Overview
GPT-6 Astra is OpenAI's next-generation flagship model **officially released on September 3, 2026 (PT)**, framed by the company as "A New Generation of Intelligence" and billed as "the world's most intelligent and aligned model." Astra means "stars" in Latin, and the pre-launch teaser read "The stars are almost aligned." At the media briefing, president Greg Brockman closed with "Welcome to the AGI era," saying that in hindsight Astra may well be seen as the moment the AGI era began — the first time OpenAI directly tied one of its releases to the AGI narrative, and the most talked-about aspect of the launch.
**The defining shift of this generation is moving from answering questions to completing work.** Astra is not limited to conversation and generation: given a high-level goal ("turn this financial analysis into a deck" or "build a game prototype and test whether it runs"), it decomposes the task, calls tools, and keeps executing based on on-screen feedback until delivery. It operates computers directly — filling web forms, updating CRM records, managing calendars, doing online research, opening Python notebooks to analyze scientific data, building websites and running front-end QA, installing software and debugging errors shown on screen. On OSWorld 2.0, Astra scores **72.6%** (GPT-5.6 Sol: 65.7%), and in latency simulations it averaged about **40 minutes per task versus roughly 75 minutes for Sol — nearly 47% faster**. This is the first time an AI computer operator has approached the productivity threshold where handing it the work beats doing it yourself.
**Reasoning shows what looks like a near-saturation jump.** On ARC-AGI-3 (abstract reasoning and learning in unfamiliar environments, widely treated as a core proxy for general intelligence), Astra scores **99.9%** versus 7.8% for GPT-5.6 Sol — a more than tenfold improvement in a single generation. On research-level mathematics (FrontierMath Tier 4 v2), Astra reaches **97.6%** (Sol: 83.0%; Claude Fable 5.1: 87.8%), and OpenAI reports Astra contributed two new theoretical results on prime gaps — improving the bound on short prime gaps from 240 to 186, and improving a long-standing bound that had stood for over 80 years. In coding, Astra scores 57.9% on Terminal-Bench 4.0 (Sol: 37.3%; Fable 5.1: 55.8%), 64.6% on the scientific-workflow benchmark Terminal-Bench Science 0.1 (Sol: 22.4%), and solves 99.2% of SRE-Bench reverse-engineering tasks within four attempts.
**Cybersecurity is the most distinctive card of this generation.** Astra is the first OpenAI model to reach the **Critical** capability threshold under the company's Preparedness Framework — meaning that, with appropriate tools and access, it can autonomously discover previously unknown security flaws and build exploit chains. It scores a perfect **100% on ExploitBench** (Sol: 78.5%); in an internal test built on real vulnerabilities disclosed between June and August 2026, its arbitrary code execution rate far exceeded Sol's, and during evaluation it autonomously discovered and used **two previously unknown zero-day vulnerabilities** (both disclosed to maintainers). OpenAI delayed parts of the release by several weeks to strengthen safeguards, restricted the most advanced offensive capabilities to the Daybreak program, and is expanding access for defensive use only. Alignment numbers are equally striking: 0% attempts to exceed authorized goals (Sol: 48%), a 2.4% misuse rate in an internal computer-use safety benchmark (Sol: 22.0%), and a 4.2% hallucination rate (Sol: 12.2%).
**Availability and pricing**: Astra first opened on 9/3 to a limited set of organizations in the Daybreak cybersecurity program, with rollout to all ChatGPT Plus / Pro / Business / Enterprise subscribers in the coming days, plus access via the OpenAI API (model id gpt-6-astra) and Amazon Bedrock; Pro and Enterprise tiers also get GPT-6 Astra Pro. Standard API pricing is **$10 per million input tokens and $50 per million output tokens** — 2.5x GPT-5.6 Sol's current price ($4/$20) and roughly on par with Claude Fable 5.1 ($10/$50) — with cache discounts, a Fast mode at up to 2.5x speed for 2x price, and five reasoning-effort levels. Independent data from Artificial Analysis adds nuance: Astra's intelligence index (61.2) is only slightly above Sol (60.9) while producing about 10% fewer output tokens, so simple single-task costs may run about 75% higher than Sol; yet within long-horizon Codex tasks, per-task cost lands close to or below both Sol and Claude Fable 5 — whether it feels expensive depends on how you use it.
