Tokenless races multiple models on every request, cancels the losers, claims 52% inference savings.

Tokenless (YC S26) launched on HN with 66 points as a drop-in replacement for OpenAI and Anthropic API calls. Rather than routing by heuristic, it fans each request out to a group of models, watches them think, then selects the one clearly on track and cancels the rest so you pay only for the winner. 52% savings in the featured example, calculator projecting 34% on a $40K monthly spend. Benchmarks reference Claude Opus, Claude Fable, GPT-5.6, and Gemini. Pricing is not published, which tells you something.