ExSitu

Launch Week: Four Major AI Companies, and One Model Classified "Critical"

Anthropic, Google, OpenAI and Meta launched new models almost in the same week. OpenAI's GPT-6 Astra is the first model officially classified "critical" in cyber capabilities.

5 min read

Launch Week: Four Major AI Companies, and One Model Classified "Critical"
Unsplash · Walls.io

In the first week of September 2026, the four leading AI labs launched new models almost simultaneously: Anthropic's Claude Fable 5.1 (September 1), Google's Gemini 3.8 Flash (September 2), OpenAI's GPT-6 Astra (September 3–4) and Meta's Muse Spark 1.3 (September 2).

What Each One Brings

  • Claude Fable 5.1 — focused on long-horizon, high-stakes tasks that run for hours and span several applications, with a roughly 25% reduction in typical cost thanks to a 75% cut in the price of cache reads.
  • Gemini 3.8 Flash — focused on software engineering, agentic tasks and multi-step reasoning, including a dedicated Cyber version for identifying vulnerabilities and generating fixes, at an entry price of $0.75 per million input tokens.
  • Meta's Muse Spark 1.3 — supports text, images, video and files in a context window of over one million tokens, and requires about 20% fewer tool calls and about 25% fewer tokens than the previous model.

GPT-6 Astra: The First Model Classified "Critical"

GPT-6 Astra is OpenAI's first model to cross a "critical" threshold in cyber capabilities under the company's internal Preparedness Framework — scoring 100% on ExploitBench and 42.4% on ExploitGym, compared with 78.5% and 30.3% respectively for the previous model.

Because of the classification, access to Astra is turned off by default for business customers and requires explicit manual activation. The public version refuses advanced offensive tasks such as generating exploit code (proof-of-concept), and OpenAI's Daybreak program grants limited access to verified defenders only.

The Test That Matters More Than the Score

OpenAI also ran a test for exceeding predefined authority: Astra scored 0% overreach, compared with 48% for the previous model, Sol. In other words, the model stayed within the boundaries set for it — but that shifts the practical responsibility to the organization using it: defining the boundaries correctly in the first place.

What This Means for Your Organization

  • The four major AI companies — Anthropic, Google, OpenAI and Meta — launched new competing models in practically the same week, at the beginning of September 2026. If you are currently comparing tools for your organization, now is the right time to refresh the comparison: the prices and capabilities of all the major vendors moved together in the same week, and the comparison you did two months ago no longer reflects the real state of the market.
  • GPT-6 Astra is OpenAI's first model officially classified "critical" in cyber capabilities under its internal Preparedness Framework, and the company has blocked it by default for business customers until a system administrator explicitly enables it manually. If you are considering using it — first check who in your organization is actually authorized to enable such access, and why it is needed at all.
  • The test point that is more concerning than the high score: Astra scored 0% on the test for exceeding predefined authority, compared with 48% for the same company's previous model — meaning it stayed within the boundaries carefully set for it. But that is exactly what shifts responsibility to you as an organization: making sure the boundaries you yourself define for the agent are correct and narrow enough from the start.
  • The professional logic emerging from the analyses around the launch is clear: governance of enterprise AI is shifting from the question "which model to choose" to the question "how much damage can a specific identity cause before some control stops it." That is a question of permissions, identities and internal monitoring — not a question of choosing the right external vendor.
  • Do not let this launch race pull you into an immediate, unconsidered upgrade decision. Meta's input format flexibility, Google's low entry price and Anthropic's significant cost reduction are all real and weighty considerations — but only after you have defined exactly what you need the model to do for you.

Sources