Fable 5, Sonnet 5, and GPT-5.6 raised the safety and routing bar

The release names are less interesting than the system around them. The Fable 5 and Mythos 5 announcement, Sonnet 5 release, and GPT-5.6 release all point at the same operational question: what changes around the model?

Fable 5, Sonnet 5, and GPT-5.6 make raw capability harder to separate from safety posture, effort levels, pricing, routing, and tool behavior. A model can be strong at long coding tasks and still be the wrong route for a policy response. A smaller model can be fine for classification and poor as the final writer.

Route the shape of the work:

• fast extraction and classification
• routine writing under source constraints
• agent planning with tools
• policy-adjacent work
• ambiguous reasoning
• background summaries where latency matters less

Safety changes the interface. A refusal may be correct, incorrect, useful, annoying, or the result of a blocked tool call. An agent refusal halfway through a run can look like a stall unless the system records what happened.

The eval should separate those outcomes:

• correct refusal
• incorrect refusal
• correct continuation
• unsafe continuation
• useful clarification
• unnecessary clarification
• blocked tool call
• allowed tool call with insufficient context

Effort controls are another budget decision. The system should record why it escalated rather than treating extra thinking as a magic fix:

{
  "workflow": "security_alert_triage",
  "initial_model": "fast_route",
  "initial_effort": "low",
  "final_model": "frontier_route",
  "final_effort": "high",
  "escalation_reason": "conflicting_entities_and_external_action",
  "outcome": "accepted_with_human_review"
}

Workflow cost is larger than one request. Planning, retrieval, tool calls, retries, verifier passes, and the final answer all contribute. Measure cost per completed workflow, not only cost per model call. A cheap route that retries twice may cost more than a stronger route that finishes once.

Routing also needs stop conditions:

• maximum retries per workflow
• maximum cost by task class
• maximum latency before handoff
• required review after policy failures
• refusal when the action is out of scope
• downgrade when the task is plainly simple

Stopping is a safety control and a quality control. Users can work with a clear refusal. They cannot easily review an agent that keeps escalating and eventually returns a vague compromise.

I would test the wrapper around the model with stale retrieval, malformed tool results, policy-adjacent requests, bad early assumptions, tasks that should refuse, tasks that should ask one narrow question, and tasks that should stay cheap.

The comparison should include the current system. A previous model with a tighter prompt, narrower tools, and a better verifier may beat a new release in a loose wrapper. Inspect pass-to-fail flips by hand, not only the launch benchmark.

The frontier is now a routing and maintenance problem. The useful upgrade is the one that can explain why it was chosen, why it escalated, and why it stopped.