We don’t build models. We pick them.
A layer that sits between you and everyone you pay has to be boring about its boundaries. Here is what we refuse to do, when we would tell you not to buy this, and every coined word on this site with a definition you can check.
Five things we don’t do
- 01We don’t train or host model weights.
- One exception, declared: a small self-hosted 8B model that does the fingerprint scoring and stands in as a last-resort rung. It exists so your request text can be scored without leaving your side, not so we can say we have a model.
- 02We don’t publish leaderboards.
- No named-vendor quality comparisons, no ranking tables, no “we beat X” charts. Tiers on this site are vendor-neutral codes and they stay that way — a routing layer with a favourite is a routing layer you can’t trust.
- 03We don’t keep your prompts or outputs.
- Fingerprints, routing decisions and metrics by default; bodies only if you turn them on, per endpoint, with a window you set.
- 04We don’t set your quality floor.
- The floor is a product decision. We will tell you what a given floor costs and what the tiers below it score — we will not pick a number for you and call it a default.
- 05We don’t lock you in.
- Ladders are plain text. Export the file and it runs without us; the thresholds, waits and prompt versions are all in it. What you would lose is the shadow history, which is the only thing here that takes time to rebuild.
When not to use us
Four cases where the honest answer is that this will not pay for itself.
- One endpoint, one model, and it works
- Routing is worth something when traffic is mixed. A single endpoint with a narrow difficulty spread has nothing to route.
- Under roughly $3k a month
- At that spend the work of declaring floors and tuning ladders costs more than the ladder saves. Come back when the bill annoys you.
- A floor near the tier ceiling
- When the floor sits just under what the top rung can reach, there is no headroom to take savings from. In this illustrative deployment that is /review, and it only saved 20%.
- A hard tail-latency budget
- Escalation means two hops. If your p99 has no room in it, you will be racing rungs in parallel and paying for both — which is a real option, but it is not the cheap one.
How we charge
Per routing decision. We do not take a margin on your token spend, because a spread would give us a reason to route you upward — and a router with a reason to pick the expensive tier is not a router. No prices on this page: this is an illustrative site, and a made-up price list would be the least useful thing on it.
Glossary
Eleven coined words. Everything else on this site is ordinary English, and where we could use an ordinary word we did.
The mechanism
| Term | Unit | Definition |
|---|---|---|
| Difficulty fingerprint | 5 × [0,100] → D ∈ [0,100] | A five-dimensional vector — reasoning depth, context span, format strictness, factual sensitivity, refusal risk — computed by a small self-hosted classifier in under 10ms, before the request leaves your application. Collapsed into one difficulty score D, which is the router’s only input. |
| Route budget | USD / 1,000 calls | A per-endpoint ceiling, enforced as a hard constraint at decision time rather than reported after the fact. When headroom runs out the ladder shifts down a rung instead of overspending. |
| Fallback ladder | 3–5 rungs | An ordered chain of rungs — each a model tier plus parameters plus a prompt version — with trigger conditions and a max wait. Timeouts, errors, schema failures or sub-floor scores move the request to the next rung. The same ladder runs both ways. |
| Quality floor | judge p10, or schema pass rate % | The minimum acceptable quality each endpoint declares for itself. It is not negotiable: savings are only ever taken from the headroom above it. |
| Shadow replay | 3% of traffic | Live traffic mirrored to candidate tiers offline. Those responses are never returned to users; they only update quality scores and confidence intervals. |
| Swap window | 5,000 samples · 7 days · −0.5 pts | The observation period a candidate clears before it can become a default: at least 5,000 shadow samples on the same endpoint, at least seven consecutive days, and a 95% CI lower bound no more than 0.5 points below the incumbent. Promotion then ramps 5% → 25% → 100%. |
| Brownout mode | state | When a provider throttles or fails, the endpoint drops to lower rungs, caps output length and disables its top tier — degrading the answer instead of returning a 5xx. Named after the grid: dim the lights, don’t cut the city. |
| Cost per resolved call | USD / 1,000 resolved calls | The all-in cost of one accepted answer. The denominator is resolved calls; the numerator includes retries, discarded lower-rung attempts, judge scoring, classifier amortisation and shadow replay. Per-token billing hides all of it. |
The three we watch
| Term | Unit | Definition |
|---|---|---|
| Under-routing | % of traffic | Requests the fingerprint scored more than a band below their actual difficulty. The main source of escalation; the ladder catches them. |
| Over-routing | % of traffic | Requests scored more than a band above their actual difficulty. Nothing catches these — this is the money we wasted. |
| Escalation churn | % of traffic | Requests that move up and down the same ladder repeatedly. A healthy deployment stays under 2%; above that, the floor and the thresholds disagree with each other. |
What we hold ourselves to
- Every percentage carries its basis — what the denominator is, which traffic mix, which window. A percentage without one is decoration.
- Every chart carries its figure number and says it is a model. A chart that cannot say where its numbers came from does not ship.
- No real provider names, no customer names, no logo wall, no testimonials. Tiers are codes, and the reference deployment is fabricated and labelled as such.
- No third-party audit or compliance claims appear anywhere on this site, because there are none to make.
- No superlatives we cannot measure, and no ranking claims about anyone else.
We don’t build models. We pick them.