The session and the assumptions
A medium refactor in Claude Code or Codex CLI: 150,000 fresh input tokens, 2,000,000 cached input tokens (the context the agent re-sends on every step) and 60,000 output tokens.
Claude is billed Anthropic-style everywhere except Model.sale: cache reads at the cache rate plus cache writes. We assume the 150,000 fresh tokens are written to a 5-minute cache once, at 1.25 times the input rate. GPT figures leave cache writes out. Rates were read on 2026-10-31 from each provider's public page; check them again before you buy, because all of them change.
How each one charges
Official APIs (Anthropic, OpenAI) charge split rates for input, cached input, output and cache writes, billed after use. OpenRouter bills inference at the provider's list price and adds a 5.5% fee when you buy credits on its Standard plan (8% on Business). CometAPI lists Claude and GPT at 20% below the official rates, with the same split structure. Model.sale charges one rate per million charged tokens and counts cached input at 40%; there is no cache-write surcharge (tokens written to the cache count like cached input, so treating them as fresh input below overstates our price). The balance is prepaid.
Claude Opus 5.5
Official ($4 input, $0.20 cached, $20 output, $5 cache write): 0.60 + 0.40 + 1.20 + 0.75 = $2.95. OpenRouter: $2.95 plus the 5.5% credit fee = $3.11. CometAPI ($3.20, $0.16, $16, $4 write): 0.48 + 0.32 + 0.96 + 0.60 = $2.36. Model.sale ($1.40 per million charged tokens, 1,010,000 charged tokens): $1.41.
Model.sale is about 52% below the official price, 55% below OpenRouter and 40% below CometAPI for this session.
Claude Sonnet 5.5
Official ($2, $0.20, $10, $2.50 write): 0.30 + 0.40 + 0.60 + 0.375 = $1.68. OpenRouter: $1.77. CometAPI ($1.60, $0.08, $8, $2 write): 0.24 + 0.16 + 0.48 + 0.30 = $1.18. Model.sale ($0.90): $0.91.
CometAPI's cheap cache reads narrow the gap here: Model.sale is 23% below CometAPI, 46% below the official price and 49% below OpenRouter.
GPT-6.1 Sol and GPT-6 Astra
GPT-6.1 Sol: official ($2, $0.10, $10) $1.10; OpenRouter $1.16; CometAPI ($1.60, $0.08, $8) $0.88; Model.sale ($0.40) $0.40, about 54% below CometAPI.
GPT-6 Astra: official ($10, $1, $50) $6.50; OpenRouter $6.86; CometAPI ($8, $0.80, $40) $5.20; Model.sale ($1.125) $1.14, about 78% below CometAPI. Astra's output and cached-input rates are very high at list price, so a flat rate saves the most here.
When the others are the better choice
OpenRouter has hundreds of models, provider routing and team features that a focused catalog does not; if you need a model Model.sale does not list, it is the simplest place to get it. The official APIs give you the vendor's contract, data terms and newest features on day one. On workloads that are almost entirely cache reads with very little output, split pricing with cheap cache reads can match or beat a flat rate, so test your own mix.
Some smaller resellers list even lower prices on some models. Whatever you compare, price the same session, include cache writes and top-up fees, and check how many requests fail or get cut off: a cheap token that needs a retry is not cheap.
How to check our numbers
Run one representative task on a small balance and read the exact charge of every request in the Model.sale dashboard; the cost calculator on the site takes fresh, cached and output tokens. The live prices are on the pricing page and in GET https://api.model.sale/v1/catalog.