A Y Combinator-backed startup called Tokenless has launched a router that sends each AI model request to several models at the same time, watches their responses as they come in, picks the one clearly producing the best answer, and cancels the others so the customer only pays for the chosen output (Tokenless). The founding team includes AI researchers from Google DeepMind, Princeton, and UC Berkeley.

The router works as a drop-in replacement for existing API calls. It exposes endpoints compatible with both OpenAI and Anthropic's APIs, so development teams can redirect their existing code to Tokenless without rewriting anything. The product comes in at least two tiers: PRO, tuned for cost savings, and MAX, which routes each task to what Tokenless considers the best available model for maximum quality (Tokenless).

The benchmark numbers the company published are notable for the spread they reveal. On TerminalBench 2.1, Tokenless PRO achieved 72% task completion 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). Put another way, 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 frontier model at comparable quality on that benchmark.

On 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 fan-out-and-cancel approach is not entirely new as a concept. Speculative decoding (where a smaller model drafts tokens that a larger model then verifies or rejects) and model cascading (where requests are routed from cheaper to more expensive models only if needed) have explored adjacent territory. What Tokenless appears to operationalize is a production-grade version of parallel speculative execution at the API layer. The idea is straightforward: try multiple frontier models simultaneously, pick the winner, kill the rest. The cost model works because providers charge per token generated, and cancelling mid-generation limits the tokens you pay for on the discarded paths.

A cache-aware routing algorithm preserves prompt cache state when switching between models, so routing decisions do not invalidate cached prefixes (HN discussion). This matters because cache invalidation is one of the hidden cost drivers in multi-model setups. A naive router that picks the cheapest model per request can destroy cache hit rates and end up more expensive than sticking with a single 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 landing page features an interactive savings calculator that references the Ramp AI Index for AI spend-per-employee 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 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 trajectory of LLM inference economics. As frontier model providers have held or raised prices while token throughput and context windows have expanded, aggregate API spend has become a material line item for any engineering organization building on these models. A router that can transparently cut per-task cost by half or more at parity quality addresses a real and growing pain point, particularly for agentic workloads where a single user-initiated task may chain dozens or hundreds of model calls.

In my view, there are two things worth watching. The first is latency: dispatching to multiple models simultaneously means the user waits for the slowest of the initial fan-out before cancellation can occur, and the monitoring overhead adds its own delay. The second is that Tokenless's benchmark claims are self-reported and benchmark-specific, and the degree to which TerminalBench and LiveCodeBench results generalize to production traffic patterns is an open question that only independent evaluation or production deployment can answer. What the company has shown, at minimum, is that the fan-out-and-cancel model can produce cost-quality tradeoff curves that no single frontier model achieves alone.