The session we price
A medium refactoring task in Claude Code or Codex CLI: 150,000 fresh input tokens (your prompts, newly read files, tool results), 2,000,000 cached input tokens (the same context re-sent on every agent step) and 60,000 output tokens. Agent sessions look like this: most input is cached, output is small.
Official numbers below use the dated rates on our comparison page (checked 2026-10-01) and leave out cache writes, which Anthropic bills on top (a 5-minute cache write costs 1.25 times the input rate). Model.sale numbers use the live rates on 2026-10-31.
Official API prices for that session
Claude Opus 5.5 ($4 input, $0.20 cached, $20 output per 1M): 0.15 × 4 + 2 × 0.20 + 0.06 × 20 = $2.20. Claude Sonnet 5.5 ($2, $0.20, $10): $1.30.
GPT-6 Astra ($10, $1, $50): 0.15 × 10 + 2 × 1 + 0.06 × 50 = $6.50. GPT-6.1 Sol ($2, $0.10, $10): $1.10. Astra is the expensive one: its output and cached-input rates are five to ten times the others.
The same session on Model.sale
Model.sale charges one rate per million charged tokens, with cached input counted at 40%: charged tokens = 150,000 + 60,000 + 0.4 × 2,000,000 = 1,010,000.
At the live rates that is $1.41 on claude-opus-5-5 ($1.40), $0.91 on claude-sonnet-5-5 ($0.90), $1.14 on gpt-6-astra ($1.125) and $0.40 on gpt-6.1-sol ($0.40). The saving against the official bill is about 36% for Opus 5.5, 30% for Sonnet 5.5, 63% for GPT-6.1 Sol and 82% for GPT-6 Astra, before Anthropic's cache-write charges, which only widen the Claude gap.
Where the difference comes from
Official tariffs make output and fresh input expensive and cached input cheap. Model.sale's flat rate is the opposite trade: output and fresh input are much cheaper, cached input is not. So the saving is largest on models with expensive output (GPT-6 Astra) and on sessions that write a lot, and smallest on very cache-heavy Claude work.
Be careful with small models. On a cache-heavy session GPT-6 Luna costs about $0.065 at the official rates and $0.08 here, so the official API is cheaper for that workload. Put your own token mix into the cost calculator rather than trusting any headline discount, including ours.
Which model to pick for coding
Start with Sonnet 5.5 in Claude Code or GPT-6.1 Sol in Codex CLI: both handle multi-file changes, and on Model.sale a medium session costs under a dollar. Move a task to Opus 5.5 when it fails twice or needs real design judgment. Use GPT-6 Astra for the hardest reasoning problems, where one correct answer replaces several attempts.
None of this is a quality ranking. Run five tasks from your own backlog on two models, count attempts until the tests pass, and compare the summed charges in the dashboard. Cost per finished task is the only number that matters.
Use both families with one key
One Model.sale key works for Claude and GPT. OpenAI-compatible clients (Codex CLI, Cursor, Aider, the OpenAI SDKs) use the base URL https://api.model.sale/v1. Claude Code and the Anthropic SDKs use https://api.model.sale with ANTHROPIC_BASE_URL and ANTHROPIC_AUTH_TOKEN. The key can carry a daily spend limit and a model allowlist, and the prepaid balance stops at zero.
Prices on this page are dated. The live table is on the pricing page, and GET /v1/models lists what your key can call right now.