OpenAI GPT-6 Astra: Computer-Use Agent Behind Critical Gate

OpenAI GPT-6 Astra: The Computer-Use Agent Gated Behind a Critical Cyber Threshold

OpenAI released GPT-6 Astra — and it is deliberately not a chat model. The company calls it its most intelligent and aligned model, and positions it primarily as a computer-use system that operates software the way a person does: across browsers, spreadsheets, desktop applications, and terminals, finishing multi-step jobs instead of describing how to finish them. It also marks a first: Astra is the first model OpenAI has designated as reaching the Critical cybersecurity threshold in its Preparedness Framework, and that designation shapes who gets to use it at all.

What Is Actually New for Developers

The most interesting change for devs is context handling. Codex previously used compaction — summarising earlier turns once the context filled up. That process discards exactly the detail an agent later needs: why a fix failed, which tests ran, which requirement got added early. Astra instead keeps notes across context windows and searches back into earlier messages and tool output. The feature ships experimental behind a config.toml setting and becomes the Codex default in the coming weeks.

# config.toml — experimental note-based cross-window memory in Codex
[agent]
notes = true              # keep persistent notes across context windows
search_history = true      # search back into earlier turns and tool output
# compaction is deprecated in favour of notes-based retention

Astra can also ask the user a question while continuing work that does not depend on the answer — removing a common agent failure mode where one unresolved decision stalls an entire job.

Model Specs and Capabilities

On the developers page, Astra lists a 1,050,000-token context window, 128,000 max output tokens, and an April 30, 2026 knowledge cutoff. Input is text and image; output is text only. reasoning.effort adds two new levels above high — xhigh and max. Tool support covers computer use, hosted shell, apply patch, skills, MCP, and tool search. Fine-tuning is not supported.

The Benchmark Picture: Impressive Until Coding

OpenAI reports 72.6% on OSWorld V2-Offline against 65.7% for GPT-5.6 Sol, with average task time falling from roughly 75 minutes to 40. Anthropic reports 77.9% for Claude Fable 5.1 but explicitly says it used a different OSWorld release and should not be compared directly — a reminder that benchmark releases drift.

Astra scores 99.9% on ARC-AGI-3. That figure is worth a dose of scepticism: it was produced with a Responses API harness that retains reasoning between turns and uses compaction for long contexts, and OpenAI has previously shown those settings move ARC-AGI-3 substantially without changing the model. The result measures the model plus the agent system.

Other reported numbers: 97.6% on FrontierMath Tier 4, 95.9% on BenchCAD Vision2Code against 84.3% for Fable 5.1, and 64.6% on Terminal-Bench Science against Anthropic's 52.6%. For context, Epoch AI notes OpenAI funded FrontierMath and holds exclusive access to part of it — worth keeping in mind when reading the headline.

Coding is the weak spot in the story. Astra scores 74.1% on DeepSWE v1.1 versus 72.7% for Sol, Meta reports 75.4% for Muse Spark 1.3 at maximum reasoning, and the public leaderboard puts Gemini 3.8 Flash and Claude Opus 5 near 74%. On a 113-task benchmark, those gaps are one or two tasks. For a model pitched partly at terminal work, the coding-results being merely "in the pack" is the honest headline.

Cyber Capability Drives the Access Model

Astra is the first model to cross OpenAI's Critical cybersecurity threshold. In testing it developed exploits for hardened browsers and operating systems and found two previously unknown V8 vulnerabilities that OpenAI says it is disclosing to maintainers.

The consequences are practical. Standard access refuses advanced cybersecurity work including exploit discovery. For API developers, a cybersecurity safety check stops a task outright rather than pausing for approval. OpenAI's Mia Glaese warned that users outside trusted-access programs may hit slowdowns, pauses, or blocks — sometimes during unrelated work. In benchmark terms OpenAI reports 100% on ExploitBench and 42.4% on ExploitGym against 30.3% for Sol, with the usual six-hour time limit removed for both.

Availability and Pricing

Astra is a closed, hosted model with no released weights — self-hosting is off the table. It is live today only for organizations in OpenAI's Trusted Access and Daybreak cyber-defense programs, with API and AWS access arriving in the coming days.

Pricing: $10 per million input tokens, $50 per million output, cached input at $1.00. Requests above 272K input tokens bill at 2x input and 1.5x output for the whole request. Batch and Flex run at 50%; Fast mode at 2x. Pro, Business, and Enterprise users also get Astra Pro.

The Takeaway for Teams Building Agents

GPT-6 Astra is a strong signal about where agent infrastructure is heading: context persistence (notes beating compaction), non-blocking user interaction, and safety gating baked into the access tier instead of bolted on. The cyber threshold is the most consequential piece — a frontier lab now refuses certain requests at the API layer by default, which changes how you'll design agent pipelines that touch security-adjacent tasks. If you are building agents for regulated or security-sensitive work, plan for the reality that the model you call may refuse a step you considered benign, and architect your pipeline to degrade gracefully rather than stall.

Key Takeaways

• Astra is a computer-use model first: 72.6% OSWorld V2-Offline, task time down from ~75 to ~40 minutes.
• Notes replace compaction in Codex, so long agent runs stop losing failure detail.
• Coding gains are marginal: 74.1% DeepSWE v1.1 sits inside the leaderboard pack.
• First model at OpenAI's Critical cyber threshold; standard access refuses exploit work.
• No open weights, $10/$50 per million tokens, 1.05M context, API and AWS in coming days.

OpenAI's broader agent bets have consequences — see OpenAI Cuts Off Cursor After the SpaceX Acquisition for how the company enforces its own ecosystem boundaries. Read the OpenAI announcement and the GPT-6 Astra model page for the full spec.

Sources

• OpenAI: GPT-6 Astra: A new generation of intelligence
• OpenAI: GPT-6 Astra model page