Model Routing Cost Calculator
Routing easy queries to a cheap model looks like free savings — until a misrouted one re-runs on the expensive model and pays twice. This finds where routing wins, and the re-run rate where it stops.
For teams routing easy queries to a cheap model and hard ones to an expensive one — computes the blended savings AND the break-even re-run rate, since misrouted queries pay twice, not a 'routing always saves' assumption.
Model prices from OpenRouter · updated 2026-07-13
01 Your cascade
$0.001400/request
$0.02/request
02 Routing vs all-expensive
Routing saves $13,195/month (59%) — you're under the 94% re-run break-even.
Break-even re-run rate: 94% · you're at 10%
GPT-5.5 costs 16× GPT-4.1 Mini per request — so you can misroute up to 94% before routing stops paying.
Related cost calculators
Prices from OpenRouter, snapshot 2026-07-13, synced daily. Blended = requests × [(1−easy)×expensive + easy×(1−rerun)×cheap + easy×rerun×(cheap+expensive)]. Break-even re-run rate = 1 − cheap/expensive per-request cost. Prompt caching and model quality aren't modeled — dollar cost only. All math runs in your browser.
How the math works
Routing sends the easy 70% of queries to GPT-4.1 Mini ($0.001400/request) and keeps the hard 30% on GPT-5.5 ($0.02/request). Running everything on GPT-5.5 would cost $22,500/month; routing brings it to $9,305. That saves $13,195/month (59%).
The number nobody prices is the re-run penalty. When GPT-4.1 Mini gets a query it can't handle, you re-run it on GPT-5.5 — so that query pays twice (the wasted cheap attempt plus the expensive redo). At a 10% re-run rate that double-charge is $1,673/month of pure waste, quietly eating into the routing savings.
Same baseline, one break-even: routing wins as long as your re-run rate stays under 94% (that's 1 − cheap/expensive cost). Because GPT-5.5 costs 16× what GPT-4.1 Mini does, you can misroute a lot before it stops paying — you're at 10%, with 84 points of headroom.
The lever isn't "route more" — it's routing accuracy. Pushing more traffic to the cheap model only helps if the cheap model actually handles it; every misroute costs a full expensive call on top of the wasted cheap one. A cheaper fallback model, or a better router that mis-sends fewer queries, moves the number more than a bigger cheap share.
Not modeled: prompt caching (applies to both models and roughly cancels out of the comparison), and any quality difference between the models — this page compares dollar cost only. Inference prices sync daily from OpenRouter (updated 2026-07-13); this is an accounting comparison of two catalog models, not a separate price source. All math runs client-side with tested code.
Frequently asked questions
Does routing to GPT-4.1 Mini actually save money for your workload?
At this page's defaults — 70% easy share, 10% re-run rate — yes: $13,195/month (59%) vs running everything on GPT-5.5. It holds as long as your re-run rate stays under 94%.
What's the break-even re-run rate?
94% — computed as 1 − (cheap cost ÷ expensive cost). Below it, routing saves; above it, the re-run double-charges outweigh the savings. Because GPT-5.5 is so much pricier per request here, the break-even is high — you have real headroom for an imperfect router.
Why do misrouted queries cost double?
Because a query sent to GPT-4.1 Mini that fails still billed for the cheap attempt — then you re-run it on GPT-5.5 and pay for that too. So each misroute is cheap + expensive, not just expensive. At 10% that's $1,673/month here.
Should I just route more traffic to the cheap model?
Only if the cheap model handles it. Raising the cheap share helps when re-run stays low, but if pushing more queries down means more of them bounce back, each bounce costs a full expensive call plus the wasted cheap one. Router accuracy beats router aggressiveness.
Are these prices current?
Both models' prices sync daily from OpenRouter (updated 2026-07-13). This mold adds no separate price source — it's an accounting comparison of two live catalog models, so it stays accurate as prices change automatically.
How much can your workload save at best?
The ceiling is routing 100% to GPT-4.1 Mini with zero re-runs: $21,100/month vs all-GPT-5.5. Real savings land below that, scaled by your easy share and re-run rate — tune both sliders to see yours.