Mythos 5.2, GPT-6 Cyber, Carbon's gated fortnight: what the restricted tier changes for teams that are not on the list
I count five frontier releases between October 5 and November 19 that most engineering teams are not permitted to use. Anthropic's Mythos 5.2, OpenAI's GPT-6 Cyber and GPT-6 Rosalind, Google's Gemini 4 Carbon for its first thirteen days, and a cyber variant of Gemini 4 Neon that Google says is going to vetted defenders. I am not on any of those lists and I assume most readers are not either. This post is about what the restricted tier does to the rest of us anyway.
Who gets what
| Model | Date | Gate | What the vendor says about it |
|---|---|---|---|
| Gemini 4 Carbon | Oct 28 | Fairwind Program, developers from Nov 10 | 88.2% on SEC-bench Pro, 54.8% on ExploitGym |
| GPT-6 Rosalind | Oct 30 | Approved organisations | Listed in the API changelog as restricted |
| GPT-6 Cyber | Nov 5 | Daybreak Red | Post-trained variant of GPT-6 Astra, rated Critical for cyber |
| Claude Mythos 5.2 | Nov 9 | Cyber and Life Sciences Verification Programs | Same underlying model as Fable 5.2, without the safeguards |
The numbers in the last column are vendor-reported, and Google notes that its external cybersecurity benchmarks were run without guardrails. OpenAI says GPT-6 Cyber developed working exploits for 948 Linux kernel vulnerabilities and found 81 previously unknown high-severity Firefox flaws. Anthropic reports Mythos 5.2 at 76.3% on CyScenarioBench against 61% for GPT-6 Cyber. Each of those figures comes from the lab that benefits from it.
The gates are real controls, not a waiting list. OpenAI's system card for GPT-6 Cyber describes identity verification, use restrictions, and organisational requirements including single sign-on, multi-factor authentication and API IP allowlisting. Individuals can apply for the lighter Daybreak Blue tier but generally not for Red.
The public model is the gated model plus a filter
Anthropic is unusually direct about this. Fable 5.2 and Mythos 5.2 are the same underlying model. Fable adds classifiers that block high-risk cyber activity, including penetration testing, exploit generation and binary vulnerability work, plus most research biology. When a query is blocked, it falls back to an older Opus model.
That design has a consequence I think teams underestimate. If part of your workload looks like security work to a classifier, some of your requests are answered by a different, older model, and your evals may not show it unless they include those requests. Anthropic reports about 60% fewer cyber blocks in 5.2 than in 5.1, which helps. It does not remove the need to log which model actually served each response.
OpenAI went the other direction with GPT-6.2. The new Astra and Sol launched with tighter cyber safeguards than their predecessors, OpenAI says those safeguards restrict some legitimate vulnerability research and security testing, and the 6.2 models are not yet available through Daybreak at all. A security team that relied on GPT-6 Astra for authorised testing should check its refusal rate before upgrading.
If a classifier can reroute your request to an older model, then which model answered is a field you need in your logs.
Retention is part of the gate
Anthropic's post also says that outside its Enterprise Frontier Safeguards program, which is in private beta and runs monitoring inside customer infrastructure, customers have 30-day retention for safety monitoring. If you chose a provider partly on zero data retention, read that sentence against your contract before you move a workload to Fable 5.2. I am not saying the term is unreasonable. I am saying it is a change that arrives bundled with a model upgrade, and model upgrades usually get approved by people who do not read retention clauses.
What defenders with access can now do
The capability claims matter even if you cannot touch the models, because people who can are pointing them at the software you run. I wrote in October that Anthropic's OSS scanner sends maintainers Mythos-class bug reports unreviewed, and that Carbon finds bugs faster than it patches them. GPT-6 Cyber adds a third pipeline producing findings at a volume no maintainer team was sized for. The practical effect for an ordinary team is more security advisories in dependencies, arriving faster, with patches of uneven quality. Patch capacity is the thing to budget.
The open-weight side has its own version of the gate. Moonshot capped offensive cyber in Kimi K3.1 Pro's training data, which I looked at in the abliteration piece, and Mistral reports that Large 4 has a higher cyber refusal rate than every other open model it tested. Restriction is moving into the weights as well as the API.
What I am adding to my own checklist
Three items. I log the serving model identifier on every response and alert when it differs from the one I requested. I keep a small set of legitimate security prompts, things like reviewing a fuzzing harness or explaining a CVE in a dependency, in my regression suite so that a safeguard change shows up as a failed test instead of a confused user. And I re-read retention terms on every major model upgrade, treating it as a contract change.
The limitation of this post is that I am describing restricted models from the outside. Everything I know about Mythos 5.2 and GPT-6 Cyber comes from the vendors' own posts, a system card and press coverage. I have not used either, and nobody without access can check the capability numbers.