Tiered routing, priced under each provider's own cache and batch rules. Directional estimate for local scenario planning. Actual cost may vary by provider terms, cache behavior, tool usage, regional routing, enterprise discounts, and workload shape.
Scenario and cached rates stay in this browser. Public model metadata refreshes on open and hourly while visible; no API key or scenario upload is required.
Scale inputs apply to every tier. Traffic is then split by each tier's share.
The baseline and tier model menus include the full loaded catalog, grouped by provider with newest additions first. Choosing a model adds its rates automatically.
Tiering means sending each request to the cheapest model that's good enough for it, instead of one model for everything. A tier is one slice of that traffic: pick its model, its entry share, and its workload weight — hover a tier's ? for what each class is for. Tier order (T1 → Tn) is the escalation path: requests that fail a tier are retried one tier up, and the failed attempt still gets billed. Cache and batch controls follow the selected model's pricing rules.
Most requests don't need the most capable model. Tiered routing splits traffic by difficulty: a cheap high-volume tier handles routine work (Pawns), mid-priced workhorses cover everyday tasks (Knights), flagship models take the heavy lifting (Rooks), and frontier models are saved for the hardest, highest-stakes problems (Queen). Requests that a tier can't handle escalate to the next tier up — you pay for the failed attempt and the retry, which is why the escalation rate is one of the levers below. Because prices differ 10–25× between tiers, the traffic mix — not the vendor — is the biggest cost decision you make. The chess names encode the logic: a set gives you eight pawns and one queen, and a healthy routing mix looks the same. The bar at the bottom shows what your mix costs versus sending everything to the Queen.
All providers with published text token prices on OpenRouter. Every listed model is already selectable in the baseline and tier menus. Add a model here to keep its rates in your rate card for comparison. Newest means added to OpenRouter; capability rankings require separate benchmark and workload evaluation.
USD per 1M tokens. Edit any figure to set a manual override — overrides are preserved on refresh until you reset them. "—" means the provider has no published price for that field.
| Model | Status | In | Out | Cache read | Write default | Write 1h | Batch in | Batch out | Ctx | Notes |
|---|
Live prices are OpenRouter's published top-provider metadata, not a quote for a direct provider contract. Cache rates and default write duration depend on the model; a 1-hour rate is used only when explicitly supplied. Unknown cache rates use ordinary input pricing. Batch uses the exact model's published :batch variant, including its own cache prices. Missing batch terms receive no assumed discount. Imported or bundled discount assumptions are labeled estimates. Published context-only pricing overrides apply to the whole request above their threshold. Time-dependent or unrecognized conditions are flagged and excluded. Not modeled: regional-routing uplifts, tools, media, cache storage duration charges, and separately priced reasoning tokens. The reasoning multiplier assumes output-token billing. Cache eligibility and hit rates remain workload assumptions. Escalation: tier order (T1 → Tn) is the escalation path; failed attempts are billed in full at their tier, escalated requests are re-billed at the destination tier's token profile with the payload overhead applied to input, and escalated volume is not batch-eligible. The all-queen baseline assumes a single model handles each entry request once, with no escalation.
Pins store the workload (drivers, tiers, shares) and are re-priced against the live rate card, so a price edit flows through every pinned scenario.
| Scenario | $/month | Δ vs current | Δ% | $/1K req |
|---|
Deterministic one-at-a-time sensitivity on the current scenario: percentage drivers are perturbed +10%, rate drivers +10 points, and the mix test moves 10 points of traffic from the cheapest tier per request to the priciest. Ranked by absolute impact.