Fast vs Deep tier.

The composer has a Fast/Deep toggle. Deep = Claude Sonnet 4.6 (LADLE's default). Fast = Claude Haiku 4.5. The models are 3× apart in price, roughly 2× apart in speed, and meaningfully different in behavior.

What Deep does well (the default)

Sonnet handles: multi-step reasoning, nuanced drafting, code review, complex explanations, long-form writing, anything that benefits from careful attention. It's slower and more expensive per turn. It's also the reason people pay for LADLE — the model quality on assistant tasks is the product.

What Fast does well

Haiku handles: quick factual questions, short reformatting, simple classifications, "just answer this directly" tasks. It replies in 1-2 seconds vs Sonnet's 5-10. For the specific class of tasks where Haiku is enough, it feels faster and less deliberate — which is the right feel.

Where the gap matters

- **Long-form drafting.** Sonnet is meaningfully better. If you're editing prose, don't switch to Fast — you'll feel the drop. - **Multi-step reasoning.** Sonnet holds context across the reasoning chain more reliably. Haiku can lose track on 3+ step problems. - **Nuanced code review.** Sonnet catches subtle issues Haiku misses. - **Voice preservation.** Sonnet mirrors tone better; Haiku tends to smooth out.

Where the gap doesn't matter

- **Titling a chat.** Fast is fine (and is what LADLE uses under the hood). - **Classifying an email.** Fast is fine. - **Formatting a list.** Fast is fine. - **One-liner factual questions.** Fast is fine, and 8 seconds saved per turn compounds.

The scoped-edit exception

In the design canvas, when you have an element selected (via the inspector), turns route to Fast (Haiku) regardless of the composer toggle. Reason: element-scoped edits ("make this button bigger") don't need Sonnet's depth — they need surgical precision + sub-second response. Full design generations still route Deep.

Budget-driven downgrade

If you're near your monthly compute cap, the router force-downgrades Deep to Fast to preserve capacity. You'll see the header "Deep model paused — Fast responses continue until [reset]." This is graceful degradation — you keep getting help, just from the smaller model.

When to explicitly pick Fast

- You want a fast, direct reply and don't care about polish. - You're on a heavy-usage day and want to conserve budget. - The task is one Haiku obviously handles (short answers, quick lookups).

When to explicitly pick Deep

- You want the better model regardless of speed. - The task benefits from careful reasoning or voice. - The reply will be pasted into a real deliverable and needs to hold up.

What you can see

Kitchen panel (Cmd+K) on any reply shows which model produced it. If a Deep reply came from Haiku, the routing reason will say why (usually budget degrade or your explicit Fast pick).