The benchmark score appeared on a leaderboard on a Thursday morning, and engineers noticed immediately.
A single number sat at the top of the column: 99.9 percent. The test it referred to was one the entire ARC-AGI series had been designed to make hard, and every frontier model had failed it badly just months before. And the workspace this new model carried into each conversation was larger than anything a flagship chat model from this lab had ever held before. How did both records land on the same launch?
Why a bigger workspace changes what an AI model can actually do
Every large language model operates inside what engineers call a context window, the total amount of text, code, or images the model can see at once during a single conversation. Think of it as the model’s working desk: the bigger the surface, the more material it can hold without having to set anything aside.
When the desk is too small, the model forgets earlier parts of a long document or loses track of code written fifty pages back. That is not a theoretical problem; it is the reason developers have had to break large projects into fragments and pass them in pieces.
GPT-6 Astra carries a 1,050,000-token context window and a maximum output of 128,000 tokens. A million tokens maps to roughly 750,000 words, or about ten full-length novels read and held simultaneously. No previous flagship chat model from this lab had ever worked at that scale inside a single session.
The mechanism that makes the window useful is how the model handles long agent runs without losing detail. In Codex, Astra introduces a new way to preserve and retrieve context when the context window fills. Historically, models used compaction to summarize work during long sessions; each compaction could leave out details about why a fix failed or how a component behaves. Astra can instead keep notes across context windows, preserving accumulated details without repeatedly compressing them into a single summary, and earlier context windows remain searchable. That shift matters most for developers debugging complex systems across hours of work.
What the model does when it takes control of a computer screen
The context window is only one half of the story. Astra is built as a computer-use model, and it scored 72.6 percent on OSWorld 2.0, setting a new frontier on computer and browser use at roughly 47 percent less time per task than its predecessor GPT-5.6 Sol. That benchmark tests how reliably an AI can navigate a real desktop, opening applications, filling forms, and moving files without a human steering it turn by turn.
Cutting roughly 47 percent off the average task time is the kind of gain that changes whether autonomous computer use is practical inside a business workflow. At the previous pace, a human watching over the model’s shoulder was often faster. The math begins to shift when task time drops that sharply.
The benchmark scores and where they came from
ARC-AGI-3 formally launched on March 25, 2026, and where earlier ARC-AGI benchmarks tested fluid pattern abstraction on static grids, ARC-AGI-3 tests agentic intelligence: the ability to explore a strange world, build an internal model of how it behaves, infer what the goal might be, then plan and execute a working solution. At launch, frontier models from Anthropic, Google DeepMind, OpenAI and xAI all scored below 1 percent on the semi-private set, while untrained members of the public solved 100 percent of the environments. Astra’s 99.9 percent score was produced using OpenAI’s Provider Adapter harness; the ARC Prize Foundation’s own Standard harness put the same model at 62.7 percent , making it the highest verified score on the series either way.
The published benchmarks also include 97.6 percent on FrontierMath Tier 4 and a perfect score on ExploitBench. Epoch AI, which runs the benchmark, says OpenAI funded its development and has exclusive access to part of it , a caveat worth keeping in mind alongside the score. FrontierMath Tier 4 covers research-grade mathematics problems previously considered beyond any AI system.
The model launched on September 3, ending months of speculation. Pricing reflects the step up in capability: $10 per million tokens for standard input and $50 per million tokens for output. That is 2.5 times higher than the previous flagship’s promotional rate. As global chip sales keep climbing, the software layer is where competitive pressure concentrates.
The part of the model that is deliberately held back
Not every capability is available to every user. The lab has adopted a two-track access mechanism: the public model refuses advanced cyber tasks such as proof-of-concept exploits, while looser safeguards go to vetted organizations through a separate program. This is the first model from the lab to reach the Critical level of cybersecurity capability under its Preparedness Framework.
A perfect ExploitBench score means the model can, in principle, identify working software vulnerabilities end to end. Allowing that capability to anyone who pays the API price would hand that power to actors with no oversight.
This tension is becoming a defining pattern across frontier AI. The push to open large-scale data runs directly into this dynamic: openness and controlled deployment pull in opposite directions at the capability frontier. The lab’s formal dual-track deployment commitment is the most explicit any major AI developer has yet put in writing, a distinction other analysts have also drawn.
What the 1 million-token era means for everyday users
For a lawyer reviewing contracts, a researcher reading a body of literature, or a developer tracing a bug through a large codebase, a million-token window changes the nature of the interaction entirely. It shifts from a series of short questions into something closer to a sustained working session with a highly capable collaborator.
Even so, larger is not always better in practice. Once input crosses into long-context territory, the lab applies higher rates to the entire request, so a long-context workflow requires careful cost control. Sending an entire codebase when only three files matter is expensive and may dilute the model’s attention. The pricing structure is, in effect, a nudge toward precision.
The record benchmark scores and the record context length arriving together suggest the two advances are complementary. A model that reasons more reliably makes better use of a very large workspace, since it can maintain coherent goals across a much longer span of material. That combination is where the practical gap between these systems and the ones they replace becomes most visible.
