GPT-6 Astra: OpenAI Unveils Frontier Model with Near-Perfect ARC-AGI-3 and ExploitBench Scores

OpenAI officially launches GPT-6 Astra, featuring a 1.05M context window, 98% FrontierMath Tier 4, 99.9% on ARC-AGI 3, and 100% on ExploitBench at $10/$50 per million tokens.

MV
OpenAIVerified Lab
3 min read·Sep 3, 2026·Original Source
GPT-6 Astra: OpenAI Unveils Frontier Model with Near-Perfect ARC-AGI-3 and ExploitBench Scores
Figure 1: Official research and architecture release visual · OpenAI

On September 3, 2026, OpenAI officially announced the release of GPT-6 Astra, their most capable foundation reasoning model to date. Positioned as a generational leap in autonomous problem-solving, science, and computer use, Astra arrives following extensive safety evaluations and alignment audits necessitated by its unprecedented capabilities in autonomous vulnerability discovery and novel abstraction reasoning. Operating across a native 1.05-million-token context window with up to 128,000 output tokens, GPT-6 Astra demonstrates benchmark-shattering results: scoring 98% on FrontierMath Tier 4, up to 99.9% on ARC-AGI-3 under specialized evaluation harnesses, and a perfect 100% on ExploitBench, establishing it as the first foundation model officially classified under OpenAI's "Critical" cybersecurity capability tier.

Key Breakthroughs.

1. Near-Saturation on

ARC-AGI-3 and Novel Abstraction ARC-AGI-3 was deliberately constructed to test genuine general intelligence by placing models in novel, interactive logic environments designed to resist memorization from internet pretraining data:

The ARC-AGI-3 Leap: While contemporary frontier models struggled below 10% on ARC-AGI-3, GPT-6 Astra achieves 62.7% on the standard semi-private benchmark harness and climbs to 98.6% – 99.9% when utilizing extended test-time search and specialized provider adapters. Autonomous Rule Induction: The model actively tests hypotheses against environment mechanics, inducing abstract geometric rules and physical invariant transformations without human demonstration.

2. Extreme

Mathematical Problem Solving on FrontierMath Tier 4 FrontierMath comprises research-grade mathematics problems authored by leading mathematicians, requiring multi-page proofs and intricate theoretical synthesis:

98% on Tier 4: Astra achieves 97.6% – 98.0% accuracy on Tier 4 problems, solving doctoral-level challenges in algebraic topology, number theory, and differential geometry that previously stumped all commercial LLMs. Dynamic Extended Deliberation: Supported across low, medium, high, xhigh, and max reasoning effort modes, Astra dynamically scales internal deliberation tokens to verify proofs before outputting symbolic solutions.

3. Critical

Cybersecurity Threshold & ExploitBench GPT-6 Astra is the first model in industry history to cross the formal "Critical" threshold under frontier safety frameworks:

100% Arbitrary Code Execution on ExploitBench: On an internal evaluation corpus containing 20 high-severity V8 engine vulnerabilities, Astra achieved a perfect 100% exploit development rate using a fraction of the output tokens required by GPT-5.6 Sol. Full Zero-Day Sandbox Escapes: In expert red-teaming evaluations, Astra autonomously identified two novel zero-day vulnerabilities in hardened operating systems and browsers, chaining them together to escape virtualization sandboxes and escalate local privileges to root. Hardened Defense & Alignment: To mitigate misuse, OpenAI implemented layered activation classifiers and cross-conversation context monitoring, achieving a 91.5% refusal rate on adversarial cyber jailbreak prompts.

4. Advanced

Desktop Agency on OSWorld 2.0 In long-horizon computer use and multimodal desktop automation:

72.6% on OSWorld 2.0: Astra establishes a new industry benchmark on full Ubuntu desktop environments, navigating arbitrary software suites, terminals, web browsers, and file systems across complex multi-step workflows without human intervention.

Technical Specifications & Benchmark Overview

& Benchmark Overview Metric / Dimension Specification Developing Lab OpenAI Release Date September 3, 2026 API Model Identifier gpt-6-astra (rolling out in ChatGPT as GPT-6 Pro) Context Window 1,048,576 (1.05M) input tokens; 128K max output tokens Reasoning Effort Levels low, medium, high, xhigh, max Key Benchmarks ARC-AGI-3: 98.6%–99.9% (62.7% std) | FrontierMath Tier 4: 98.0% | ExploitBench: 100% | OSWorld 2.0: 72.6% Safety Classification Critical Threshold (Cybersecurity & Autonomous Replication Audited) Pricing $10.00 / 1M Input Tokens, $50.00 / 1M Output Tokens (Standard Tier) Availability Trusted Access Program, ChatGPT Pro ($100/$200), OpenAI API

Verified Integration & API Usage

Developers with access can invoke gpt-6-astra using the standard OpenAI client:

python
import os

from openai import OpenAI

client = OpenAI(

    api_key=os.environ.get("OPENAI_API_KEY"),

)

response = client.chat.completions.create(

    model="gpt-6-astra",

    messages=[

        {

            "role": "system",

            "content": "You are an expert AI research scientist specializing in abstract reasoning.",

        },

        {

            "role": "user",

            "content": "Formalize the topological proof for the invariant manifold transition problem.",

        },

    ],

    reasoning_effort="max",

)

print(response.choices[0].message.content)
Primary Foundation Model

This technical digest directly references verified architecture specifications documented in TheModelverse Foundation Model Catalog.

OpenAIReasoningProprietary Commercial API

GPT-6 Astra

Params: UndisclosedContext: 1049k tokensRate: $10/1M
View Full Specs & Benchmarks
Advertisement