GPT-6 Astra: A new generation of intelligence

What changed

Source summary

OpenAI released GPT-6 Astra, calling it its most intelligent and aligned model yet, with major gains in computer use, software engineering, mathematics and scientific research (98% on FrontierMath Tier 4, 99.9% on ARC-AGI-3). It is rolling out to ChatGPT Plus, Pro, Business and Enterprise users over the coming days, plus via the API and cloud partners. OpenAI says Astra's cybersecurity capability is strong enough to cross its "Critical" risk threshold, so it ships with extra alignment training and monitoring, safeguards OpenAI admits can sometimes slow or block legitimate work and which do not fully solve the harder problem of monitoring the model's own reasoning.

Why it matters to your work

Practical interpretation

This is OpenAI's newest and most capable ChatGPT model, rolling out to Plus, Pro, Business and Enterprise plans. If you use ChatGPT for coding, research or reports, expect noticeably better results without doing anything differently. Because this model is also powerful enough to raise cybersecurity risk, OpenAI has added extra safety checks that can occasionally slow down or block a request — that is a safety guardrail working as intended, not a bug.

When this could help

Worth evaluating for your hardest reasoning and agentic tasks — financial modelling, multi-step browser/computer control, cybersecurity analysis, scientific research — where no cheaper model meets the accuracy or safety bar.

GPT-6 Astra is OpenAI's original flagship launch: SOTA across computer use, browsing, software engineering, cybersecurity, and science. Benchmarks include FrontierMath T4 98%, ARC-AGI-3 99.9%, ExploitBench 100%, Terminal-Bench 57.9%. It achieved 0% scope-overreach on computer use (vs 48% for Sol). API pricing: $10/M input, $50/M output.

How to get started

1. Select GPT-6 Astra in ChatGPT (available on Plus/Pro/Business/Enterprise), or use model ID gpt-6-astra in the API. Also accessible via Azure OpenAI and Amazon Bedrock.

   Provider guidance ↗

Build it into your routine

SCALA suggested workflow

1. Flagship model for hardest tasks

   Reserve Astra for tasks where accuracy, safety, and multi-step reasoning are non-negotiable. Use Sol or Luna for routine work to manage costs.

What to check

• At $10/M input and $50/M output, Astra is significantly more expensive than Sol ($2/$10) and Luna. Reserve for tasks that justify the cost. Source ↗
• OpenAI identifies Astra as reaching the Critical cybersecurity threshold — access controls and monitoring are important for cybersecurity use cases. Source ↗