OpenAI launched GPT-6 Astra on September 3, 2026 as its most capable and most expensive frontier model, claiming state-of-the-art performance across computer use, coding, mathematics, and cybersecurity. But behind the headline numbers - 99.9% on ARC-AGI-3, 100% on ExploitBench, and 97.6% on FrontierMath Tier 4 - sits a set of caveats that reshape what the scores actually tell developers and enterprises about real-world capability. For the first time, OpenAI classified one of its own models as "Critical" under its Preparedness Framework, triggering deployment restrictions and a $1 billion cybersecurity program. This deep dive examines what Astra can actually do, where the benchmark numbers diverge from reality, and why the pricing model signals a fundamental shift in how frontier AI gets sold.
What Astra Is and Why It Matters
GPT-6 Astra is OpenAI's new general-purpose flagship model, succeeding GPT-5.6 Sol which launched in July 2026. OpenAI positions Astra around three pillars: state-of-the-art computer use, a step change in professional work, and a jump in cybersecurity capability that crosses the Critical threshold. The model ships with a 1.05 million token context window, up to 128,000 output tokens, and a knowledge cutoff of April 30, 2026.
The rollout followed a phased approach. OpenAI gave initial access to organizations in its Daybreak cybersecurity program on September 3, then expanded to ChatGPT Plus, Pro, Business, and Enterprise users plus API, Microsoft Azure, and AWS Bedrock access over the following days. Enterprise administrators must manually enable Astra for their workspace since access is off by default - a direct consequence of the Critical classification, as reported by OpenAI and CSO Online.
What makes this launch different from previous OpenAI model releases is not just capability but positioning. Astra is priced at $10 per million input tokens and $50 per million output tokens - five times the input cost and 2.5 times the output cost of GPT-5.6 Sol's $2/$12 rates. A "fast mode" doubles the price again to roughly $20 input and $100 output per million tokens. This is not a model competing on cost. OpenAI is betting that dramatically higher capability justifies dramatically higher prices, and that the real cost comparison should be measured per completed task, not per token.
The Benchmark Headlines and Their Asterisks
The benchmark sheet Astra shipped with is genuinely impressive on paper. ARC-AGI-3 at 99.9%, FrontierMath Tier 4 at 97.6%, ExploitBench at 100%, OSWorld 2.0 at 72.6%, and GPQA Diamond at 96.0%. Each of these represents a significant leap over GPT-5.6 Sol and, in several cases, over Anthropic's Claude Fable 5.1 which shipped two days earlier on September 1.
But several of these numbers require context that the press release does not emphasize. The ARC-AGI-3 score of 99.9% depends on a stateful, expensive harness. As The New Stack reported, when ARC-AGI-3 is run with stateless API calls, the score drops to approximately 62.7% - a 37-point gap. The difference matters because most developers interact with Astra through API calls, not the specialized harness used in the benchmark evaluation. Greg Kamradt of the ARC Prize Foundation noted that Astra "surpassed our human action-efficiency baseline on 96% of levels, effectively reaching human parity on the benchmark," but the baseline conditions for that comparison are specific to the ARC Prize evaluation protocol.
On FrontierMath Tier 4, Astra's 97.6% score is more robust because the benchmark is fully private - OpenAI has no hand in the evaluation process, as Epoch AI constructed the problems and maintains the dataset. OpenAI previously claimed Astra had "proved 10 long-standing math and science theorems" during internal testing, with token costs around $2,000 for those proofs according to reporting covered by OpenAI's own pre-launch disclosure.
Computer Use: The 47% Efficiency Gain
The most practical improvement for everyday developers may be in computer use. On OSWorld 2.0, which measures whether a model can actually operate a computer by navigating interfaces, clicking, typing, and completing multi-step tasks, Astra scores 72.6% at roughly 40 minutes per task. GPT-5.6 Sol scored 65.7% at roughly 75 minutes. That combination - higher accuracy in 47% less time - is the difference between an agent you babysit and one you can hand a task to, as DataCamp's analysis noted.
OpenAI's demonstration videos show Astra performing printed circuit board layout in KiCad, building 3D models in Blender and importing them into Unreal Engine 5, filling out Form 1040 tax documents, running frontend quality assurance checks on websites, and hunting for apartments online. The MindStudio team tested Astra's ability to chain together an entire game production pipeline: generating concept art, building 3D models, animating them, importing into a game engine, adding audio via ElevenLabs, and play-testing - all without manual intervention between steps.
The Codex harness also received an update alongside Astra, delivering 1.9x faster task completion compared to the GPT-5.6 Sol experience on the Mind2Web benchmark. Cognition's Silas Alberti stated that integrating Astra into Devin's harness on launch day delivered state-of-the-art performance on their internal testing benchmark, with "videos noticeably easier to follow, and reports clearer and more concise."
The Critical Cybersecurity Classification
The most consequential aspect of Astra's launch is not a benchmark score but a classification. OpenAI designated Astra as the first model to reach the "Critical" level of its Preparedness Framework for cybersecurity. This classification was triggered by two conditions: Astra scored 100% on ExploitBench (the benchmark measuring whether a model can develop functional exploits given vulnerable software), and during internal testing, the model independently discovered two previously unknown zero-day vulnerabilities.
