Tokenless bets per-turn model racing halves AI spend; YC S26 launch hit 71 HN points

Tokenless is a YC Summer 2026 startup selling an API gateway that routes AI-agent traffic dynamically, turn by turn, between different models. Its approach is fan-out rather than rule-based: a single request is sent to several frontier and open models at once, their progress is watched mid-generation, the losers are cancelled and the customer pays only for the model that finished. The company claims the result matches Claude Fable 5-level quality at roughly half the cost.

The founding bet grew out of founder Rohit's own pain as a Princeton PhD student watching his academic Cursor budget disappear while he agonized over which model to use; a routing technique developed during NeurIPS-recovery time 'hit SOTA pretty fast,' and with co-founders Andrew and Kev he turned it into an OpenAI/Anthropic-compatible endpoint. Launching through YC S26 gave the company instant distribution: the Launch HN post on July 29, 2026 drew 71 points and 63 comments.

The Hacker News reception captured both sides of the bet. Practitioners who had built similar internal routers shared cost and caching data, while skeptics called it a metoo OpenRouter clone and asked hard questions about latency, auditability and the danger of quality silently regressing when a cheaper model wins a turn. Automations Cookbook covered the launch the same week, and EarlyTerms classified the term as validating, listing OpenRouter, Ramp Router and Martian as competitors.

Tokenless remains an open bet rather than a proven company: no public pricing, revenue or defensible distribution had emerged by early September 2026, and the strategic question is whether per-turn fan-out routing data and developer trust can become switching costs before model labs or incumbents bundle routing themselves.