GPT-6 Astra: the first Critical-cyber model, priced exactly like Fable 5.1

GPT-6 Astra went GA between September 3 and 5, priced at $10 per million input tokens and $50 per million output, with a 1.05M context window and 128K output. That is, to the cent, what Anthropic charges for Claude Fable 5.1. The frontier tier has converged on a price, which means the vendors now have to compete on something else. But the more interesting fact about Astra is not the price, it is the label: OpenAI says it is the first model to cross the Critical threshold for cyber capability in its Preparedness Framework, and the first one broadly deployed at that level.

A strange five weeks

Astra was announced on August 1 as a model that had solved ten open problems in mathematics and theoretical computer science. No model card, no price, no availability, and it was unclear whether it was GPT-6 or a GPT-5 point release. Then on September 1 OpenAI published the cyber evaluation: 100% on ExploitBench; two previously unknown vulnerabilities in V8, Chrome's JavaScript engine, found during the eval; and an exploit chain the model built that escaped the browser sandbox and ran commands on the host after the browser opened an HTML file. Days later it had a name, a price and a rollout to ChatGPT Plus, Pro, Business and Enterprise, plus the API, Azure and Bedrock. Greg Brockman is framing it as the arrival of the AGI era. On llm-stats it sits at #3 overall with a 96.0% average, behind GPT-5.1.

What Critical means operationally

OpenAI's response is a set of controls, not a refusal to ship: stricter isolation, encrypted checkpoints, trajectory monitoring, and a phased cyber program called Daybreak for the offensive capabilities. Google did the equivalent the same week with Gemini 3.8 Flash Cyber gated behind its Fairwind program. The pattern is worth naming: intelligence and permission to use it are now separate products. For a security team the threat model changes from AI helps an attacker to AI is the attacker, and the dependencies nobody ever reviewed by hand are now reviewable by a model in minutes. Run that against your own supply chain before somebody else does.

When every frontier model costs $10 and $50, the price list stops being a differentiator and the cache, the safeguards and the uptime become the product.

The convergence, and what to compare instead

With input and output priced identically, three things decide the bill. Cache economics first: Anthropic cut Fable 5.1 cache reads by 75% to $0.25 per million, 2.5% of the input price, and claims roughly 25% savings on typical loads and up to 45% on long agentic runs. OpenAI's equivalent numbers are what I would ask for before signing anything; I have not seen them published at GA. Safeguards second: the Critical tier means some capabilities are gated and some traffic is monitored, which matters if your agents do anything that looks like security research. Availability third: Astra is on Azure and Bedrock at launch, and whether your region and quota are there is a sharper question than which model is smarter.

• Benchmark on your task, not the leaderboard; #3 at 96.0% says nothing about your codebase. The benchmark page has the published numbers side by side.
• Model the cache hit ratio before the token price; that is where the 2x differences hide now.
• Read the Daybreak and Preparedness gating terms if agents touch security tooling.
• Keep the router: price parity is the best argument yet for not hardcoding a vendor.

The honest limitation

The ten solved problems from August 1 are still, as far as I can find, unaccompanied by anything verified outside OpenAI, and the AGI framing is marketing until the distance-to-AGI measures move. The V8 zero-days are the concrete claim, and they are OpenAI's own account of its own eval. The rest is a price list and a promise, and the price list is the part I can check.