Launch HN: Tokenless (YC S26) – Automatic model switching to save money
usetokenless.com rohaga
Tokenless (YC S26) built a router that swaps AI models mid-conversation to cut token costs. It claims to match Claude Fable 5's performance for half the price by only calling in the expensive model when the task actually needs it.
Based on reporting by usetokenless.com, rohaga — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Rohit, Andrew and Kev built Tokenless because everyone paying for AI agents has hit the same wall: frontier models are great but pricey, and open-source alternatives are cheap but not quite good enough yet. Their pitch is an API gateway that routes agent traffic turn by turn, sending easy steps to cheaper models and only escalating to a smarter one when a task genuinely calls for it.
The idea traces back to Rohit's time as a PhD student at Princeton, where he says he obsessively rationed his academic Cursor account, second-guessing which model to use on every coding task. Around the same time, while recovering from a NeurIPS submission crunch, he stumbled onto a small routing technique that turned out to perform surprisingly well, fast enough that the team was caught off guard by how much a simple idea could accomplish.
The technical trick, according to the founders, is querying several models simultaneously and using their in-progress outputs to decide where to send the next turn, a method they believe hasn't been tried elsewhere. They also built the router to be aware of when a model's cache is warm or cold, so switching between models mid-task doesn't wipe out the caching benefits that usually make repeated calls cheaper.
The team says this approach currently matches the performance of Claude Fable 5 at half the cost, and points to companies like Uber and Salesforce publicly griping about blowing through their annual AI budgets faster than planned as evidence the problem is real and expensive. Support for Kimi K3 and additional GPT models is on the roadmap.
For now, Tokenless is inviting people to sign up and try routing their own agent traffic through the platform, with $20 of free credit to test it out. The founders are mostly asking for feedback on where the router gets things wrong and whether the underlying problem, picking the right model for the right moment, actually resonates with other builders.
My take — AI-written commentary, not fact-checked reporting
Every company burning through AI budgets faster than expected is basically begging for a middleman that promises to make the spending smarter, and that's exactly the gap Tokenless is stepping into. The novelty claim around querying multiple models at once to steer routing decisions is the interesting bit here, more than the usual YC pitch of cheaper-and-just-as-good; if it genuinely holds up outside a demo video, it's a real technical contribution rather than just a billing trick. The real test isn't the benchmark against one named model, it's whether the router quietly screws up on the messy edge cases nobody benchmarks for, which is precisely what the founders are asking Hacker News to go find.
Read more about this at: usetokenless.com