A startup called Tokenless, backed by the startup accelerator Y Combinator, has launched a product that sends every AI request to several AI models at the same time, watches their answers as they come in, picks the one clearly doing the best job, and cancels the rest. The customer only pays for the chosen answer (Tokenless). The team behind it includes AI researchers from Google DeepMind, Princeton, and UC Berkeley.

The product is designed to be easy to adopt. It works as a replacement for existing connections to OpenAI and Anthropic's AI services, so companies can switch to Tokenless without changing their code. It comes in at least two tiers: PRO, designed to save money, and MAX, which sends each task to what Tokenless considers the best available model for the highest quality (Tokenless).

Think of it like calling three taxi companies at once, watching which car arrives first, and paying only that one driver. The others get sent home unpaid.

The benchmark numbers the company published are striking for the spread they reveal. On a test called TerminalBench 2.1, Tokenless PRO completed 72% of tasks at $0.32 per task. GPT 5.5 scored 76% at $0.74, Opus 4.8 hit 72% at $2.41, and Tokenless MAX reached 82% at $6.70 (Tokenless). In other words, PRO matched Opus 4.8's completion rate at roughly one-eighth the cost per task. Tokenless also claims PRO is 57% cheaper than the cheapest competing top-tier model at comparable quality on that benchmark.

On another test called LiveCodeBench, the gap narrows but holds. Tokenless PRO scored 89.0% completion at a total cost of $74.78, edging out GPT 5.5 (85.7% at $84.27) and Opus 4.8 (88.9% at $80.01). Tokenless MAX led the group at 90.4% and $85.33 (Tokenless).

The idea of trying multiple models and cancelling the losers is not entirely new, but Tokenless appears to have turned it into a polished product. The cost model works because AI providers charge based on how much text each model generates, and cancelling a model partway through means you only pay for the portion it produced before being stopped.

The router also includes a feature that preserves cached information when switching between models (HN discussion). This matters because a router that thoughtlessly picks the cheapest model each time can accidentally break cached data and end up costing more than just sticking with one provider. Tokenless also claims its router matches the performance of Claude Fable 5 at half the cost (HN discussion), though the TerminalBench and LiveCodeBench figures on the company's site are the more detailed performance disclosures available.

The company was founded by Rohit, alongside co-founders Andrew and Kev (HN discussion). Its website features an interactive savings calculator that references the Ramp AI Index for AI spending data covering June 2025 through June 2026, letting prospective customers estimate savings against their current provider spend (Tokenless). New users receive $20 in free credit on sign-up (HN discussion).

Tokenless plans to add a model called Kimi K3 to its router in the future and has published a technical blog post detailing how the router was built, along with a YouTube demo video (HN discussion).

The broader context here is the rising cost of using AI models. As the companies behind these models have held or raised their prices while the models handle more data per request, the total bill for any organization building on them has become a real budget item. A router that can cut per-task cost by half or more while maintaining the same quality addresses a genuine and growing pain point, especially for systems where a single user action triggers dozens or hundreds of AI calls in sequence.

In my view, there are two things worth watching. The first is speed: sending a request to multiple models at once means the user has to wait for the slowest model to start responding before the system can cancel the others, and the monitoring process adds its own small delay. The second is that Tokenless's benchmark results are self-reported and specific to those tests, and whether they hold up in everyday real-world use is an open question that only independent testing or actual deployment can answer. What the company has shown, at minimum, is that this try-multiple-and-cancel approach can produce cost and quality results that no single AI model achieves on its own.