A cheaper model that drops quality is debt, not a saving.

Both routes run on your tasks against the check you ratified. The cheaper one ships only when it clears the bar and costs less; otherwise the verdict is hold. The bar is ours today. What an outcome is worth to you should set it, and we will never guess that number.

Cheaper by 69%, and read task by task: the saving forfeits 20 outcomes the incumbent lands, and recovers 4.Figure 1

116 tasks from a live tagging workflow, both arms run in full. Each cell is one task: the gate reads the pairing, not the averages.

88 both arms landed20 only the incumbent (what the saving forfeits)4 only the candidate4 neither

Held. The cheaper arm keeps 88 of the incumbent’s outcomes, drops 20, and finds 4 of its own: a net 16 forfeited to save $0.0372. Break‑even: $0.0023 an outcome. Value a finished outcome above that and refusing was the right trade.

n = 116 paired taskslanded 108 → 92net −16 outcomesis_simulated=falsecaptured 2026-08-19

Source: the committed artifact provenance/autoroute_defended_savings_frac.json. Every cell derives from its per-call rows at build time; none is typed.

The same gate, three more rulings.

One promotion, one rival house, one card that is not ours. It reads each the way it read Figure 1.

The gate says yes when the evidence does: −64% with quality up, and promoted: the paired test over 115 shared tasks holding.Figure 2

The same model at a lower thinking budget, 115 trials on a live reasoning workflow. The first route the gate has passed in a paired eval.

A. What each budget cost, all 115 tasks
gemini-2.5-flash · full thinking budget$0.0970
gemini-2.5-flash · thinking budget 128$0.0352
B. What each budget landed
full thinking budget89%
thinking budget 12894%
the drawn rule is the full budget’s own rate: the lower budget cleared it

Promoted. The bar is parity with the incumbent, not an absolute rate: the lower budget was not worse task for task, so the gate passed it on 2026-09-19.

n = 115 taskscost −64%quality 89% → 94%is_simulated=falsecaptured 2026-09-19

Source: the committed artifact provenance/reasoning_effort_savings_frac.json. Both panels derive from it at build time.

Under these verdicts sits one piece of algebra.

Every mix of models that lands the same outcome is a curve, and the cheapest mix that still lands it is where your budget line touches it. The gate keeps you on the curve. A point below it costs less and delivers less, and no per-token dashboard can tell the two apart.

isoquant · cost mix vs. the outcome it buys
small-model callsfrontier callsbudget too smallsame outcomecheapest mix that holds paritycheaper, and short of the outcome
anywhere on the curvethe same outcome
where the budget line touches it4 small : 1 frontier
below the curvecheaper, and short
Buy the mix, not the model. Every point on the curve buys the same outcome; the cheapest is where the budget line touches it. Below the curve you pay less and get less, and that line is our parity gate. Drawn from the algebra; your own mix is measured.

Cheaper, and refused. Then their bill.

A cheaper tier of the incumbent’s own vendor, run in full. Beside it, seven models on someone else’s bench.

The tier cut: 69% cheaper, and refused because quality fell 13.8 pts.Figure 1

A live tagging workflow, both arms run in full.

A. What each arm cost, all 116 tasks
claude-sonnet-4-6$0.0535
claude-haiku-4-5$0.0164
B. What each arm landed
claude-sonnet-4-693%
claude-haiku-4-579%

Refused. Saved $0.0372, forfeited 16 outcomes. Break‑even: $0.0023 an outcome; worth more to you, and refusing was right. We never assume your number.

n = 116 paired tasksis_simulated=false

Source: provenance/autoroute_defended_savings_frac.json, re-derived from its per-call rows on every build.

7.0× the bill, for pass rates inside a single point of each other.Figure 2

All seven on one suite.

We thought token efficiency was the driver of cost, but really it’s just the cost per token. Given the constant shifts in both variables, we have to just track dollars.

Matt MacInnis, President and Chief Product Officer, Rippling
n = 7 models · ~2,100 scored runs eachtheir bench, their data, not our meterpublished

Source: “RipplingBench: A comparison of open‑weight and frontier models in real‑world use cases”, Rippling, 29 July 2026.

Rippling is not a customer and has not evaluated Margin.

The gate’s whole record

A test fails the build if the numbers slip.

Five checks run on every commit and fail closed.

provenance check

A headline saving reproduces from a committed is_simulated=false artifact, or fails closed.

auto-route adversarial check

Promote only past a high parity bar and cheaper; revert when parity slips.

regression gate

Cheaper by dropping the pass rate fails, whatever the saving.

pricing accuracy check

Per-token math checked against published prices, cached reads included.

outcome integrity check

An outcome is positively determined; an ambiguous signal raises, never passes.

What they let through on our own agent runs: 280 of 306 outcomes passed at 0.003722 USD each, is_simulated=false, captured 2026-09-03. The meter’s full read

Down a chain, quality compounds.

Cost adds. Quality multiplies. Five steps, each holding our 0.97 bar, deliver 0.859 to your user. Margin advises per step, at the bar your chain’s length demands.

Every step holding our 0.97 bar, compounded
Steps in the chainWhat the user actually getsevery step passes at 0.97What each step must holdto deliver 0.97 end‑to‑end
1 (one call)0.9700.970
30.9130.990
50.8590.994
120.6940.997
This is the identity 0.97n, not a measurement of your pipeline. It treats each step as independent: a validating step that catches an upstream error does better, and steps that fail together do worse. Either way the gap widens with every step, and a per-call gate cannot see it, because a single call never knows how long its chain is.

Watch the loop run on real spend.

The console runs it on open-source agents we metered ourselves, every number tagged real or sample.