In June, Anthropic released Claude Fable 5, the first publicly available model of its Mythos class, a tier that had previously been limited to a small group of cyber defenders and critical-infrastructure providers. In July, OpenAI answered with the GPT-5.6 family, and followed in August with an updated GPT-5.6 that exposes a slider controlling how much reasoning effort the model spends on a response.
Tiering is the real story
Fable 5 and Mythos 5 share the same underlying model; what differs is the safety layer. The broadly available tier ships with safeguards that gate dual-use capabilities in areas like cybersecurity and biology (triggering, per Anthropic, in under 5% of sessions), while the unrestricted tier is reserved for vetted organizations. Frontier capability now arrives with an access policy attached, and the major clouds (AWS Bedrock, Google Cloud, Microsoft Foundry) distribute it that way.
OpenAI's effort slider points at the same trend from the cost side: one model, many operating points. The question 'which model should we use?' is turning into 'which tier, at which effort, for which task?'
What this means for your stack
- Route by task, not by habit: pair each workload with the cheapest tier and effort level that passes your quality bar, and reserve frontier calls for the steps that need them.
- Make evals the gatekeeper: with releases landing quarterly, a regression suite over your own tasks is the only reliable way to adopt new models quickly and safely.
- Budget with effort in mind: effort controls turn latency and cost into tunable parameters; treat them as part of the product design, not an ops afterthought.
- Keep a thin abstraction over providers: tiering and gating policies differ and shift; your application logic shouldn't have to care.
We build model-agnostic LLM stacks with routing and evaluation harnesses baked in, so when the frontier moves again next quarter, adopting it is a config change, not a rewrite.