Tokenless (YC S26): Intelligent Model Routing to Slash Inference Costs

Tokenless is a sophisticated router designed to drastically reduce your LLM expenditures—potentially cutting your inference bill by 50%. It functions as a drop-in replacement for your existing API calls, ensuring that every request is automatically directed to the most efficient model for the job.

How It Works: The "Fan-Out" Mechanism

Rather than guessing which model is best, Tokenless employs a unique strategy: it broadcasts your request to a cluster of models simultaneously and monitors their "thought process" in real-time.

As soon as the system detects that a specific model is clearly on the right path to a correct answer, Tokenless selects that model and immediately terminates the others. This ensures you only pay for the tokens absolutely necessary to achieve the result.

"Measured, not marketed."

---

Performance Benchmarks

Tokenless has been tested against public agentic benchmarks—specifically $\tau^3$-Banking Terminal-Bench 2.1 and DeepSWE 1.1—comparing its performance against the top published runs of frontier models.

---

Financial Impact & Projections

For a company spending $40,000 per month on LLMs, the transition to Tokenless can be transformative.

Monthly Cost Breakdown:

• Current Monthly Bill: $40,000
• New Monthly Bill: $26,000
• Total Monthly Savings: $14,000

Using the formula for blended savings: $\text{Blended Savings Rate} = \frac{\text{Current Bill} - \text{New Bill}}{\text{Current Bill}} = \frac{40,000 - 26,000}{40,000} = 34\%$

Annual Outlook: Based on the Ramp AI Index (which tracks AI spend per employee from June 2025 to June 2026), the current trajectory for AI spending is increasing by approximately $+11.0\%/\text{mo}$. By implementing Tokenless now, a company could see $344,000 saved over the next year.

Key Metrics:

• Requests Rerouted: 42%
• Blended Savings: 34%

---

Technical Implementation & Pedigree

The platform was engineered by a team of AI researchers hailing from Google DeepMind, Princeton, and UC Berkeley.

To integrate Tokenless, you don't need a massive overhaul. You can simply refactor your CLI options into an enum and update your routing logic.

# Example: Swapping to Tokenless
# OLD: response = frontier_model.generate(prompt)
# NEW:
import tokenless
response = tokenless.route(prompt, strategy="Pro") 

Ready to optimize?

• Book a personalized demo
• Run the numbers on your actual traffic
• Swap two lines of code to test the system