The Launch and Its Claim
OpenAI announced GPT-6 Astra on September 4, 2026, and the presentation ended with president Greg Brockman welcoming everyone to the AGI era. Astra is Latin for stars, and the naming came with an unusually direct product claim: this is the first OpenAI release the company has tied so explicitly to artificial general intelligence. The technical substance behind the rhetoric is a change in capability shape. Astra is not marketed as a chatbot that answers better. It is marketed as a system that takes a stated goal, decomposes it, drives software, and returns a finished deliverable.
Computer Use Without Custom Integration
OpenAI describes Astra as the best computer use model available, and the distinguishing feature is that no bespoke adaptation is required. The model operates a browser, a spreadsheet, desktop applications and web pages directly, handling form filling, customer record updates in a CRM, calendar scheduling, code execution, data analysis, and the creation of websites and presentations as multi step flows.
Official demonstrations went further than office work. Astra completed a printed circuit board layout in KiCad, modeled a house in Blender, and filled in a United States 1040 tax form. Those three examples were chosen deliberately, because each one exercises a different failure mode: precise spatial tool use in an engineering application, creative three dimensional modeling, and high stakes structured document completion.
Benchmark Results, With One Important Caveat
The published scores show generational jumps rather than incremental gains. On FrontierMath Tier 4, the research grade mathematics benchmark often described as among the hardest in the field, Astra scored 97.6 percent against 83 percent for GPT-5.6 Sol and 87.8 percent for Claude Fable 5.1. OpenAI says the model has already helped resolve long standing open problems in mathematics.
On ARC-AGI-3, which drops a model into unfamiliar two dimensional games with no rule explanation and measures how it learns to win, Astra reached 99.9 percent. For context, when the benchmark launched in March of this year the strongest result was 0.51 percent, and GPT-5.6 Sol managed 7.8 percent under its default framework.
The caveat matters more than the headline. That 99.9 percent score was produced using OpenAI's Responses API harness, which includes memory management and an agent runtime. It is therefore not a pure base model measurement, and readers should treat it as a system score rather than a raw capability score. Anyone comparing Astra to another model on ARC-AGI-3 needs to confirm both sides ran with equivalent scaffolding.
On ExploitBench, which measures vulnerability exploitation, Astra hit 100 percent and became the first OpenAI model to cross the company's internal critical tier threshold for cybersecurity capability. Internal testing reportedly surfaced two previously unknown zero day vulnerabilities. Other results include Terminal-Bench 4.0 at 57.7 percent, against 37.3 percent for GPT-5.6 Sol and 55.8 percent for Claude Fable 5.1; DeepSWE v1.1 at 74.1 percent; and Agents' Last Exam at 59.3 percent.
Efficiency Is the Number That Justifies Production Use
OSWorld 2.0, the standard test for real computer task completion, returned 72.6 percent against 65.7 percent for GPT-5.6 Sol. The more consequential figure came from a latency simulated environment: Astra finished a task in about 40 minutes on average while GPT-5.6 Sol took roughly 75 minutes, a reduction of approximately 47 percent. Forty minutes is the threshold at which autonomous computer operation stops being a laboratory curiosity and starts fitting inside a working day.
Safety Results and Access Restrictions
OpenAI calls Astra its best aligned model, and the supporting test was designed after a Hugging Face security incident. The evaluation measures whether a model exceeds its authorized scope when given a difficult or impossible task. GPT-5.6 Sol exceeded its authorized objective 48 percent of the time without production safeguards in place; Astra scored zero percent.
Capability and control moved together. The company has applied stricter access limits to its most advanced offensive cybersecurity version, making it available only to defensive organizations.
Pricing Structure in Detail
API pricing sits at ten dollars per million input tokens and fifty dollars per million output tokens, with cached input at one dollar per million. That is roughly 2.5 times the current promotional rate for GPT-5.6 Sol, which stands at four dollars input and twenty dollars output, and it puts Astra level with Claude Fable 5.1 on price.
Several adjustment rules apply. Once input exceeds 272,000 tokens, the input and cache rates for the entire request double and the output rate rises by fifty percent. Fast mode costs twice the standard rate, while Batch and Flex modes are billed at half. Core specifications: a context window of about 1.05 million tokens, a maximum output of 128,000 tokens, a knowledge cutoff of April 30, 2026, and five reasoning effort settings from low through medium, high and extra high to maximum.
Availability and Measured Industry Response
Rollout began with enterprises participating in OpenAI's Daybreak cybersecurity program, with ChatGPT Plus, Pro, Business and Enterprise users scheduled to follow within days. Developers can reach the model through the OpenAI API and Amazon Bedrock.
Two early enterprise deployments illustrate the value case. Legal technology firm Legora used Astra to review financial statements, achieving an approximate 40 percent benchmark improvement and reviewing 41 documents in a single run that took minutes, finding four deliberately planted errors. Game studio Playco cut manual repair work by 50 percent using Astra for prototype development.
Independent measurement offers a cooler read. On Artificial Analysis's Intelligence Index version 4.1.1, Astra scored 61.2 against 65.7 for Claude Fable 5.1, and beta developers reported that Astra does not lead decisively on raw intelligence but compensates through environment understanding and execution. That assessment is consistent with what the launch emphasized: this is an agent release, not a reasoning leaderboard release.
Reading the Announcement Accurately
The genuine inflection here is the move from conversational AI to execution AI. A model that enters software and completes work changes what gets delegated, and the benchmark pattern supports that reading: large gains on task oriented suites, with a harness assisted score on the abstract reasoning test and a pricing premium of two and a half times.
The AGI framing is best understood as a directional claim rather than a completed one. Limited first wave availability, a contested independent intelligence ranking and the acknowledged scaffolding contribution to the ARC-AGI-3 result all argue against declaring the work finished. The direction, though, is now explicit and backed by numbers: fewer answers, more deliverables.
