Loading benchmark data…
EyesInAI·Loading live benchmark data
The router fits high-volume, repeatable work billed per token: bots, back-office jobs and agents. It does not fit a few staff chatting on flat-rate plans. Below are nine places AI spend leaks, each tagged with how far we have actually got.
No saving here has been measured on a client’s real traffic, and we have no savings receipt yet.
Many support questions are simple: order status, return policy, “do you rent X”. On our synthetic acme-support test, Haiku cleared our quality bar on quick questions, FAQs and refund requests; on open-ended analysis problems only Opus cleared it (Opus 15/15, Sonnet 14/15, Haiku 12/15).
A route pack sends each type of request to the model that cleared the bar for it. When no model clears the bar, the route keeps serving the best fallback it has, marked degraded and flagged for a person to review.
Examples: tagging tickets, scoring leads, pulling fields from invoices, summarising calls.
In our tests on DLR’s internal lookup routes, groq/gpt-oss-120b, a much cheaper open model, passed every lookup test it was given on pricing lookup (6/6), invoice lookup, open invoices and account summaries.
That model is still blocked on all four routes until a person reviews it (the routes serve Haiku meanwhile), and lookup tests say nothing yet about tagging, scoring, extraction or summaries. Each of those needs its own test set.
An automation retries and re-reads its whole context all night. Today the router can cap the size of each request (input and output) per client, but no client has those caps switched on.
A daily or monthly spend budget per client is not built. A loop of normal-sized requests would still run. This is a gap worth building.
The router caches its routing decision, which saves the classification step, but the chosen model still answers every time. Answer caching is not built.
It would only suit answers that don’t change per customer (store hours, a standard policy), never invoices or account data. Another gap worth building.
On Claude models, “thinking” is billed as output tokens, the most expensive kind. In our 2026-10-14 test (45 synthetic support problems, run twice: 90 answers per model per setting), the lowest thinking setting scored Haiku 85 vs 86 at default, Sonnet 90 vs 90, Opus 89 vs 90.
On automatic routes the gateway applies a low default, and it rewrites some caller settings to the form each model accepts.
Providers charge much less for context they have already seen, but only on the same model with a matching prefix. On automatic Claude conversations, the gateway normally holds the chosen model to help preserve prompt-cache reuse; an upgrade or the pin expiring can change it.
In a live test chat (billed per token through an API key), later turns were read from cache.
Route packs are recompiled daily from a client’s latest benchmark run, so a model that passes that client’s tests and is healthy can take the route with no code change. A new model is not tested automatically; it has to be added to the client’s benchmark.
The compile job once went more than three weeks without a scheduled run, and a partial failure can still look like success.
Providers discount batch jobs that don’t need an immediate answer.
The two gaps that matter most for “spend is out of control” are a per-client spend budget (case 3) and answer caching for stable answers (case 4). Both are unbuilt today. Savings are only ever claimed on traffic billed per token: a Pro or Max seat does not get cheaper by routing.
Before we suggest any change, we measure it. From your numbers we rank where the money is going. For any fix we then agree to test, you get its test results and an estimated saving, and nothing is switched over until it has been checked on your traffic.
Please send what you can of the following. Summaries are fine, and you can black out account numbers, names and addresses.
Never send API keys, passwords or customer records. We don’t need them.
Sample requests for your 2–3 biggest uses, so we can test cheaper models on your real work. We’ll agree a secure way to send them first. Made-up examples that look like your real ones are a good start. Note that test samples are sent to AI model providers to be answered, so leave out attachments and anything confidential.
From the numbers alone, you’ll get back where the money is going and which fixes look worth testing. If we go on to test a fix together, you’ll also get how it scored on test examples and an estimated net saving with its assumptions written down. If the honest answer is “you’re already doing this well”, we’ll say so.
Tell us what you run and roughly what it costs. Don’t paste bills, exports or keys into this form.
Prepared 2026-10-17.