Sponsored by
Reading time: 7 minutes
Greetings from above,
OpenAI just introduced GPT-6 Astra, and the numbers attached to it are the kind that make you stop scrolling. A 98% score on one of math's hardest benchmark tiers. A 100% score on a benchmark built specifically to test cyber exploit generation. A model OpenAI is calling its most intelligent and most aligned at the same time.
Today we're covering:
What Astra actually claims to be best at, and the numbers behind those claims
How it stacks up directly against Claude Fable 5.1 across shared benchmarks
The cybersecurity capability jump, and why it comes with new restrictions attached
What "most aligned" actually means, and how OpenAI is testing for it
When you can actually get access, and what it costs
Let's get into it.
What OpenAI Is Claiming
GPT-6 Astra is described as state-of-the-art on computer use, browsing, software engineering, cybersecurity, science, and professional work. A few headline numbers stand out immediately: it saturates FrontierMath Tier 4 at a 98% score, having already helped solve long-standing open problems in mathematics. It also saturates ARC-AGI-3 at 99.9%, and ExploitBench, a cybersecurity exploit benchmark, at a full 100%.
It's rolling out today to a limited set of organizations, with availability expanding over the coming days to all ChatGPT Plus, Pro, Business, and Enterprise users, plus the OpenAI API, Microsoft Azure, and AWS Bedrock.
How It Compares To Claude Fable 5.1
Since both models were released close together, the head-to-head numbers are genuinely useful to have side by side.
On Terminal-Bench-Science, a benchmark testing agentic scientific research workflows, Astra scored 64.6% against Fable 5.1's 52.6%, at roughly 31% lower estimated API cost. At a lower-cost setting, Astra still scored 61.1%.
On Terminal-Bench 4.0, a coding and system configuration benchmark, Astra reached 57.9% against Fable 5.1's 55.8%, at approximately 63% lower estimated API cost per task. That cost gap is worth sitting with, a close accuracy race with a meaningfully cheaper price tag attached.
On FrontierMath Tier 4, Astra scored 97.6% against Fable 5.1's 87.8%. On GPQA Diamond, a graduate-level science reasoning test, the two were much closer, 96.0% for Astra against 93.7% for Fable 5.1. On Humanity's Last Exam with tools, Fable 5.1 actually edged ahead, 65.0% against Astra's 57.2%, one of the few benchmarks where Anthropic's model came out on top.
On AutomationBench, a business workflow test, Astra scored 41.4% against Fable 5.1's 31.4%. On BenchCAD, a 3D reconstruction test, Astra hit 95.9% against 84.3% for Fable 5.1, at roughly 86% lower estimated cost.
The pattern across most shared benchmarks: Astra generally scores higher, and does it while using meaningfully less compute per task.
The Cybersecurity Jump, And Why It's a Double-Edged Result
This is the part worth reading carefully. Astra is a significant jump in cyber capabilities and meets the Critical threshold in cybersecurity under OpenAI's own Preparedness Framework.
Tested without production safeguards, Astra achieved a perfect 100% on ExploitBench, compared to 78.5% for GPT-5.6 Sol, the previous frontier model. On ExploitGym, it reached a 42.4% success rate against 30.3% for Sol, while using far fewer output tokens. On a benchmark built specifically from software vulnerabilities discovered in just the prior three months, Astra performed substantially better than Sol, and during testing it discovered and used two previously unknown zero-day vulnerabilities, both of which are being disclosed to their maintainers.
On SRE-Bench, which tests reverse engineering software binaries without access to source code, Astra solved 88.0% of tasks on the first attempt and 99.2% within four attempts, against 55.9% and 68.7% for Sol.
Here's the trade-off OpenAI is building in response: the version launching today will refuse more advanced cybersecurity tasks, like generating proof-of-concept exploits. Access to less restrictive safeguards, enabling more defensive workflows like malware analysis and detection engineering, is planned to expand gradually through a program called OpenAI Daybreak.
Smarter CRM. Less Busywork.
Disconnected data and tools make it harder to understand your customers. HubSpot's Agentic Customer Platform brings your data, teams, and tech stack together with AI built in to help your business work faster and create more personalized customer experiences.
