{"ok":true,"engine":"aicost.gateway-stage-advisor","inputs":{"monthly_spend_usd":6000,"workload_count":3,"attestation":"none","has_eval_data":false,"traffic_mix":"uniform","router":"openrouter","eligible_for_cheaper_pct":40,"expected_saving_pct":60,"engineering_hours_month":4,"hourly_rate_usd":145},"result":{"ok":true,"verdict":"POLICY ROUTING","verdictReason":"More than one workload or more than trivial spend. A gateway gives spend attribution and a budget cap, which is worth having before any routing question.","decision":"Assign models per workload, and prove the saving before adding a router.","decisionDetail":"Most businesses belong at Stage 2 or 3. A fixed model policy captures the large majority of the practical value without evaluating every prompt, and it is far easier to defend to an auditor.","stage":3,"stageName":"Policy routing","monthlySpendUsd":6000,"eligibleSpendUsd":2400,"grossSavingUsd":1440,"routerFeeUsd":330,"operatingCostUsd":580,"netMonthlyUsd":530,"breakEvenSpendUsd":3135.14,"breakEvenNote":"Below about $3,135 a month, at these assumptions, the layer costs more than it saves.","routerName":"OpenRouter","paysForItself":true,"reasons":["More than one workload or more than trivial spend. A gateway gives spend attribution and a budget cap, which is worth having before any routing question.","Three or more distinct workloads. Assigning a model per workload captures most of the available saving, is explainable to an auditor, and needs no per-request evaluation.","NOT Stage 4, because prompt difficulty is uniform, so there is little for a router to decide; there is no production evaluation data, so nobody can prove a cheaper model preserved quality."],"gives":["most of the cost benefit","explainable and auditable choices"],"costs":["someone has to maintain the assignments"],"stageFourBlockedBy":["mixed","eval"],"memo":"This client sits at Stage 3, Policy routing. More than one workload or more than trivial spend. A gateway gives spend attribution and a budget cap, which is worth having before any routing question. At $6,000 a month with 40 percent of traffic eligible for a cheaper model, the decision layer breaks even at about $3,135 a month. They are above that line, so it pays for itself at the stated assumptions.","questions":["What share of this client's traffic could genuinely run on a cheaper model without a quality complaint?","Who maintains the model assignments once they exist, and is that time billable?","Does the client hold an attestation that requires an enforced allowlist regardless of spend?","Is prompt difficulty actually varied, or is it uniform enough that a router has nothing to decide?","If spend halved next quarter, would this layer still pay for itself?"],"assumptions":["40 percent of traffic eligible for a cheaper model *","60 percent saving on that eligible traffic *","4 engineering hours a month to operate the layer *","Engineering time valued at $145 an hour *","Router fees captured 2026-07-26 and re-verified before quoting."],"caveat":"Stage 2 is a governance decision and does not depend on spend. Stages 3 and 4 are economic and do.","inputs":{"monthly_spend_usd":6000,"workload_count":3,"attestation":"none","has_eval_data":false,"traffic_mix":"uniform","router":"openrouter","eligible_for_cheaper_pct":40,"expected_saving_pct":60,"engineering_hours_month":4,"hourly_rate_usd":145}}}