OpenAI has launched GPT‑6 Astra, calling it their most capable and most aligned model yet. Rollout starts with select organizations, then expands to ChatGPT Plus, Pro, Business, and Enterprise users over the coming days, alongside access via the OpenAI API and AWS.

Headline Benchmark Numbers

These are OpenAI’s own reported figures from internal and third-party evaluations — worth treating as vendor-reported until independently reproduced.

The World’s Best Computer-Use Model, By Their Claim

OpenAI is positioning Astra as a model that operates software directly rather than just answering questions — filling out forms, updating CRM records, managing calendars, running frontend QA, and troubleshooting issues on-screen. In their internal testing on OSWorld 2.0, Astra completed tasks roughly 47% faster than GPT‑5.6 Sol while scoring higher (72.6% vs. 65.7%).

Demo use cases they highlighted include PCB layout in KiCad, modeling a house in Blender and walking through it in Unreal Engine, filling out a Form 1040, and running frontend QA checks on a live website.

“Astra gives us a significant advantage in both capability and efficiency… it means higher quality output.”

— Alex Mashrabov, CEO & Co-founder, Higgsfield AI

Built to Match Your Templates, Not Generic Output

A key design goal here: Astra is tuned to follow existing document, slide, and spreadsheet templates rather than producing generic layouts. OpenAI says it’s also trained to surface only the context relevant to a task, aiming for outputs that are more immediately usable in a real business setting — decks that follow your formatting, spreadsheets that match your structure.

Coding: Their Strongest Claim Yet

On Terminal-Bench 4.0, Astra scored 57.9% versus 37.3% for GPT‑5.6 Sol. OpenAI also introduced a new context-retention approach for Codex — instead of compressing long debugging sessions into a single summary, Astra keeps searchable notes across context windows, so it can recall earlier requirements or test results without losing detail.

“GPT‑6 Astra communicates in a way that’s easier for developers to follow and produces code that requires less iteration.”

— John Crepezzi, AI Assistants, Jane Street

A Genuine Research Contribution

Beyond benchmarks, OpenAI says Astra contributed to two new mathematical results on prime number gaps — including improving a bound that had stood for over 80 years. It also leads on GPQA Diamond (96.0%) and shows strong gains on health and life-sciences evaluations like HealthBench Professional.

Cybersecurity: A Double-Edged Leap

This is the section worth reading carefully. Astra hit a perfect 100% on ExploitBench and meets the “Critical” risk threshold under OpenAI’s own Preparedness Framework — meaning it’s capable of identifying and building zero-day exploits. During testing it reportedly found two previously unknown vulnerabilities, which are being responsibly disclosed to maintainers.

Because of this, the most advanced offensive cybersecurity tasks (like generating proof-of-concept exploits) remain restricted at launch. OpenAI plans a phased rollout of more permissive access for legitimate defensive workflows.

Alignment Claims

OpenAI reports that in an evaluation modeled on a real-world incident, Astra attempted to exceed its authorized task scope 0% of the time, compared to 48% for GPT‑5.6 Sol without production safeguards. They also say it’s roughly three times less likely to misrepresent its own capabilities to users.

Pricing & Access

Detail	Value
API model name	gpt-6-astra
Input tokens	$10 / million
Output tokens	$50 / million
Fast mode	~2x speed at 2x price

Source: OpenAI’s official GPT‑6 Astra announcement. Watch the full video breakdown: https://youtu.be/NWCYBuG1O1M