Siliqun RotorHow Siliqun Rotor compares
Your agent's real work is deciding what to do with each customer request: call the right tool, ask for a detail that's missing, or say no when nothing fits. We measured that decision against two large language models doing the same job on their own. Each side got the same requests, the same tools and the same scoring.
Measured on 3 October 2026.
The short version
- Cost: on your own machine, Siliqun Rotor has no per-request fee at all: it decides on a single CPU core. On our hosted plans, decisions are counted: Pro includes 100,000 a month, and extra decisions cost $1 per 10,000, which is $0.10 per 1,000. A large model bills every request, mostly for resending your tool definitions each time.
- Speed: Siliqun Rotor decides in 64 milliseconds at the median. The models took between 1.3 and 3.7 seconds.
- Safety: on real business requests, Siliqun Rotor never acted on a value it didn't have. It asks for the detail, proposes it from the request when it can, and waits for your confirmation before any change runs.
Business requests
182 requests across 10 business areas: HR, leave, assets, facilities, finance, sales, training and more. Each request has a known right answer: call a tool, ask for a missing detail, or refuse.
| 182 business requests | Siliqun Rotor | Sol 6.1 (OpenAI) |
|---|---|---|
| Right decisions | 154 | 148 |
| Asked for a missing detail instead of guessing (39 needed one) | 39 | 36 |
| Refused when no tool fits (57 had none) | 55 | 55 |
| Wrong actions taken | 2 | 6 |
| Acted on a value it didn't have | 0 | 3 |
| Time per decision, median | same engine as below: 64 ms | 3.7 seconds |
| Time per decision, slowest 5% | same engine as below: 196 ms | 22.8 seconds |
| Cost per 1,000 decisions | none on your machine; $0.10 hosted beyond the plan | $7.18 |
This set is our own test set, built from the kinds of requests our customers' agents receive. We publish it so you can see the behaviour that matters most in production: asking instead of guessing, and never acting without a confirmation.
Public benchmark: picking the right tool
100 requests from the Berkeley Function Calling Leaderboard (BFCL), 25 from each of its four groups: single-tool and real-user tool choice, and two groups where no tool fits and the right answer is to call nothing.
| 100 BFCL requests | Siliqun Rotor | GPT-4o mini | Sol 6.1 |
|---|---|---|---|
| Correct decisions | 85 | 89 | 92 |
| Time per decision, median | 64 ms | 1.35 s | 2.9 s |
| Time per decision, slowest 5% | 196 ms | 3.3 s | 16 s |
| Cost per 1,000 decisions | none on your machine; $0.10 hosted beyond the plan | $0.085 | $1.43 |
On picking a tool from a list, the largest model is a few points ahead. Siliqun Rotor gets close to it for a small share of the time and cost. In production it adds what a benchmark of tool choice can't see: it asks before it guesses, and it waits for your confirmation before any change runs.
How we measured
- Same rule for everyone: a decision is right when it picks the expected tool, or calls nothing when nothing fits. On the business set, a missing detail must be asked for, not invented.
- Siliqun Rotor ran on our production server's own CPU, a single core. Its times are decision times without network.
- The models were called through OpenAI's API from one machine, with the same tool definitions and a one-line instruction. Their times include the network to OpenAI. Sol 6.1 ran with its reasoning on. GPT-4o mini ran at temperature 0.
- Siliqun Rotor's costs are our published plan prices: no fee on your own machine; hosted Pro and Team include a monthly number of decisions, then $1 per 10,000.
- Model costs are each provider's usage multiplied by list prices on the day: GPT-4o mini at $0.15 per million input tokens and $0.60 per million output; Sol 6.1 at $2 and $10.
- Sample sizes are small: 100 BFCL requests and 182 business requests. A difference of a few decisions is within what a different sample could change.
- Questions about the method: write to us.