We find savings on your agent workload and prove them.

How it works

1. Baseline. We measure your agent workload’s $ per million tokens with a 95% confidence interval, on your own GPUs, with your real traffic shape. Repeat checks of the current config also measure how much it swings run to run; that noise bound is what every later change is judged against.
2. Test. We try config changes (engine flags, batching, quantization, GPU count, engine version) and keep only the winners that beat the noise with unchanged outputs. A change that is cheaper but alters what your agents produce does not count.
3. Prove. Each change comes with a verified-savings statement from throttle savings, line by line, counted at the conservative end of the 95% interval.

Pricing

• The pilot: 3 months, free. Setup on your stack, the baseline, the tests, and proof for every config change that lowers your cost per token.
• After the pilot: pricing is based on the savings we verify together.
• Verified savings = (baseline $ per million tokens − current $ per million tokens) × your production tokens, measured by Throttle and shown line by line. We count at the conservative end of the 95% interval, so the saving is the smallest one the data supports.

Why the interval matters

Same-server runs swing. Four cost checks of one untouched server came back 27% more expensive, then 11% cheaper, with nothing changed (the write-up). If savings were billed off single runs, you would pay for noise. That is why only changes that clear your measured noise count, and why the statement uses the conservative end.

What we need from you

• An agent workload served from an OpenAI-compatible endpoint (vLLM, SGLang, Ollama, TGI or similar) on GPUs you own or rent, or time on a box with the same model, engine and GPU type.
• What your GPUs cost per hour, so tokens and time can be turned into dollars.
• A representative prompt sample or a description of your agents’ traffic shape, so the baseline reflects your real workload.
• Your monthly production token volume, so verified savings can be put in dollars.

Confidentiality

Throttle runs on your side and only sends requests to your own endpoint. Prompts, responses and keys stay in your environment, and shared summaries leave out URLs, hostnames and anything that looks like a secret. We’re happy to work under your NDA.

What a result looks like

On a rented A100 PCIe 80GB running vLLM 0.16.0 with Qwen2.5-0.5B-Instruct at $1.39/hr, changing max_num_seqs from 1 to 8 took cost from $0.746 to $0.234 per million output tokens (−68.6%), measured with a six-position counterbalanced protocol (B-C-B / C-B-C) so warm-up and drift could not fake the result.

The baseline was deliberately bad: one sequence at a time on a small model. It shows the measurement, not a saving you should expect. The raw artifacts are in the repo, and the full write-up is on the blog. More measured results, including AMD MI300X runs, are in the Cost Index.

Questions