AINEWS
A Y Combinator S26 startup called Tokenless races multiple AI models in parallel, cancels the ones falling behind, and reports 34% blended savings on its own benchmarks.
Tokenless is an LLM router, a thin layer that sits between an app and the model providers and decides, per request, which model to call. Its trick is racing several models in parallel and then canceling the ones falling behind, so a customer pays only for the work the winning model actually did.
The company's product page describes the system as an OpenAI- and Anthropic-compatible endpoint that fans each request out to multiple candidate models, watches them begin to answer, and cuts the rest once one is clearly on track. A "refactor CLI options into an enum" request shown on the homepage was routed to Fable 5. The screen marked "−52% · $0.0101 never billed," a small example of the cost the user does not pay. For an existing app, the practical fit is that the router takes the same API calls a customer's code already sends to OpenAI or Anthropic, so it slots in without code changes.
The numbers Tokenless is anchoring the launch on come from its own "public agentic benchmarks" page. Tokenless Pro solved 40.2% of tasks, ahead of GPT-5.6 at 33.0%, Sol at 32.8%, Claude Opus 5 at 30.9%, Tokenless Ultra Saver at 26.8%, and Claude Fable 5 at 24.5%. Average cost per task on the same page ran $0.57, $1.50, $1.64, $2.25, $2.58, and $3.32. The source layout does not perfectly line up the model legend with the per-task cost column, so the cost-vs-model mapping should be read as approximate rather than a precise 1:1.
The savings figure the company leads with is 34% blended, not the "halve your bill" the homepage hero copy pushes. A calculator on the Tokenless homepage takes an illustrative $40,000 monthly AI spend and lands on a $26,000 bill, with 42% of requests rerouted to a cheaper model and $14,000 a month "off your bill." A "+11.0%/mo" trendline is attributed to the Ramp AI Index cohort; Tokenless describes the figure as illustrative, not a named-customer outcome.
The team is three AI researchers from Google DeepMind, Princeton, and UC Berkeley, and the company is in Y Combinator's S26 batch. The company's launch blog, dated July 29, 2026, argues that "most AI inference is wasted" on routine agentic calls and claims a state-of-the-art result on τ³-Banking using a mix of GPT-5.6, Claude Opus 5, Gemini 3.1 Flash-Lite, and Deepseek V4 Pro. The τ³-Benchmark context was not surfaced in the public hydration of the launch, so the SOTA claim should be read as a company claim rather than an independent result.
The cost story is narrower than the hero copy suggests. "Same quality" is what Tokenless says, and 40.2% on the company's own benchmark is what the company measured. The savings depend on traffic mix, reasoning-token volume, and cache-rate assumptions the customer can edit in the calculator. Independent benchmarks and real production traffic, with their own cache rates and model mixes, may or may not reproduce the numbers. Watch items: a third-party τ³-Banking reproduction, a published cache-rate sensitivity, and any named-customer traffic mix. The launch points to the wider category: a thin routing layer priced on the gap between what a frontier model is asked to do and what a frontier model is actually needed for.