GPT-6 Astra
by OpenAI
OpenAI's frontier model for autonomous computer use, gated cyber capability and long-horizon coding
GPT-6 Astra is OpenAI's flagship frontier model, released in September 2026 for enterprises that need agents to drive computers, reason across large codebases and run security work autonomously. It targets platform, security and engineering teams who have hit the ceiling of cheaper models on long-horizon, multi-step tasks and can absorb a premium per-token price.
GPT-6 Astra is OpenAI's frontier reasoning model, which began rolling out on September 3, 2026 through the company's enterprise Trusted Access Program before reaching the OpenAI API, ChatGPT Plus, Pro, Business and Enterprise plans, Microsoft Azure and Amazon Bedrock. It succeeds GPT-5.6 Sol and is positioned around three capabilities rather than general chat: autonomous computer and browser use, offensive and defensive security work, and long-horizon software engineering across large codebases. OpenRouter lists the model with a 1,050,000-token context window and 128,000-token maximum output, accepting text, images and documents such as PDFs and returning text, with function calling and JSON-schema structured outputs. Reported benchmarks include 72.6% on OSWorld 2.0 computer use, 96.0% on GPQA Diamond, 97.6% on FrontierMath Tier 4 and 57.7% on Terminal-Bench 4.0. OpenAI gated the model's exploit-development capability at launch after it reached a Critical cybersecurity threshold, restricting the strongest use to its Daybreak security program. The launch is contested on two fronts. TechCrunch reported that Astra uses an opaque recurrence reasoning technique that degrades chain-of-thought monitorability, a regression flagged by the UK AI Safety Institute. ARC Prize's independent evaluation scored it 62.7% on ARC-AGI-3 under a standard harness against the 99.9% obtained with OpenAI's proprietary provider adapter, at roughly $19,000 to $26,000 in inference per benchmark run. API pricing is $10 per million input tokens and $50 per million output, about 2.5x GPT-5.6 Sol.
Platform and security engineering leaders who already run agents in production and have specific tasks — cross-file code review, GUI automation of legacy apps, authorized security testing — where cheaper models measurably fail rather than merely underperform.
A frontier model that finishes long-horizon computer-use and cross-codebase tasks other models abandon, at a price that only pays back when it is routed to those tasks selectively.
At a Glance
- Category
- AI Models & APIs
- Pricing
- Usage-based, Subscription, Contact for pricing
- Target Market
- CTOs, CISOs, Enterprise Developers, Data Scientists, VP Engineering
- Deployment
- API-based, Cloud-only, Multi-cloud
- Founded
- 2015
- Headquarters
- San Francisco, United States
- Team Size
- 500+
Key Features
- ✓Autonomous computer and browser use
Drives desktop and browser workflows end to end, scoring 72.6% on OSWorld 2.0 and completing tasks about 47% faster than GPT-5.6 Sol.
- ✓~1.05M-token context window
Accepts roughly 1,050,000 input tokens and returns up to 128,000, with 96.3% reported accuracy on 8-needle retrieval between 512K and 1M tokens.
- ✓Gated cybersecurity capability
Scores 100% on ExploitBench, but OpenAI restricts exploit generation to its Daybreak program so defenders get the capability under contract rather than openly.
- ✓Cross-file software engineering
Independent CodeRabbit testing measured 57.1% actionable bug coverage on complex cross-file reviews, 20% above Sol and 33% above Claude Opus 5.
- ✓Fast mode API tier
Trades price for latency, delivering roughly twice the standard processing speed at twice the standard per-token cost for time-critical agent runs.
- ✓Enterprise rollout controls
Supports zero data retention for eligible API customers and stays off by default until a workspace administrator explicitly enables it.
- ✓Structured output and tool calling
Supports function calling via tools and tool_choice plus JSON-schema-constrained responses, so agent output parses reliably downstream.
Capabilities
Use Cases
- •Defensive vulnerability research
Security teams inside OpenAI's Daybreak program use Astra to find and triage bugs in their own codebases before attackers reach them.
- •Selective large-codebase code review
Engineering organisations route only hard cross-file pull requests to Astra, because its advantage over cheaper models narrows sharply on routine reviews.
- •Legacy GUI automation
Operations teams hand Astra multi-step desktop and browser workflows in applications that expose no API, replacing brittle recorded RPA scripts.
- •Whole-corpus document analysis
Legal, finance and compliance teams load entire contract sets or filings into the million-token context and query across all of them in a single pass.
- •Agentic coding inside Codex
Developers run Astra in Codex, where it can ask clarifying questions asynchronously without losing the state of a long-running task.
