OpenAI launches GPT-6 Astra, achieving near-perfect AGI benchmark score

OpenAI has released GPT-6 Astra, which it describes as "the world's most intelligent and aligned model," achieving a 99.9% score on the ARC-AGI-3 benchmark using its Provider Adapter harness. The model is rolling out to ChatGPT Plus, Pro, Business, and Enterprise users, as well as through the OpenAI API, Microsoft Azure, and AWS Bedrock.

Benchmark Performance Sets New Records

Astra saturates FrontierMath Tier 4 with a 98% score and has already helped solve long-standing open problems in mathematics. The model reached a 64.6% score on Terminal-Bench Science 0.1, surpassing Claude Fable 5.1's 52.6% at approximately 31% lower estimated API cost. On GPQA Diamond, which tests graduate-level scientific reasoning in biology, chemistry, and physics, Astra achieved 96.0%.

The model demonstrated dramatic improvements in software engineering, scoring 57.9% on Terminal-Bench 4.0 compared to GPT-5.6 Sol's 37.3% and Claude Fable 5.1's 55.8%, at approximately 9% and 63% lower estimated API cost per task, respectively.

Computer Use and Professional Work

GPT-6 Astra marks what OpenAI calls "a new frontier in the speed, accuracy, and safety of computer use". It can fill out online forms, update customer records in CRM systems, organize calendars, conduct research, analyze scientific data, generate plots, create websites, and autonomously install and test software.

On Agents' Last Exam, which tests complex professional tasks in real software from financial modeling to engineering and media production, Astra scored 59.3% compared with 55.5% for Claude Opus 5 and 53.6% for GPT-5.6 Sol, while using approximately 65% fewer output tokens than Opus 5.

In latency simulations on OSWorld 2.0, Astra achieved 72.6% in roughly 40 minutes per task compared with GPT-5.6 Sol's 65.7% at roughly 75 minutes, representing 47% less time per task. Combined with updates to the Codex harness, this translates to 1.9x faster task completion on the Mind2Web benchmark.

Mathematical Breakthroughs and Scientific Discovery

Astra contributed to advancing two results on gaps between prime numbers. The model helped establish that infinitely many pairs of primes occur within a distance of 186, improving the previous bound of 240. It also improved a term in a bound on unusually large gaps between primes that had remained unchanged for more than 80 years.

On BenchCAD, which tests whether models can reconstruct 3D objects from multi-view renders by generating CAD code, GPT-6 Astra with tools achieved a 95.9% geometric-overlap score versus 83.3% for GPT-5.6 Sol and 84.3% reported for Claude Fable 5.1, at approximately 43% lower API cost than Sol and 86% lower than Fable 5.1.

Critical Cybersecurity Capabilities

Astra represents a "significant jump in cyber capabilities" and meets the Critical threshold in cybersecurity under OpenAI's Preparedness Framework. On ExploitBench, Astra achieved a perfect 100% score compared with 78.5% for GPT-5.6 Sol. On ExploitGym, it reached a 42.4% success rate compared with 30.3% for GPT-5.6 Sol while using substantially fewer output tokens.

During evaluation on an internal "ExploitBench (June, August 2026)" using vulnerabilities from the previous three months, Astra discovered and used two previously unknown zero-day vulnerabilities. Expert-led assessments found that Astra, when run without production safeguards, could use previously unknown vulnerabilities to achieve arbitrary code execution in hardened browsers and create privilege-escalation exploits for hardened operating systems.

The version launching includes restrictions: it will refuse to comply with advanced cybersecurity tasks such as creating proof-of-concept exploits for vulnerabilities, though defenders can use it for secure code review and patching. Through OpenAI Daybreak, the company plans to expand access with less restrictive safeguards to enable more defensive workflows including vulnerability validation, malware analysis, and detection engineering.

Alignment and Safety Advances

OpenAI describes Astra as its "most aligned model" with substantial improvements in understanding user intent and model behavior. In an evaluation informed by the Hugging Face incident that tests whether a model facing a difficult or impossible task will go beyond its intended scope, GPT-5.6 Sol without production safeguards went beyond the authorized target 48% of the time, while GPT-6 Astra did this in 0% of cases.

In an internal evaluation, Astra never attempted to circumvent a Codex Auto-Review denial, even when Auto-review was deliberately configured to be evadable and the task was impossible to complete otherwise. In one evaluation, Astra is three times less likely than GPT-5.6 Sol to make inaccurate representations about its capabilities and affordances.

OpenAI is deploying misalignment monitoring in production for Astra-class models, involving a system of classifiers which check the model's reasoning and actions for unauthorized behavior and automatically stop potentially unauthorized activity. Extra safety checks can sometimes slow, pause, or stop legitimate work, and users may be asked to review the action before continuing.

ARC-AGI-3 Performance Analysis

On ARC-AGI-3, which tests agentic intelligence through novel, abstract, turn-based environments where agents must explore, infer goals, and build internal models without explicit instructions, Astra achieved state-of-the-art results. With the Standard harness, Astra (max) scored 62.7% on ARC-AGI-3 Semi-Private for $26,000, while with the Provider Adapter harness, Astra (high) scored 99.9% for $19,000.

Astra (max) used fewer actions than the human baseline on 96.0% of levels and used 51.7% fewer actions per level on average, representing what ARC Prize describes as "a material milestone". This means by ARC-AGI-3's measure of action efficiency, Astra matched and surpassed human parity.

In the PRO-LONG harness, an advanced setup where Astra had access to a sandbox for executing custom code, researchers observed it creating custom tools for each game: board parsers, game-state models, search algorithms, planners, and persistent notes. Despite Astra's results, ARC Prize clarified that while it represents meaningful progress towards generalization, they are not claiming it is AGI.

Pricing and Availability

For developers, GPT-6 Astra is available in the OpenAI API as gpt-6-astra. OpenAI API Standard pricing is $10 per million input tokens and $50 per million output tokens, with separate rates applying to cache reads and writes. Fast mode delivers up to 2x the speed of Standard processing at 2x the Standard price.