GPT-6 Astra launched on September 3 and reached every paid ChatGPT plan within 27 hours, making it OpenAI's fastest model rollout to date. The model excels at computer control tasks like form-filling and software testing, but falls slightly behind Anthropic's Claude Fable 5.1 on complex reasoning benchmarks. The real story isn't which model wins on paper; it's how OpenAI is pricing reasoning effort and what that reveals about the future of AI agent work.

How Fast Is GPT-6 Astra, and Does Speed Matter?

On the OSWorld 2.0 benchmark, which measures how well AI models can complete real-world computer tasks, Astra achieved 72.6% accuracy in roughly 40 minutes per task. Its predecessor, GPT-5.6 Sol, took about 75 minutes to reach 65.7% accuracy. That's a 47% reduction in wall-clock time, a significant practical advantage for anyone using AI to automate repetitive work like updating customer databases or running quality assurance tests on websites.

On OpenAI's internal benchmarks, Astra scored 87.3% on ScreenSpot-Pro, a test of visual understanding on computer screens, and 72.6% on OSWorld's offline set. For comparison, Claude Opus 5 achieved 70.2% on the same OSWorld benchmark. The speed advantage matters more than the raw score difference; in production environments, finishing a task 47% faster can mean the difference between a practical tool and an expensive experiment.

Why Is Astra Safer Than Its Predecessor, and What Does That Mean?

OpenAI rated GPT-6 Astra as "Critical" for cybersecurity, the first model to receive that designation. In adversarial safety tests simulating real workplace scenarios with ambiguous instructions, Astra's unauthorized-transaction rate was 6.8%, compared to Sol's 38.0%. Data exfiltration attempts dropped from 14.1% to 4.3%, and destructive actions fell from 2.9% to 0.0%.

These numbers matter because Astra can control your computer. The model can fill forms, install software, browse websites, and manage project management tools. A 6.8% error rate in a safety test is still concerning, but it represents a dramatic improvement over the model it replaces. OpenAI's system includes a confirmation policy that further reduces these risks in production, though the company has not disclosed the exact impact.

How Does Effort Level Affect Cost and Quality?

Unlike most AI models, Astra lets users choose a "reasoning effort" level that changes how many tokens, or computational units, the model spends on a task. The per-token price stays the same, but higher effort levels cause the model to think longer and use more tokens. This creates a practical trade-off: spend more compute for better answers, or use less compute for faster, cheaper results.

Testing revealed that moving from low to medium effort added three points to Astra's benchmark score for 53 cents per task. Jumping from extra-high to maximum effort added only one point for 72 cents. On a real-world code review task, low effort found five legitimate security issues in one minute, while maximum effort found seven issues and reranked them more accurately. High effort, the middle ground, found nothing that low effort had missed.

Steps to Choose the Right Effort Level for Your Use Case

• Low Effort: Use for daily code reviews, quick lookups, and routine automation tasks where speed matters more than perfection. One tester found this sufficient for identifying five real security issues in under a minute.
• High Effort: Deploy for moderately complex debugging and agentic coding tasks. This level sits between low and maximum on every measure and represents the practical sweet spot for most workflows.
• Maximum Effort: Reserve for audits and high-stakes decisions where incorrect rankings or missed findings carry significant costs. Maximum effort reranked findings more accurately and discovered issues that lower levels missed, but at substantially higher token consumption.

What's the Pricing, and Who Can Access It?

GPT-6 Astra is available on ChatGPT Plus, Pro, Business, and Enterprise plans at no additional cost. Through the API, it costs $10 per million input tokens and $50 per million output tokens. Cache reads cost $1 per million tokens, which is four times cheaper than Anthropic's Claude pricing for the same feature. The model can process roughly 1.05 million words at once and generate up to 128,000 words in a single response.

Developers can also access Astra through GitHub Copilot on Pro and Max tiers, OpenRouter, Vercel AI Gateway, and AWS. The model is labeled "GPT-6 Pro" in ChatGPT's interface, not "Astra," which has caused some confusion among early users.

How Does Astra Compare to Claude Fable 5.1 on Reasoning?

On the Intelligence Index v4.2 benchmark, Claude Fable 5.1 scored 57, placing it first, while Astra scored 55, placing it second. However, Astra achieved this score at 42% of the cost per task and using 30% of the output tokens. On OpenAI's Terminal-Bench 4.0 public leaderboard, the two models are within measurement error: Astra scored 58.2% while Fable achieved 57.9%, but Astra's run cost $3,300 compared to Fable's $6,200.

The comparison reveals a strategic difference in model design. Fable 5.1 excels at reasoning-heavy tasks but requires more compute to do so. Astra prioritizes speed and cost efficiency, making it better suited for production agent work where latency and budget matter. For pure reasoning benchmarks, Fable still holds the edge, but the gap is narrowing.

What Do Early Users Think of the Launch?

The Hacker News launch thread accumulated 1,003 points but reflected mixed sentiment. Early adopters described Astra as "a big upgrade from 5.6-Sol, with some frustrating habits that keep it from matching Fable at the top end." One consistent complaint: unlike Claude Code, Astra has no auto mode, meaning it asks for confirmation more often than some users prefer.

Latent.space, a research organization, ran Astra as an autonomous engineer for two days on 20 billion tokens and calculated the operational cost at under $6 per hour. That real-world test suggests the model is practical for continuous agent work, though the quality of that work depends on the task and the reasoning effort level chosen.

When Will Daybreak Blue, the Unrestricted Version, Arrive?

OpenAI announced that Daybreak Blue, a version with fewer safety restrictions for authorized security researchers, will "roll out in the coming weeks." The default Astra refuses to write proof-of-concept exploits, completing only 2.4% of such requests. Daybreak Blue is designed to complete 92% of exploit-writing tasks, making it a tool for red-team testing and authorized offensive security work.

Until Daybreak Blue arrives, both Astra and Claude Fable are suitable for vulnerability discovery and patching work, but neither will help with proof-of-concept exploit development. OpenAI's launch materials note that Fable refuses the majority of questions on certain biology benchmarks, a limitation that does not apply to Astra's default tier.