OpenAI has officially introduced GPT-6 Astra and they just declared that it the world’s most intelligent, capable, and aligned artificial intelligence model to date.
This is succeeding the GPT-5.6 Sol and Astra marks a massive architectural leap forward across pre-training, reinforcement learning, autonomous computer use, coding, mathematics, and cybersecurity.
Saturating FrontierMath and ARC-AGI-3 Benchmarks
GPT-6 Astra sets records across academic and reasoning evaluations, nearly saturating tests previously deemed impossible for AI.
On FrontierMath Tier 4, an evaluation of unsolved research-grade mathematical problems, Astra scored an astonishing 97.6%, leapfrogging GPT-5.6 Sol’s 83.0%.
On the visual and abstract reasoning benchmark ARC-AGI-3, the model achieved 99.9% accuracy, demonstrating true spatial generalization.
In scientific reasoning, Astra scored 96.0% on GPQA Diamond, outperforming human PhD experts across quantum physics, organic chemistry, and biology.
Autonomous Computer Use at 1.9x Speed and ChatGPT Sites
A major focus of Astra is native computer use. Astra executes GUI and browser navigation 1.9x faster than previous agentic systems, minimizing unnecessary clicks and correcting navigational errors on the fly.
Working alongside an updated Codex harness, the model can independently manage multi-step computer tasks, research complex itineraries, book appointments, and browse the web with human-like judgment.
In professional environments, Astra excels at generating production-ready spreadsheets, polished presentation slide decks adhering to corporate templates, and 3D CAD modeling with a 95.9% geometric-overlap score on BenchCAD.
With the new Sites feature in ChatGPT, Astra can also write, build, host, and deploy interactive websites, web applications, and playable 3D games directly from a single natural language prompt.
State-of-the-Art Software Engineering and Persistent Context Memory
For devs, GPT-6 Astra comes with a new standard in software engineering, hitting 57.7% on Terminal-Bench 4.0 and 74.1% on DeepSWE v1.1. The model produces production-grade code requiring far fewer debugging cycles, communicating clearly during multi-turn refactoring sessions.
OpenAI is also introducing an experimental persistent memory architecture for Codex that eliminates destructive context compaction.
Pricing and Global Availability
GPT-6 Astra begins rolling out today to select organizations and will become available across ChatGPT Plus, Pro, Business, and Enterprise tiers over the coming days, included within existing subscription allowances.
In the OpenAI API, it is available under the identifier gpt-6-astra priced at $10 per million input tokens and $50 per million output tokens, alongside an accelerated Fast Mode delivering up to 2.5x speed.
It is also launching concurrently on Amazon Bedrock for enterprise cloud developers.
OpenAI GPT-6 Astra Key Benchmark Performance Summary
| Evaluation Benchmark / Domain | GPT-6 Astra | GPT-5.6 Sol | Claude Fable 5.1 |
|---|---|---|---|
| FrontierMath Tier 4 (Mathematics) | 97.6% | 83.0% | 87.8% |
| ARC-AGI-3 (Abstract Reasoning) | 99.9% | ~85% | ~88% |
| GPQA Diamond (Graduate Science) | 96.0% | 94.6% | 93.7% |
| ExploitBench (Cybersecurity) | 100.0% | 78.5% | — |
| Terminal-Bench 4.0 (Coding & CLI) | 57.7% | 37.3% | 55.8% |
| DeepSWE v1.1 (Software Engineering) | 74.1% | 72.7% | 67.4% |
| BenchCAD (3D CAD Reconstruction) | 95.9% | 83.3% | 84.3% |
| Computer Use Speed | 1.9x Faster | Baseline | — |
| API Pricing (Input / Output per 1M) | $10 / $50 | — | — |
Leave a Reply