Claude Model Picker

A decision engine for choosing which Claude model to run per task across the 2026 lineup — Fable 5, Opus 4.8, Sonnet 4.6, and Haiku 4.5 — with the cost math worked out so you stop paying Opus rates for Haiku work (and stop sending Haiku at problems that need a reasoning model). Built for engineers wiring up agent loops, RAG pipelines, and batch jobs who keep hard-coding one model everywhere.

When to use it

• You default claude-opus-4-8 into every call and your bill reflects it.
• A task fans out into sub-steps of wildly different difficulty (plan with one model, grep/classify with another) and you want a routing rule.
• You're sizing a high-volume endpoint (classification, extraction, moderation) and need the per-1M-token math before committing.
• You hit a refusal or a latency wall and want to know whether a different tier or a fallback chain is the right fix.

What it does

The skill encodes the current price and capability table and applies it to your described workload:

It reasons about the levers that actually move cost and quality: output_config.effort (low→max, plus xhigh on Fable 5/Opus 4.8/4.7), adaptive thinking (thinking: {type: "adaptive"} — budget_tokens is gone on the 4.7+ tier), Task Budgets for agentic loops, and prompt caching break-even. It outputs routing patterns — model cascade (cheap model first, escalate on low confidence), fallback chains (Fable 5 → Opus 4.8 via the server-side fallbacks param), and subagent-on-Haiku splits that preserve the main loop's cache.

Usage example

curl -fsSL --create-dirs -o ~/.claude/skills/claude-model-picker/SKILL.md https://claudeskil.com/skills/claude-model-picker/SKILL.md

> I summarize ~40k support tickets/day, 1.5k tokens in, 200 out.
  Currently all on Opus 4.8. What should I route to?

Haiku 4.5 handles ticket summarization well.
  Opus 4.8:  40k × (1500×$5 + 200×$25)/1M  = $500/day
  Haiku 4.5: 40k × (1500×$1 + 200×$5)/1M   = $100/day
Route to claude-haiku-4-5; cascade to claude-opus-4-8 only when
the summary confidence flag trips. Batch API halves it again.

Tips & gotchas

• Effort beats model swaps for cost. Dropping Opus 4.8 from xhigh to medium often saves more than downgrading to Sonnet while keeping more capability. The skill checks effort before tier.
• Fable 5 is not the default upgrade. At $10/$50 it's for genuinely hard long-horizon work; for "use the best model" the answer is usually claude-opus-4-8.
• Caching changes the math. Cache reads cost ~0.1×; a large shared prefix can make Opus cheaper per-request than uncached Sonnet. The picker factors in your read/write ratio.
• Don't switch models mid-conversation — it invalidates the prompt cache. Use a subagent on the cheaper model instead.

FAQ

Does it call the API? No. It's a reasoning/routing layer over the price-and-capability table. It recommends model IDs and emits routing code; you keep your own client.

Fable 5 vs Opus 4.8 — when does the 2× price pay off? When the task is long-horizon agentic or first-shot implementation of a well-specified system that Opus measurably can't complete. For most coding, analysis, and chat, Opus 4.8 at high/xhigh is the sweet spot.

Can it design fallback chains? Yes — including Fable 5's refusal stop_reason handling with the server-side fallbacks: [{model: "claude-opus-4-8"}] parameter, and cascade-on-confidence patterns for cost.