Ideal For
Best For
- ✓Cross-file code review and refactoring where a change's blast radius is spread across a large repository
- ✓Authorized defensive security research and vulnerability triage under OpenAI's Daybreak program
- ✓GUI and browser automation of legacy enterprise applications that expose no API
- ✓Single-pass analysis of very large document sets using the roughly 1M-token context window
- ✓Long-horizon agent tasks where cheaper models lose task state after a handful of steps
Not Ideal For
- ✗Cost-sensitive high-volume workloads — at $10/$50 per million tokens it is roughly 2.5x GPT-5.6 Sol, and CodeRabbit measured only a 4% actionable-bug-coverage gain on routine code reviews
- ✗Teams with a hard chain-of-thought auditability requirement — Astra's opaque recurrence reduces reasoning monitorability, a regression the UK AI Safety Institute flagged at launch
- ✗Latency-sensitive interactive products — llm-stats measured p95 time-to-first-token at roughly 13.7 seconds
- ✗Buyers who need reproducible third-party benchmarks — ARC Prize's standard-harness result was 62.7% against the 99.9% headline figure
Integrations
Deployment
Market Analysis
Pros
- ✓Genuine step change on autonomous computer use — 72.6% on OSWorld 2.0 with roughly 47% faster task completion than GPT-5.6 Sol
- ✓Its largest measured advantage lands exactly where cheap models fail: 57.1% coverage on complex cross-file code review, 20% above Sol in independent CodeRabbit testing
- ✓Roughly 1.05M-token context with 96.3% 8-needle retrieval accuracy between 512K and 1M tokens
- ✓Enterprise controls are in place at launch — zero data retention for eligible API customers, admin-gated rollout, and delivery through Azure and Amazon Bedrock
Cons
- ✗The headline 99.9% ARC-AGI-3 score depends on OpenAI's proprietary provider-adapter harness; ARC Prize's own standard harness produced 62.7%, and the runs cost roughly $19,000-$26,000 each
- ✗2.5x the token price of GPT-5.6 Sol for only about a 4% actionable-bug-coverage gain on ordinary code reviews, per CodeRabbit's measurement
- ✗Opaque recurrence degrades chain-of-thought monitorability — flagged as a regression by the UK AI Safety Institute and the core criticism in TechCrunch's launch coverage
- ✗Not a clean sweep: it trails Anthropic's Claude Fable 5.1 on Humanity's Last Exam with tools, 57.2% against 65.0%
- ✗Safety gating around cyber capability can interrupt legitimate security work, and the strongest capabilities require Daybreak program access rather than a standard contract
Pricing
API — Standard
From $10/M input tokens
- ✓$10 per million input tokens
- ✓$50 per million output tokens
- ✓$1 per million cached input tokens
- ✓$12.50 per million cache-write tokens
- ✓Zero data retention for eligible customers
API — Fast mode
From $20/M input tokens
- ✓Roughly 2x standard processing speed
- ✓2x standard per-token price
- ✓Same context and tool-calling support
ChatGPT Business / Enterprise
Contact for pricing
- ✓Astra and Astra Pro access
- ✓Admin-gated enablement, off by default
- ✓SSO and SCIM provisioning
- ✓Data residency in 11+ regions
Metered per token at $10 in / $50 out per million, with cached input at $1 and cache writes at $12.50; web search calls bill separately at $10 per 1,000. That is roughly 2.5x GPT-5.6 Sol and about 2x Claude Opus 5, and llm-stats puts blended cost near $11.90 per million tokens at a 20:1 input-to-output ratio. Fast mode doubles the rate. ChatGPT plan pricing for Astra access is not published for Enterprise; the strongest cyber capability is gated behind the separately contracted Daybreak program rather than any listed tier.
Security & Compliance
Connect
Sources
This page was written from 8 sources, 7 on domains other than openai.com.
- 1.openai.com — gpt 6 astravendor
- 2.techcrunch.com — openai launches astra its powerful and controversial new mod
- 3.arcprize.org — astra
- 4.coderabbit.ai — gpt 6 astra code review evaluation
- 5.datacamp.com — gpt 6 astra
- 6.llm-stats.com — gpt 6 astra
- 7.openrouter.ai — gpt 6 astra
- 8.hn.algolia.com — hn.algolia.com
Stay Ahead of the Curve
Weekly enterprise AI insights for technology leaders. No spam, no vendor pitches—unsubscribe anytime.
SubscribeRelated Products
Inkling
Apache-2.0 975B-parameter open-weights model from Mira Murati's lab, built to be fine-tuned rather than rented
Fundamental NEXUS
A foundation model built for tables, not text — enterprise prediction without feature engineering
Meta Muse Glimmer
Apache 2.0 open-weight 30B model built to run always-on AI agents on a single consumer GPU
Solar Pro 4
Agentic enterprise LLM tuned for multi-step document work at a tenth of frontier-model token cost