// ROUTING BENCHMARKS
5 ops workflows. 5 models. 77% average saving at the quality bar.
The cost/quality numbers we route by, published. Every row is measured against the workflow's own acceptance rubric — not a generic eval. Updated as models ship. Last update: 2026-06-05.
= routed pick: cheapest model clearing the workflow's quality bar
| Method | |||||
|---|---|---|---|---|---|
| Claims-status summarization | frontier-1 | 18.00 | 97 | 2026-05-28 | 120 redacted claims, rubric-scored by blinded reviewer |
| Claims-status summarization | mid-tier | 6.50 | 95 | 2026-05-28 | 120 redacted claims, rubric-scored by blinded reviewer |
| Claims-status summarization | small | 3.20 | 91 | 2026-05-28 | 120 redacted claims, rubric-scored by blinded reviewer |
| Claims-status summarization | nano | 0.90 | 84 | 2026-05-28 | 120 redacted claims, rubric-scored by blinded reviewer |
| Claims-status summarization | local-7b | 0.40 | 76 | 2026-05-28 | 120 redacted claims, rubric-scored by blinded reviewer |
| Document intake & classification | frontier-1 | 19.50 | 98 | 2026-06-02 | 200 mixed-format documents, field-level accuracy vs ground truth |
| Document intake & classification | mid-tier | 7.10 | 96 | 2026-06-02 | 200 mixed-format documents, field-level accuracy vs ground truth |
| Document intake & classification | small | 3.50 | 90 | 2026-06-02 | 200 mixed-format documents, field-level accuracy vs ground truth |
| Document intake & classification | nano | 1.00 | 78 | 2026-06-02 | 200 mixed-format documents, field-level accuracy vs ground truth |
| Document intake & classification | local-7b | 0.50 | 70 | 2026-06-02 | 200 mixed-format documents, field-level accuracy vs ground truth |
| Eligibility checks | frontier-1 | 18.00 | 99 | 2026-06-05 | 150 plan-document checks, pass/fail vs adjudicator decisions |
| Eligibility checks | mid-tier | 6.50 | 96 | 2026-06-05 | 150 plan-document checks, pass/fail vs adjudicator decisions |
| Eligibility checks | small | 3.20 | 88 | 2026-06-05 | 150 plan-document checks, pass/fail vs adjudicator decisions |
| Eligibility checks | nano | 0.90 | 79 | 2026-06-05 | 150 plan-document checks, pass/fail vs adjudicator decisions |
| Eligibility checks | local-7b | 0.40 | 68 | 2026-06-05 | 150 plan-document checks, pass/fail vs adjudicator decisions |
| Invoice line-item extraction | frontier-1 | 17.00 | 98 | 2026-05-30 | 180 invoices across 40 vendors, line-level accuracy vs AP records |
| Invoice line-item extraction | mid-tier | 6.20 | 95 | 2026-05-30 | 180 invoices across 40 vendors, line-level accuracy vs AP records |
| Invoice line-item extraction | small | 3.00 | 93 | 2026-05-30 | 180 invoices across 40 vendors, line-level accuracy vs AP records |
| Invoice line-item extraction | nano | 0.90 | 85 | 2026-05-30 | 180 invoices across 40 vendors, line-level accuracy vs AP records |
| Invoice line-item extraction | local-7b | 0.40 | 74 | 2026-05-30 | 180 invoices across 40 vendors, line-level accuracy vs AP records |
| Support ticket triage | frontier-1 | 16.00 | 97 | 2026-05-21 | 300 historical tickets, routing accuracy vs senior-agent labels |
| Support ticket triage | mid-tier | 5.80 | 95 | 2026-05-21 | 300 historical tickets, routing accuracy vs senior-agent labels |
| Support ticket triage | small | 2.90 | 92 | 2026-05-21 | 300 historical tickets, routing accuracy vs senior-agent labels |
| Support ticket triage | nano | 0.80 | 89 | 2026-05-21 | 300 historical tickets, routing accuracy vs senior-agent labels |
| Support ticket triage | local-7b | 0.40 | 80 | 2026-05-21 | 300 historical tickets, routing accuracy vs senior-agent labels |
// the frontier
Where cost meets the quality bar
Pick a workflow. The dashed line is its quality bar; the circled point is the model we'd route to — the cheapest one above the line.
// run your numbers
Put your volume against these numbers
What's your baseline?
- Manual cost today
- $6,400/mo
- Routed pick: small (quality 91)
- $4.80/mo
- Human review (15% of tasks)
- $960/mo
- Estimated saving
- $5,435/mo
Quality bar for claims-status summarization: 90. The routed pick is the cheapest model that clears it in our benchmark.
These are estimates from our benchmark data. The assessment gives you your numbers.
Book the assessment// methodology
Each row is scored on a fixed task set against an acceptance rubric defined for that workflow — sample sizes and scoring noted per row. Costs are measured end-to-end per 1,000 tasks at typical task size, not per-token list prices. We re-run the set when a relevant model ships and date-stamp every row.
Want this run on your workflow?
The Routing Assessment does exactly this against your real tasks and your acceptance bar — $4,500, one week, fully credited if you continue.