Tokenless API Gateway Routes AI Traffic Between Models to Cut Spend in Half

Tokenless (YC S26) is a drop-in API gateway that cuts AI inference spend by routing each turn of an agent conversation to the cheapest adequate model. The founders (Rohit, Andrew, Kev) came from Princeton, Google DeepMind, and UC Berkeley, and they claim their router matches Claude Fable 5 performance at half the cost.

How It Works

Tokenless fans out a request to multiple models simultaneously and watches their progress. Once one model is clearly on track, it cancels the others. You only pay for the winning model's usage. The technique is novel: the router queries multiple models in parallel and uses their intermediate outputs to make the routing decision. The team also notes that switching models doesn't destroy the cache if the routing algorithm is aware of cache hot/cold states.

Benchmarks

On the τ³-Banking, Terminal-Bench 2.1, and DeepSWE 1.1 agentic benchmarks, Tokenless Pro achieves a 40.2% solve rate at $0.57/task, vs Claude Fable 5 at 24.5% solve rate and $3.32/task. The Ultra Saver mode routes more aggressively and gets 30.9% solve rate at $2.25/task. The numbers are presented as "measured, not marketed" — they claim to beat all frontier models on cost-adjusted quality.

Getting Started

Tokenless exposes an OpenAI- and Anthropic-compatible endpoint. You point your existing agent at their endpoint, and they handle the routing. New users get $20 in free credit. The team plans to add Kimi K3, GPT efforts, and more models to the router.

Cost Savings Projection

Tokenless provides a calculator: for a team spending $40K/month on LLMs, they project a new bill of $26K/month (a 34% savings, $14K/month off), totaling $344K saved over the next year assuming 11% monthly spend growth.