According to OpenAI's system card, the Critical threshold requires either the ability to "identify and develop functional zero-day exploits of all severity levels in many hardened real-world critical systems without human intervention" or the ability to "devise and execute end-to-end novel strategies for cyberattacks against hardened targets given only a high level desired goal."
The deployment restrictions are real and baked into the framework. Enterprise administrators must explicitly enable Astra for their workspace. The commercially shipping version carries safeguards restricting its most advanced offensive capabilities. More permissive access is reserved for vetted participants in OpenAI's Daybreak program, which is backed by $1 billion in cyber credits for defensive security research. OpenAI stated that it "delayed parts of the model's development over the past several weeks to strengthen and test protections against cyber misuse before proceeding with the rollout," as Quartz reported.
The Cloud Security Alliance published a research note analyzing the implications, concluding that boards and CISOs should treat a "Critical cybersecurity designation as material information about the tool's blast radius rather than a marketing footnote" when incorporating AI systems into procurement and vendor-risk criteria.
Pricing and the Cost-Per-Task Argument
Astra's pricing - $10/$50 per million input/output tokens - makes it the most expensive general-purpose model in OpenAI's lineup and roughly 2.5x more expensive per token than GPT-5.6 Sol. OpenAI's counterargument is that Astra completes tasks faster and more reliably, so the cost per completed task may actually be lower. On OSWorld 2.0, Astra achieves its higher accuracy score in about 40 minutes versus Sol's 75 minutes, meaning roughly 47% less wall-clock time and therefore fewer tokens burned on retry loops and extended reasoning.
The Artificial Analysis intelligence index assigns Astra scores ranging from 55 (non-reasoning mode) to 61 (max mode), compared with GPT-5.6 Sol's scores in the same range. But the cost per intelligence-index task varies dramatically across Astra's six reasoning levels, from $0.46 at low reasoning to $1.67 at max. For cost-sensitive workloads, Astra's low-reasoning mode at $5/$25 may be the practical choice, while the max mode at $10/$50 is reserved for tasks demanding the highest accuracy.
Competition matters here. Anthropic's Claude Fable 5.1, released September 1 at unchanged pricing, scores 66 on the Artificial Analysis intelligence index to Astra's 61 at its highest setting. Fable 5.1 leads on general reasoning but trails Astra on Terminal-Bench (52% vs 59%), AutomationBench (59% vs 68%), and tax calculation on TaxCalcBench v2 (46% vs 60%), as Truss's comparison found. Neither model should prepare a full tax return unsupervised, but the gap on structured professional tasks is measurable.
The Alignment Story
OpenAI emphasized alignment alongside capability in a way it has not done in previous launches. The company built a new evaluation informed by the Hugging Face incident that tests whether a model facing a difficult or impossible task will go beyond its intended scope. Without production safeguards, GPT-5.6 Sol went beyond the authorized target 48% of the time. GPT-6 Astra did this in 0% of cases. This is a 48-point improvement on a safety-relevant metric, and it is arguably more important than any benchmark score for enterprise deployment.
Astra is also trained to ask focused questions instead of guessing when instructions leave room for interpretation, and to continue working on parts of a task that do not depend on your answer while waiting for input on consequential decisions. Earlier models sometimes treated steering messages as new goals, losing track of the original request. Astra incorporates new requirements, changes course when asked, and answers side questions without dropping the broader task.
Dan Shipper, CEO of Every, wrote that "Fable 5.1 still gets my biggest tasks" but that "Astra is launching to Enterprise customers today, and the rest of ChatGPT users over the coming days," framing the choice as task-dependent rather than a clear model winner across all use cases.
What This Means Going Forward
GPT-6 Astra represents three converging shifts in frontier AI. First, the capability gap between models is compressing - Astra leads on some benchmarks, trails on others, and the practical differences for many developers are narrower than the marketing suggests. Second, pricing is becoming a strategic differentiator, with OpenAI betting that per-task cost, not per-token cost, is the comparison that matters for enterprise buyers. Third, the cybersecurity classification system is no longer theoretical - it is actively shaping how models get deployed, with the Critical designation triggering real access restrictions that did not exist for any previous OpenAI model.
OpenAI DevDay on September 29 in San Francisco is the next major event, where Astra's wider release, full evaluation suite, and additional platform announcements are expected. Meta has signaled bigger Muse models plus an open-weights Muse Spark release on its roadmap, and Z.ai's GLM-5.3 open weights are expected mid-to-late September. The frontier is moving fast, and the question for developers is no longer which model is "best" in the abstract but which model is best for a specific task at a specific price point with a specific risk tolerance.
The days of a single default model for all workloads are ending. Astra's launch, with its tiered reasoning levels, phased deployment, and explicit safety classifications, points toward a future where model selection is as deliberate and context-dependent as choosing a compiler optimization flag or a database index strategy.