Key Features
- First Critical cybersecurity rating: OpenAI's first model to reach the 'Critical' threshold under its Preparedness Framework: autonomously discovers unknown flaws and builds exploit chains; 100% on ExploitBench, with protections raised accordingly
- 99.9% on ARC-AGI-3: Jump from 7.8% (GPT-5.6 Sol) to 99.9% on abstract reasoning and learning in unfamiliar environments — a tenfold gain in one generation
- 97.6% on FrontierMath Tier 4: Research-level mathematics nearly saturated; OpenAI reports two new theoretical results on prime gaps (bound of 186; an 80+ year old bound improved)
- A real computer-use agent: 72.6% on OSWorld 2.0 and ~40 min per task (47% faster than Sol); operates forms, CRM, calendars, scientific software, KiCad, Blender and more end to end
- Major alignment and control gains: 0% unauthorized goal-seeking (Sol: 48%), 2.4% misuse rate (Sol: 22.0%), 0% on ExploitGym honeypot (Sol: 48.2%) — OpenAI's most 'obedient' model to date
- Codex long-horizon memory: Keeps working notes searchable across context windows instead of repeatedly compressing them; Mind2Web + Codex task completion at 1.9x the Sol experience
- 1.05M token context: 1.05 million total context tokens and up to 128K output tokens; knowledge cutoff 2026/4/30; streaming, structured outputs, function calling, file search and web browsing
- Professional software workflows: 95.9% geometric overlap on BenchCAD (Sol: 83.3%); handles PCB layout, Blender-to-Unreal 3D modeling, financial modeling, legal documents and more
Use Cases
- Long-horizon agentic work: hand it a goal and let it decompose and execute (computer use, software engineering, data workflows)
- Scientific research: operating professional software for data analysis, experiment design and model fitting (64.6% on Terminal-Bench Science 0.1)
- Authorized cybersecurity defense: vulnerability research, validation and detection (Daybreak program; offensive use restricted)
- Professional design and engineering: PCB layout, 3D modeling, CAD and financial-modeling work that was hard to automate before
- Deep math and code reasoning: FrontierMath / Terminal-Bench / SRE-Bench scenarios
- Flagship upgrade for heavy ChatGPT and API users (Astra Pro for Pro/Enterprise tiers)
Pros
- The 99.9% ARC-AGI-3 and 97.6% FrontierMath Tier 4 results represent a real generational gap
- First Critical-rated cyber model — offensive and defensive capability in a league of its own
- Best-ever alignment numbers: 0% unauthorized goal-seeking, ~10x lower misuse rate, ~2/3 lower hallucination rate
- Computer use now approaches real productivity (~40 min/task vs Sol's ~75 min)
- 1.05M context plus Codex cross-window memory — long tasks no longer lose details
- Mathematics produced real results (two new prime-gap theorems), not just benchmark scores
Pricing
**Standard API**: $10 per million input tokens and $50 per million output tokens — 2.5x GPT-5.6 Sol ($4/$20) and roughly on par with Claude Fable 5.1 ($10/$50). Cached input ~$1 per million tokens (~90% discount); cache writes carry ~25% premium. **Fast mode**: up to 2.5x standard processing speed at 2x the price. **Reasoning effort**: five levels from low to max to trade depth against cost. **Subscriptions**: Astra usage is included in existing allowances, with paid credits beyond; Pro/Business/Enterprise tiers include GPT-6 Astra Pro. **Channels**: OpenAI API (gpt-6-astra) and Amazon Bedrock. **Independent note**: Artificial Analysis measures Astra's intelligence index at 61.2 (Sol: 60.9) with ~10% fewer output tokens; because unit price is 2.5x, simple single-task costs may run ~75% higher than Sol, while long-horizon Codex tasks land close to or below Sol and Claude Fable 5.
Summary
GPT-6 Astra, released by OpenAI in September 2026, is the sixth-generation flagship model that shifts the focus from conversation and generation to complete agentic work: it operates computers directly, decomposes goals and executes long-horizon tasks. It scores 99.9% on ARC-AGI-3 and 97.6% on FrontierMath Tier 4, is OpenAI's first model to reach the Critical cybersecurity threshold (100% on ExploitBench), and posts the company's best alignment and control numbers to date. API pricing is $10 per million input and $50 per million output tokens — 2.5x GPT-5.6 Sol and on par with Claude Fable 5.1. It suits developers and enterprises that need deep reasoning, long-horizon agent execution, or a high safety boundary; cost-sensitive or lightweight users may still prefer Sol/Terra/Luna.
Version History
- GPT-6 Astra official launch (2026-09-03): OpenAI unveiled GPT-6 Astra, billed as the world's most intelligent and aligned model. It is OpenAI's first model to reach the Critical cybersecurity threshold (100% on ExploitBench; two zero-days found), scores 99.9% on ARC-AGI-3 and 97.6% on FrontierMath Tier 4, and is positioned for long-horizon agentic work and computer use (72.6% on OSWorld 2.0, ~40 min per task). Rollout began 9/3 to Daybreak organizations, expanding to all ChatGPT subscribers, the API and Amazon Bedrock within days. API $10/$50, 2.5x GPT-5.6 Sol.