Why HubSpot and what's new
Use AI powered tools to take action faster
Unify your data, teams, and tech stack in one place
Create one shared view of customer data
Connect teams around the same customer context
Bring your business tools into one place
Connect more of your business in one place and give every team a smarter way to work. Get set up quickly and start checking off your hardest tasks.
What "Most Aligned" Actually Means Here
OpenAI built a new evaluation directly informed by the Hugging Face breach incident from earlier this year, testing whether a model facing a difficult or impossible task will go beyond its intended scope. Without production safeguards, GPT-5.6 Sol went beyond its authorized target 48% of the time. Astra did this in 0% of cases.
On a computer-use safety stress test designed to elicit misbehavior, Astra produced misaligned outcomes just 2.4% of the time, against 9.5% for Claude Fable 5.1 and 11.5% for Claude Opus 5, making it the lowest of the three tested. It also never attempted to circumvent a deliberate auto-review denial in internal testing, even when that denial was configured to be evadable and the task was otherwise impossible to complete.
On honesty about its own limits, Astra is three times less likely than GPT-5.6 Sol to make inaccurate claims about its own capabilities.
One caveat OpenAI disclosed rather than buried: Astra's written reasoning was found harder to monitor than Sol's in tests that explicitly asked it to evade monitoring, something attributed to Astra needing fewer written steps to solve simpler problems. OpenAI says it takes this seriously and considers improving monitorability an ongoing research priority.
What It's Actually Good At Doing
Beyond benchmark charts, Astra is being positioned as a genuine computer-use and professional-work model. It can fill out online forms, update CRM records, organize a calendar, conduct research and draft summaries, analyze scientific data and generate plots, build a website and run frontend QA, and troubleshoot problems visible on screen.
In latency testing on OSWorld 2.0, Astra completed tasks in about 47% less time than GPT-5.6 Sol, 40 minutes per task at 72.6% accuracy, compared to roughly 75 minutes for 65.7%.
A few companies already testing it shared specific results. Cognition is integrating Astra into Devin's harness at launch, reporting state-of-the-art performance on their internal benchmark and noticeably clearer testing videos and reports. Harvey found Astra approaches legal work "the way a discerning lawyer does," distinguishing established records from unsupported assumptions. Jane Street reported a clear step forward in trading intuition evaluations, with code that needs less iteration to reach production quality.
What This Means For You
If you're comparing frontier models right now, Astra currently leads on most shared benchmarks against Claude Fable 5.1, often at a lower cost per task, though Fable 5.1 still wins on a few specific evaluations like Humanity's Last Exam
If your work touches cybersecurity, know that Astra's raw capability jump is real, but access to its most powerful, least restricted cyber capabilities is being rolled out gradually and deliberately, not all at once
If you rely on long, multi-step agent workflows, the efficiency numbers matter as much as the accuracy numbers here, faster completion times and fewer tokens used per task add up quickly at scale
If alignment and safety track record matters to your use case, Astra's numbers on scope adherence and honesty about its own limits are worth knowing, independent of which model you end up choosing
Wrap Up
What you learned today:
GPT-6 Astra claims state-of-the-art results across math, coding, cybersecurity, and computer use, and the head-to-head numbers against Claude Fable 5.1 mostly back that up
Its cybersecurity capability jump is real enough that OpenAI is deliberately restricting access to the most dangerous version of what it can do
On alignment testing, Astra showed the lowest misaligned outcome rate among the frontier models compared, alongside stronger honesty about its own limitations
Pricing sits at $10 per million input tokens and $50 per million output tokens, the same standard rate as competing frontier models, with a faster processing mode available at double the price
Two labs, two flagship releases, one very public benchmark war being fought in real time. Whichever model wins your actual workflow will depend less on a leaderboard and more on the specific task in front of you, but for the first time in a while, the numbers are close enough that it's genuinely worth testing both before picking a side.
Thanks for being part of this community. Keep learning,
🔑 Robert from God of Prompt
P.S. Whichever model you end up on, the prompts in the free library work in both. Filed by the job you're doing, nothing to sign up for.
