One model. On purpose.
LADLE runs on Claude from Anthropic. Two models, one published rule for which turn gets which, the same rule on every plan and every seat — and a toggle to force the bigger one whenever you want it.
Sonnet, Haiku, Opus — what we ship, when, and what we don’t.
What you get for anything with real work in it — longer prompts, code, multi-turn threads, analysis — and any time you pick Deep in the composer. Strong on prose, code, long-context reasoning, and instruction-following. 200K context window.
Anthropic's higher-capability tier, priced meaningfully more per token. LADLE doesn't route to Opus today — including it at $20/mo would either cut the meal donation or lose us money. May change if usage patterns and Opus pricing warrant it.
What you get on short, casual turns — a quick question, a one-line reply, the first couple of messages in a thread before it gets substantive. Haiku 4.5 is a genuinely capable model, and routing these turns to it is what keeps $8 of every $20 in the meal fund. You can always override it: pick Deep in the composer.
What runs your query, and what doesn’t.
We route per turn between two Claude models, and we tell you which one you got. A turn goes to Sonnet if the prompt is over 800 characters, if it's the sixth message or later in the thread, if it contains a code block, if it reads like a code or analysis request, or if any earlier turn in that chat already used Sonnet — once a conversation goes heavy it stays heavy. Everything else goes to Haiku. The composer shows AUTO / FAST (Haiku) / DEEP (Sonnet), and picking Deep forces Sonnet on any turn.
We don't route silently — every response reports the model that produced it, and you can force Sonnet on any turn. We don't vary routing by plan: Base, Max 5x and Max 20x get the identical rule, and nobody gets quietly downgraded at peak for being on the cheap tier. We don't drop a conversation from Sonnet back to Haiku mid-thread.
Sonnet on every turn, with Haiku only where you ask for it. That's the goal, and it's an economics problem rather than a technical one: at $20/mo with $8 committed to the meal fund, Sonnet-on-everything doesn't currently clear. If inference prices fall or subscriber volume grows, routing to Haiku is the first thing we retire — and we'll announce it on /changelog rather than quietly flipping it.
What it can hold. What it can’t.
The 200K window means the whole 10-K fits. Well past that, the model still tries — but its own reasoning about early parts of a very long context degrades measurably. Practical rule: if your document exceeds 300 pages, chunk to the section you actually care about.
Not opt-out. Not conditional. Not “up to”.
LADLE runs on Anthropic’s API tier where customer data is not used to train Anthropic’s public models by default. We ourselves don’t have models — there’s no scenario where your chats train a “LADLE model” because no such model exists.
If Anthropic ever changes their API tier terms in a way that affects this commitment, we would tell you first — before the effective date — and offer cancellation with a pro-rata refund. Full details in the training-data policy article.
One model. Same one for you. On purpose.
Haiku and Sonnet, 200K context, a published routing rule and a manual override.