OpenAI just dropped GPT-6 Astra, and if the benchmarks are real, we're looking at the first model that genuinely saturates the tests we built to measure AGI. Not "approaches." Not "shows strong performance." Saturates. 99.9% on ARC-AGI-3. 98% on FrontierMath Tier 4. 100% on ExploitBench.
This isn't incremental. This is OpenAI saying the benchmark era is over and the deployment era has begun.
The Benchmark Massacre
Let's start with the numbers, because they're genuinely absurd. Astra scores 99.9% on ARC-AGI-3, the benchmark François Chollet designed specifically to resist memorization and test genuine reasoning. It hits 98% on FrontierMath Tier 4 and has already helped solve long-standing open problems in mathematics, according to OpenAI.
On GPQA Diamond—graduate-level scientific reasoning across biology, chemistry, and physics—Astra reaches 96.0%. For context, GPT-5.6 Sol topped out at 94.6%. That's not a huge jump, but the cost efficiency is: a lower-cost Astra configuration beats Sol's best score at roughly 37% lower estimated API cost.
The cybersecurity results are where things get spicy. Astra achieves a perfect 100% on ExploitBench, up from 78.5% for Sol. On ExploitGym, it hits 42.4% versus Sol's 30.3%. OpenAI is transparent that this meets their "Critical threshold" under the Preparedness Framework—meaning Astra can identify and develop zero-day exploits at a level that requires serious safeguards.
These aren't just better scores. These are "we need new tests" scores.
Computer Use That Actually Ships
The real story isn't the benchmarks—it's that Astra is the first model where computer use feels like a product, not a demo. OpenAI claims it's state-of-the-art on computer use, browsing, and software engineering, and the evaluation results back that up.
On Agents' Last Exam, which tests complex professional tasks in real software (financial modeling, engineering, media production), Astra scores 59.3% versus 55.5% for Claude Opus 5 and 53.6% for GPT-5.6 Sol. Crucially, it does this while using approximately 65% fewer output tokens than Opus 5.
The efficiency gains matter more than the raw scores. In latency simulations on OSWorld 2.0, Astra achieves 72.6% accuracy in roughly 40 minutes per task, compared to Sol's 65.7% at roughly 75 minutes. That's 47% less time per task.
OpenAI also claims that combined with an updated Codex harness, Astra delivers 1.9x faster task completion on Mind2Web. The demo videos show it doing PCB layout in KiCad, modeling houses in Blender and turning them into walkable Unreal Engine scenes, and creating playable games with vivid graphics.
This is the kind of computer use that actually matters: not just clicking buttons, but handling specialized professional software with context and judgment.
Alignment By Design, Not By Prayer
Here's the part that made me sit up: OpenAI built a new evaluation based on the Hugging Face incident to test whether a model facing a difficult or impossible task will exceed its authorized scope. GPT-5.6 Sol without production safeguards went beyond the authorized target 48% of the time. Astra? 0%.
Zero percent.
That's not "we prompted it really well." That's architectural. OpenAI calls Astra "our most aligned model" with "substantial improvements in understanding user intent and model behavior." The model uses context to fill in routine gaps, asks focused questions when the answer could change the outcome, and in Codex can ask asynchronously while continuing work that doesn't depend on a reply.
The examples show Astra incorporating new requirements and answering side questions without dropping the broader task—something earlier models consistently fumbled. This is what delegating to an agent actually requires: not just capability, but judgment about when to proceed and when to wait.
Professional Work, Actually
Astra pairs computer use with targeted training for professional environments. On BenchCAD, which tests whether models can reconstruct 3D objects from multi-view renders by generating CAD code, Astra with tools reaches 95.9% geometric-overlap score versus 83.3% for Sol and 84.3% for Claude Fable 5.1. The estimated API cost is approximately 43% lower than Sol and 86% lower than Fable 5.1.
The model creates slide decks that follow your templates, captures the correct tone and layout, and produces documents that match your business standards. It's trained to pull only the context that matters instead of repeating unnecessary information—which means outputs are more immediately usable.
OpenAI integrated this with Sites in ChatGPT, so Astra can create, host, and share websites, web apps, and games directly from a prompt. The demos show non-technical people creating custom games with engaging gameplay and accurate motion in minutes.
Coding: Notes Across Context Windows
On Terminal-Bench 4.0, which tests agents on complex terminal-based tasks including software engineering, system configuration, and data analysis, Astra reaches 57.9% compared with 37.3% for Sol and 55.8% for Claude Fable 5.1—at approximately 9% and 63% lower estimated API cost per task, respectively.
The real innovation is how Codex preserves context when the window fills. Historically, models used compaction to summarize work during long sessions. Each compaction loses details about why a fix failed or how a component behaves.
Astra can keep notes across context windows, preserving accumulated details without repeatedly compressing them. Earlier context windows remain searchable, so it can find requirements or test results from previous messages and tool outputs—even if that information wasn't captured in its notes. This is experimental now but will become the default in coming weeks.
That's a genuinely clever solution to the context-limit problem that doesn't involve waiting for 10M-token windows.
Science: Software + Reasoning
Astra combines scientific reasoning with computer use to work directly in specialized software. OpenAI shows it navigating scientific software to inspect sequencing quality and visualize genetic variation, helping researchers assess data and identify where to focus further analysis.
On Terminal-Bench Science 0.1, which tests whether agents can complete scientific research workflows using code and terminal tools, Astra reaches 64.6% versus 52.6% for Claude Fable 5.1, at approximately 31% lower estimated API cost. At a lower-cost setting, Astra scores 61.1% versus Sol's best result of 22.4%, at approximately 27% lower cost.
OpenAI also announced two new results on gaps between prime numbers, suggesting Astra is already contributing to mathematical research. That's the kind of claim that will be fascinating to watch play out.
What This Actually Means
If these results hold up in production—and OpenAI is rolling Astra out to ChatGPT Plus, Pro, Business, and Enterprise users, plus API, Azure, and AWS Bedrock—we're at an inflection point. Not because the model is perfect, but because it's good enough to ship for real work.
The combination of saturated benchmarks, efficient computer use, strong alignment, and cost improvements means Astra isn't just more capable—it's more deployable. That's the difference between a research milestone and a product.
The cybersecurity implications are real. A model that scores 100% on ExploitBench and meets the Critical threshold under the Preparedness Framework is a dual-use tool that requires serious access controls and monitoring. OpenAI has been transparent about this, but transparency doesn't eliminate risk.
The bigger question is what happens when every knowledge worker has access to a model that can autonomously handle professional software, maintain context across hours of work, and make judgment calls about when to proceed versus when to ask. That's not AGI in the sci-fi sense, but it's certainly AGI-complete for a huge range of economically valuable tasks.
We built benchmarks to measure the path to AGI. Astra just saturated them. Now we find out what happens next.