Model selection for the agent workflow

Recommended claude-code model/effort settings per stage of the grill → PRD → issues → triage → AFK → land workflow, in the provider catalog's own values (as of the Claude 5 family, 2026-08):

Stage	Lab surface	Model	Effort	Why
Grilling, /to-prd, /to-issues	manual instance	fable	high	Highest-leverage thinking stage: design sessions are low token volume, a human is present, and question/slicing quality compounds through every later stage. Fable's edge — scoping ambiguous problems — is exactly what grilling is.
Triage, interactive /land-pr	manual instance	opus[1m]	high	Brief-writing and bug reproduction want strong reasoning, but not frontier pricing.
AFK runs	repo AFK model/effort defaults	opus[1m]	xhigh	Opus is the agentic-coding workhorse and xhigh its documented setting for unattended coding; nobody waits on latency. A good agent brief removes exactly the ambiguity Fable would otherwise be for.
Autoland lander + escalate	repo lander_model / lander_effort	opus[1m]	high	The unattended gate in front of the default branch. Review accuracy holds at high; don't go below it on an auto-merging path.

• Keep fable off unattended surfaces (AFK, lander): ~2× Opus pricing, and its safety classifiers can occasionally refuse benign security-adjacent work — interactively that's a nudge, in a pipeline it's a stalled run.
• sonnet at xhigh is a credible budget tier for small, well-briefed AFK slices (dependency bumps, mechanical changes) if per-issue model routing ever exists; while model/effort stay per-repo, the practical AFK default remains opus[1m].
• haiku has no role in the workflow (lab uses it internally only for the credential-refresh poke).
• Roadmap idea: a Fable escalation tier — rerun an issue on fable after Opus has failed it twice, before routing it to ready-for-human.