Frontier AI vendors now ship one model at multiple access levels: a generally available tier with capability restrictions, and gated tiers granting vetted organizations capabilities withheld from the public. Which tier an organization holds is a governance fact, because it determines what its agents can actually do.
Capability tiering became explicit in 2026. When a model's abilities cross a vendor's own danger thresholds, the vendor no longer chooses between shipping and not shipping; it ships different authorities to different users. That decision used to be internal to labs. Now it is a procurement variable, and a governance record has to capture it.
The tiers, by vendor
| Vendor and model | Tiering, as published |
|---|---|
| OpenAI: Astra | First model designated Critical for cybersecurity under OpenAI's Preparedness Framework (September 2026). The broadly available tier refuses advanced cyber requests, including proof-of-concept exploits. Advanced cybersecurity capability is limited to an approved group of testers, expanding to defensive users through the Daybreak Blue program.<sup><a href="#r1">[1]</a></sup> Astra shipped generally as GPT-6 Astra in September 2026, described by OpenAI as state-of-the-art on computer use, browsing, software engineering, and cybersecurity, with this tiering in effect at launch.<sup><a href="#r4">[4]</a></sup> |
| Anthropic: Claude Fable 5 / Claude Mythos 5 | One underlying frontier model, two products. Fable 5 is generally available with additional safety measures for dual-use capabilities. Mythos 5 is available only to approved organizations under trusted access, with cybersecurity capability extended to defenders through dedicated programs.<sup><a href="#r2">[2]</a></sup><sup><a href="#r3">[3]</a></sup> |
| Other vendors | Added when tiering is announced and verifiable against primary sources. A vendor without published tiering is not listed as untiered; it is listed as unverified. |
What tiering means for governance
Tiering moves a safety decision from the lab into your delegation records. Two organizations running "the same model" may hold different authorities, and an agent built on a gated tier can do things the generally available tier refuses. So the five questions of AI agent governance begin one step earlier than most programs expect: before asking what your agent was authorized to do, record which tier of capability your organization was granted, by whom, under what program, and with what conditions attached. That grant is part of the provenance of every action the agent takes.
Frequently asked questions
What is a frontier model access tier?
A vendor-defined level of model capability and access: typically a generally available tier with restrictions, and gated tiers granting vetted organizations capabilities withheld from general availability.
Why does the access tier matter for governance?
Because it determines what an organization's agents can actually do. The tier grant belongs in the governance record the same way delegated authority does: which tier, granted by whom, under what program, with what conditions.
References
- OpenAI, Path to Astra: critical capabilities and frontier safeguards, September 1, 2026. View source ↗
- Anthropic, Claude Fable 5 and Claude Mythos 5, 2026. View source ↗
- Anthropic, Bringing the cybersecurity capabilities of Claude Mythos 5 to more defenders, 2026. View source ↗
- OpenAI, GPT-6 Astra: a new generation of intelligence, September 2026. View source ↗
Cite this page
Bertram, M. (2026). Frontier Model Access Tiers. Digital Information Governance® (DIG), Framework v1.1. https://digitalinformationgovernance.com/compare/frontier-model-access-tiers
@misc{dig-compare-frontier-model-access-tiers,
author = {Bertram, Matthew},
title = {Frontier Model Access Tiers},
year = {2026},
howpublished = {Digital Information Governance (DIG), Framework v1.1},
url = {https://digitalinformationgovernance.com/compare/frontier-model-access-tiers}
}