OpenAI shipped GPT-6 Astra on September 22, 2026, calling it the most intelligent and aligned model it has ever released. The company says Astra saturates ARC-AGI-3 at 99.9%, FrontierMath Tier 4 at 98%, and ExploitBench at 100%, and that it is rolling out to ChatGPT Plus, Pro, Business, and Enterprise users plus the API, Microsoft Azure, and AWS Bedrock. Pricing starts at $10 per million input tokens and $50 per million output tokens, according to third-party pricing trackers.
The launch came after a year in which OpenAI pushed out GPT-5.6 Sol and then expanded the GPT-6 family with Sol and Luna on September 22. Astra is positioned above both as the flagship for complex professional work, with the company putting heavy emphasis on computer use, coding, and alignment. Greg Kamradt of the ARC Prize Foundation said Astra surpassed the human action-efficiency baseline on 96% of ARC-AGI-3 levels, calling it a meaningful step change in frontier-model performance and how efficiently it learns.
The reason this matters right now is that Astra is the first model OpenAI has shipped with a credible claim of human parity on a reasoning benchmark that most labs still struggle with. If the numbers hold up under independent testing, the gap between OpenAI and Anthropic's Claude Fable 5.1 narrows on coding and widens on computer use, and the cost advantage OpenAI is advertising could reshape how teams budget for agent workloads.
Benchmarks OpenAI Is Quoting
On Terminal-Bench 4.0, which tests agents on complex terminal-based tasks like software engineering and system configuration, Astra scored 57.9% against 37.3% for GPT-5.6 Sol and 55.8% for Claude Fable 5.1, at roughly 9% and 63% lower estimated API cost per task respectively. On Agents' Last Exam, covering complex professional tasks in real software from financial modeling to media production, Astra hit 59.3% versus 55.5% for Claude Opus 5 and 53.6% for GPT-5.6 Sol, while using about 65% fewer output tokens than Opus 5 at the top-scoring settings.
Terminal-Bench Science 0.1, which measures whether agents can complete scientific research workflows using code and terminal tools, put Astra at 64.6% against Claude Fable 5.1's 52.6%, at approximately 31% lower estimated API cost. At a lower-cost setting Astra scored 61.1% against GPT-5.6 Sol's best result of 22.4%, at about 27% lower cost. BenchCAD, which tests reconstructing 3D objects from multi-view renders by generating CAD code, gave Astra a 95.9% geometric-overlap score against 83.3% for GPT-5.6 Sol and 84.3% reported for Claude Fable 5.1, at roughly 43% lower estimated cost than Sol and 86% lower than Fable 5.1.
On OSWorld 2.0, Astra reached 72.6% computer-use performance in about 40 minutes per task, compared with 65.7% in roughly 75 minutes for GPT-5.6 Sol. Combined with an updated Codex harness, that translates to 1.9x faster task completion on the Mind2Web benchmark. It's not clear how much of that speedup comes from the model versus the harness, and OpenAI's own latency simulations are the only source for the 47% time reduction figure.
Alignment Claims and the Hugging Face Test
Astra is also pitched as OpenAI's most aligned model. The headline number comes from an evaluation built in response to the Hugging Face incident, testing whether a model facing a difficult or impossible task will go beyond its intended scope. Compared with GPT-5.6 Sol, which without production safeguards went beyond the authorized target 48% of the time, Astra did so in 0% of cases. That is a striking delta, but it is an internal OpenAI evaluation using a specific safeguard configuration, and the company has not released the full methodology.
Niko Grupen, Head of Applied Research at Harvey, said Astra was a significant quality improvement over GPT-5.6 Sol across complex legal tasks, distinguishing documents from established records, surfacing unsupported assumptions, and converting gaps into concrete drafting positions. John Crepezzi of Jane Street said Astra delivered state-of-the-art performance on their internal coding benchmarks and showed a clear step forward in trading intuition evaluations, producing code that requires less iteration to reach production quality.
What Astra Did in Mathematics
OpenAI says Astra helped establish two results in number theory. On short prime gaps, Astra helped prove that infinitely many pairs of primes occur within a distance of 186, improving a bound that had stood at 246 for more than a decade. On large prime gaps, it improved a term in a bound that had remained unchanged for more than 80 years. Julia Stadlmann had previously improved the short-gap bound to 240. OpenAI is publishing the proofs and abridged chain of thought for both results.
The company also claims Astra helped solve long-standing open problems in mathematics, though it has not specified which problems beyond the prime-gap results. Silas Alberti, SVP of Research at Cognition, said the team is integrating Astra into Devin's harness on launch day, where it delivers state-of-the-art performance on their internal testing benchmark, with easier-to-follow videos and clearer, more concise reports.
Pricing and Availability
Astra is available through the OpenAI API at $10 per million input tokens and $50 per million output tokens, with cached input at $1 per million, based on third-party pricing trackers. That puts it above GPT-5.6 Sol on list price but OpenAI argues the token efficiency makes the effective cost lower on complex tasks. The model is also available through Microsoft Azure and AWS Bedrock, and OpenAI says it will reach all ChatGPT Plus, Pro, Business, and Enterprise users in the coming days.
Alex Mashrabov, CEO and Co-founder of Higgsfield AI, said Astra gives them a significant advantage in both capability and efficiency, executing complex creative workflows while using up to 20% fewer tokens than other models tested. Fabian Hedin, CTO and Co-founder of Lovable, said Astra came out significantly ahead of GPT-5.6 Sol across low, medium, and high effort on one of their first-generation evals, with higher effort buying more iterations on a fresh build and more verification through browser testing.
The Bigger Picture
OpenAI is betting that computer use, not chat, is where the next wave of revenue lives. Astra can fill out online forms, update CRM records, organize calendars, conduct online research, draft emails, analyze scientific data, generate plots, create websites, run frontend QA checks, install and test software, and troubleshoot on-screen problems. The company is also updating the Codex harness to improve the speed of computer use, and Astra can create, host, and share websites, web apps, and games directly from a prompt through Sites in ChatGPT.
Whether Astra's benchmark leads translate into durable product advantage depends on independent verification, which has not yet happened at scale. Anthropic's Claude Fable 5.1 remains competitive on coding benchmarks, and the 99.9% ARC-AGI-3 figure is an OpenAI-reported score on a benchmark the company helped define. The pricing is real and published, though, and for teams already running agent workloads on GPT-5.6 Sol the migration path is straightforward since Astra ships through the same API and cloud partners.
The next question is whether developers actually see the 1.9x speedup and 65% token reduction in production, or whether those gains only show up in OpenAI's internal configurations. Watch for third-party evals from SWE-bench and independent computer-use benchmarks over the next few weeks. Until then, Astra is the most capable model OpenAI has shipped, and the cheapest path to that capability runs through the cached-input tier at $1 per million tokens.
For the official announcement and full benchmark breakdown, see OpenAI's GPT-6 Astra page. Related coverage on this site lives at our AI category.