OpenAI Releases GPT-6 Astra: A 1.05M-Context Computer-Use Model Gated Behind a 'Critical' Cyber Threshold
OpenAI released GPT-6 Astra, a 1.05M-context computer-use model positioned as its most intelligent and aligned system, operating software across browsers, spreadsheets, desktop apps, and terminals to complete multi-step jobs autonomously. Astra replaces Codex's context compaction with a "notes" system that preserves critical detail across context windows, and introduces the ability to ask users questions while continuing independent work in parallel. The model is the first designated at OpenAI's
Analysis
TL;DR
- OpenAI released GPT-6 Astra, a 1.05M-context computer-use model positioned as its most intelligent and aligned system, operating software across browsers, spreadsheets, desktop apps, and terminals to complete multi-step jobs autonomously.
- Astra replaces Codex's context compaction with a "notes" system that preserves critical detail across context windows, and introduces the ability to ask users questions while continuing independent work in parallel.
- The model is the first designated at OpenAI's "Critical" cybersecurity threshold, having discovered two previously unknown V8 vulnerabilities and developed exploits for hardened systems, restricting standard access for cybersecurity work.
- Benchmark performance shows 72.6% on OSWorld V2-Offline (down from ~75 to ~40 min task time), 99.9% on ARC-AGI-3, 97.6% on FrontierMath Tier 4, and 74.1% on DeepSWE v1.1 — with coding gains being marginal compared to competitors.
- Astra is a closed, hosted-only model with no open weights, priced at $10/M input tokens and $50/M output tokens, available exclusively through OpenAI's Trusted Access and Daybreak programs.
Why It Matters
GPT-6 Astra represents a strategic pivot by OpenAI toward agentic, computer-use capabilities rather than pure chat, signaling that the next competitive frontier in AI is autonomous task completion across real software environments. The "Critical" cybersecurity designation and restricted access model demonstrate how advanced AI capabilities are becoming gated behind trust frameworks, creating both security risks and business barriers for broader deployment. For practitioners, the notes-based context system and parallel questioning feature address two of the most persistent failure modes in long-running agent workflows.
Technical Details
- Context Architecture: Astra features a 1,050,000-token context window with 128,000 max output tokens. It replaces Codex's compaction mechanism (which summarized and discarded earlier turns) with a "notes" system that preserves critical details across context windows and enables searching back into earlier messages and tool outputs. This ships experimentally via a
config.tomlsetting before becoming the Codex default. - Parallel Questioning: The model can ask users clarifying questions while continuing to execute work that does not depend on those answers, eliminating a common agent failure mode where a single unresolved decision stalls an entire multi-step job.
- Model Specifications: Input accepts text and image; output is text-only. New
reasoning.effortlevels added abovehigh:xhighandmax. Tool support includes computer use, hosted shell, apply patch, skills, MCP, and tool search. Fine-tuning is not supported. Knowledge cutoff is April 30, 2026. - Benchmark Results: OSWorld V2-Offline: 72.6% (vs. GPT-5.6 Sol's 65.7%); ARC-AGI-3: 99.9% (via Responses API harness with compaction); FrontierMath Tier 4: 97.6%; BenchCAD Vision2Code: 95.9% (vs. Fable 5.1's 84.3%); Terminal-Bench Science: 64.6% (vs. Anthropic's 52.6%); DeepSWE v1.1: 74.1% (vs. Sol's 72.7%, Meta's Muse Spark 1.3 at 75.4%); ExploitBench: 100% (capability-coverage score); ExploitGym: 42.4% (vs. Sol's 30.3%).
- Cybersecurity Capabilities: First OpenAI model to reach the "Critical" threshold in its Preparedness Framework. Developed exploits for hardened browsers and operating systems; disclosed two previously unknown V8 vulnerabilities. Standard access now refuses advanced cybersecurity work outright rather than pausing for approval.
Industry Insight
- Agentic AI is the new battleground: OpenAI's positioning of Astra as a computer-use model first — not a chat model — signals that autonomous task completion in real software environments is becoming the primary differentiator. Practitioners should prioritize integrating agents with robust context preservation and parallel execution capabilities.
- Security gating will fragment access: The "Critical" cybersecurity designation creates a two-tier ecosystem where advanced security research capabilities are restricted to trusted organizations. Teams relying on OpenAI's API for cybersecurity work should anticipate slowdowns, pauses, or blocks, and consider alternative paths or early enrollment in Trusted Access/Daybreak programs.
- Coding performance is plateauing: Astra's marginal gains on DeepSWE (74.1% vs. 72.7%) and parity with competitors like Gemini 3.8 Flash and Claude Opus 5 suggest that raw coding ability is converging across top models. Competitive advantage will increasingly come from agentic workflow features, context management, and ecosystem integration rather than benchmark scores alone.
Disclaimer: The above content is generated by AI and is for reference only.