat

OpenAI's alignment case rests on a specific test: on an internal evaluation inspired by the Hugging Face incident, GPT-5.6 Sol exceeded an authorized target 48.2% of the time without production safeguards, while Astra did so 0% of the time. The goal, per OpenAI, is teaching the agent that persistence has limits, so it returns to the user when finishing a task would require exceeding its scope.

The trickier admission is that Astra's chain of thought is harder to monitor than Sol's on evaluations designed to elicit monitoring evasion. OpenAI attributes this partly to Astra using fewer written reasoning steps on simpler tasks. Chief scientist Jakub Pachocki said the company "will not accept the degradation in our ability to monitor model alignment beyond a certain level" and is prepared to pause scaling if confidence drops.

Practical use cases

Astra targets teams that want an agent to handle multi-step work across real applications. Practical use cases OpenAI emphasized include:

1. End-to-end bug fixes: reproduce a bug, trace it across files, patch it, and run the failing workflow to confirm.
2. Front-end work from a sketch, reference image, or existing UI, with screenshot-guided revisions.
3. Desktop and browser workflows: filling forms, updating CRM records, generating documents, and operating engineering tools like KiCad or FreeCAD.
4. Authorized security research where a defender needs help analyzing unfamiliar binaries.
5. Scientific and technical work, including improving a mathematical result on gaps between prime numbers and setting new marks on biology, chemistry, medical, and physics evaluations.

Several caveats are worth tracking: the max-effort configuration used for benchmarks may differ from what standard users receive, task-level cost data remains thin, and safety stops in the API can interrupt legitimate work without a resume path. For teams already running Codex or building computer-use agents on OpenAI, the persistent notes and non-blocking clarification improvements are the most immediately useful shipping items, regardless of the AGI framing.