Worked examples from live market floor prices (data date 2026-11-20)
Formula: cost = input_tokens × floor_input_price + output_tokens × floor_output_price (USD per 1M tokens, per model, from https://321ai.cc/api/public/model-prices). Platform fee is 0 and failures are never billed, so this is the whole bill.
| Workload | Model (live cheapest lane) | Volume | Estimated cost |
|---|---|---|---|
| Claude Code — heavy coding day | claude-sonnet-5 | 10M in / 1M out | $0.5024 |
| Chat app — busy bot, one day | gpt-5.6-sol | 5M in / 1M out | $0.0750 |
| Batch eval — 1,000 prompts | glm-5.3-flash | 0.5M in / 0.5M out | $0.0000 |
Volumes are illustrative; pick your own model and multiply with the formula above. Interactive calculator (Chinese UI): https://321ai.cc/calculator.
Floor price is what the cheapest healthy supplier quotes right now — the marketplace routes to it automatically unless you pin a lane. Cache reads are cheaper than input tokens for models that support prompt caching; the price API exposes the cache-read rate too.
cheap optimizes for price; stable for stability; fast for latency; guaranteed for merchant-supplied lanes. Conditional routes re-pick the best lane within 5 minutes as prices move.