Glonce

Tokenless launches AI model router that halves inference bills

toolsThe Glonce editorsHacker News🔥61Glonce rating61Impact60Novelty60Relevance90Surprise55Players70Credibility80

Tokenless, a Y Combinator (S26 batch) startup, launched on Hacker News with a product it calls a drop-in router for LLM API calls. The pitch: point existing OpenAI- or Anthropic-format API calls at Tokenless instead of directly at a model provider, and the service automatically picks a cheaper model when a cheaper model can do the job, cutting the bill without a rewrite.

The routing mechanism described on the launch page fans a single request out to a group of models at once and watches their output as it is generated. Once one model is judged to be clearly on track to a correct answer, Tokenless keeps that model's output and cancels the rest, so the customer is billed only for the model that was kept, not for the models that were tried and dropped. The site shows a worked example: a request to refactor CLI options into an enum was routed to a model called Fable 5, with a note that the call saved 52 percent: $0.0101 that was never billed.

To back the same-quality-half-the-cost claim, Tokenless publishes a table comparing two of its own offerings, Tokenless Pro and Tokenless Ultra Saver, against GPT-5.6 Sol, Claude Opus 5, Claude Fable 5 and Gemini 3.6 Flash on solve rate and average cost per task, measured on unspecified public agentic benchmarks against what the company describes as the best published run of each frontier model. By its own numbers, Tokenless Pro posts a 40.2 percent solve rate at $2.25 per task, ahead of GPT-5.6 Sol (33.0 percent at $2.58) and Claude Opus 5 (32.8 percent at $1.50); Tokenless Ultra Saver posts 30.9 percent at $1.64 per task. Claude Fable 5 and Gemini 3.6 Flash trail on solve rate at 26.8 percent and 24.5 percent respectively, though Gemini 3.6 Flash is the cheapest option overall at $0.57 per task.

The launch page also carries an interactive savings calculator: a business with a stated $40,000-a-month LLM bill is shown a projected new bill of about $26,000 a month under Tokenless, a 34 percent blended savings rate with 42 percent of requests rerouted to cheaper models. The company notes these estimates use published token prices, including cache rates, plus routing assumptions a visitor can edit, and that actual savings depend on the customer's own traffic; the trend line behind the calculator cites the Ramp AI Index, tracking AI spend per employee from June 2025 to June 2026, extrapolated at its trailing 12-month growth rate of 11.0 percent a month.

Tokenless says its team is made up of AI researchers from Google DeepMind, Princeton, and UC Berkeley, and that it is backed by Y Combinator. The Hacker News Launch post, submitted by user rohaga, had gathered 56 points and 50 comments roughly 13 hours after posting.

Key facts

  • Tokenless, a Y Combinator S26 startup, launched on Hacker News with an inference router that claims to cut LLM API bills by about half while matching quality.
  • It exposes OpenAI- and Anthropic-compatible endpoints; it fans a request out to several models, watches which one is already on track, cancels the rest, and bills only for the winner.
  • On its own benchmark table, Tokenless Pro claims a 40.2 percent solve rate on public agentic benchmarks at $2.25 per task, versus 33.0 percent at $2.58 for GPT-5.6 Sol and 32.8 percent at $1.50 for Claude Opus 5.
  • A savings calculator projects a customer spending $40,000 a month on LLM APIs down to about $26,000 a month, a 34 percent blended savings rate with 42 percent of requests rerouted.
  • The company says its founders are AI researchers from Google DeepMind, Princeton and UC Berkeley; the Launch HN post drew 56 points and 50 comments within about 13 hours.

Why it matters

LLM inference bills for API-based apps and agents have become a significant operating cost, and Tokenless targets that line item directly by automatically downgrading a request to a cheaper model whenever a cheaper model can already solve it, without asking developers to rewrite their integration. It turns which model to call from a manual, per-feature decision into an automatic, per-request one, done by racing several models against each other rather than by pre-classifying the prompt.

Who it affects

Teams calling GPT, Claude or Gemini APIs directly, especially those running agentic workloads made of many small calls, where Tokenless claims to reroute 42 percent of requests. The OpenAI- and Anthropic-compatible endpoints matter specifically to developers already using those SDK formats, since it lets them change provider by changing an endpoint rather than a codebase.

How to use it

Sign up and repoint API calls at Tokenless's OpenAI- or Anthropic-compatible endpoint, described on the site as swapping two lines of code, or book a demo where the company runs the projected savings against the customer's actual traffic. The site does not publish a subscription price; savings are shown per request, as in the $0.0101 saved on the CLI-refactor example, and through the calculator, which is built on the customer's own stated usage rather than a flat rate.

How solid is it

All performance and savings figures come from Tokenless's own landing page, not from an independent benchmark. The solve-rate and cost table compares Tokenless's two products against the best published run of each frontier model, without naming the specific agentic benchmarks used or the number of tasks, and the savings calculator is presented as an editable estimate built on assumptions a visitor can change. The one outside data point is Hacker News reception itself, 56 points and 50 comments about 13 hours after the Launch HN post, which is real engagement but not verification of the underlying numbers.

Risks and caveats

The fan-out-and-cancel routing means every request briefly runs on multiple models before one is picked, and the launch page does not explain how on track is judged or what that costs in latency. Comparing a company's own product against the best published runs of competitors is a favorable framing by construction. The savings calculator's growth trend leans on a third-party spend index extrapolated forward, not on Tokenless's own customer data, and none of the benchmark or savings figures are independently verified in the source.