claude-all-models

Route tasks across the Claude lineup — model selection, delegation, cost/speed trade-offs, cache discipline. Trigger for any model-selection, routing, multi-model, or cost-optimization decision.

SKILL.md

GitHub

All Models — Fleet Protocol

Treat the lineup as one machine with four gears. The single biggest lever for correctness AND cost is routing each task to the cheapest model that reliably does it — then never re-doing work on a bigger model that a smaller one already finished.

Staleness guard

The lineup table below was verified 2026-10-08. Model lineups change: before load-bearing routing/cost decisions, confirm the current lineup (Models API / platform.claude.com or the claude-api skill). New models slot into the same gear logic; update the table and log the change in learnings.md.

Self-improvement (do this first and last)

  1. At start: read learnings.md in this skill's folder if it exists. Apply relevant lessons.
  2. At end of every use: append one dated bullet — routing decisions that paid off or backfired, cost surprises, delegation patterns that worked. Merge instead of duplicating; delete disproven bullets.

The lineup (current, verified)

Model ID Context / Max out $/MTok in–out Role
Fable 5.1 claude-fable-5-1 1M / 128K 10 – 50 Hardest problems only; long-horizon autonomy
Opus 5.5 claude-opus-5-5 1M / 128K 4 – 20 Default flagship: serious coding, agents, review
Sonnet 5.5 claude-sonnet-5-5 1M / 128K 2 – 10 Near-Opus coding at volume; interactive work
Haiku 5.5 claude-haiku-5-5 1M / 128K 0.10 – 0.50 (≤100K prompt; 0.50 – 2.50 above) Fan-out, classification, routing, latency paths

Previous generation, still served (pinned projects, fallback targets): Fable 5 (claude-fable-5, 10 – 50), Opus 5 (claude-opus-5, 5 – 25), Opus 4.8 (claude-opus-4-8, 5 – 25), Sonnet 5 (claude-sonnet-5, 2 – 10), Haiku 4.5 (claude-haiku-4-5, 1 – 5, 200K / 64K). New work targets the four rows above.

Per-model depth lives in the sibling skills: fable-5-1, opus-5-5, sonnet-5-5, haiku-5-5 (previous generation: fable-5, opus-5, opus-4-8, sonnet-5, haiku-4-5) — load the one matching the model you're actually driving.

Routing rules

  1. Start one gear lower than instinct says. Escalate on demonstrated failure, not anticipated difficulty. Sonnet 5.5 at medium/high handles most of what people reflexively send to Opus.
  2. Never retry a failure on the same gear more than once. Two failures → escalate with the failure context attached ("Sonnet tried X, output was wrong because Y").
  3. Never send finished work back up. If Haiku classified 500 items, have the bigger model spot-check a sample, not redo them.
  4. Orchestrator high, workers low. One Opus/Fable orchestrator delegating to Haiku/Sonnet subagents beats one giant model doing everything serially — cheaper and faster.
  5. Match effort before matching model. Bumping Sonnet medium → high → xhigh is cheaper than switching to Opus; try it first. Effort levels are recalibrated per model (Opus 5.5 medium ≈ Opus 5 high) — re-sweep after every model change, never carry a level over.

API-surface differences that bite when switching models

  • Thinking: Fable 5/5.1 and Opus 5.5 = always on (omit param; disabled 400s) · Sonnet 5.5 = adaptive when omitted, disabled 400s, off only via {type:"between_tools"} at high or below · Opus 5 / Sonnet 5 = adaptive when omitted, disabled allowed (Opus 5: high or below) · Opus 4.8 = off unless {type:"adaptive"} set · Haiku 5.5 = adaptive when omitted, disabled allowed at high or below, budget_tokens 400s · Haiku 4.5 = legacy budget_tokens only.
  • Forced tool_choice (any/tool) 400s on Fable 5.1, Opus 5.5, Sonnet 5.5 — use auto + strict: true + a prompt naming the tool. Still allowed on older rows.
  • effort: supported Fable/Opus/Sonnet/Haiku 5.5 (low→max); errors on Haiku 4.5. Default is medium on Opus 5.5 and Haiku 5.5, high elsewhere — set it explicitly.
  • Sampling params: rejected on Fable/Opus/Sonnet 5+ and Haiku 5.5; allowed on Haiku 4.5. Assistant prefill and manual budget_tokens also 400 on Haiku 5.5. Haiku 5.5 has no server-side refusal fallback.
  • Thinking blocks are bound to their producing model and conversation on the 5.5 / Fable 5.1 tier: keep histories append-only, and expect a fallback model to run without them.
  • Switching models mid-conversation invalidates the prompt cache — spawn a subagent on the other model instead of swapping the main loop.
  • Tokenizers differ — re-baseline max_tokens with count_tokens per model, never reuse counts.

Cost discipline

  • Prompt caching first: stable prefix (tools → system) frozen, volatile content last; verify cache_read_input_tokens > 0.
  • Batches API = 50% off anything non-latency-sensitive; pair with Haiku for the cheapest possible per-item cost.
  • Refusal fallback (Fable 5.1, Opus 5.5, Opus 5, Sonnet 5.5): ship betas: ["server-side-fallback-2026-07-01"] + fallbacks: "default" so classifier declines degrade instead of failing (Claude API only for Sonnet 5.5; other platforms use SDK middleware). reasoning_extraction declines are never retried. Older array form: fallbacks: [{"model": "claude-opus-4-8"}].
  • Log which model produced what; when reviewing costs, cut by moving traffic down-gear before cutting features.

Regression eval — not yet built

No harness exists to catch a skill or routing change regressing across the lineup. Not worth building speculatively. Build it when: the same skill shows behavior drift after 2+ routing/model changes to it — that's the signal a fixed prompt-set eval would have caught earlier and cheaper than debugging it live.