Model cost calculator

For each model, the per-request input cost depends on whether your token is served from the prefix cache. So the effective input rate is a weighted average:

effective_input_$_per_Mtok =
    cache_hit_rate × cached_input_price
  + (1 − cache_hit_rate) × full_input_price

monthly_cost =
    input_tokens × effective_input_$_per_Mtok / 1_000_000
  + output_tokens × output_price / 1_000_000

If you're running RAG or any workflow with a stable system prompt, your real hit rate is 60-90% — far above the 0% the vendor price table implicitly assumes. That gap is the whole reason a Sonnet workload that "should" cost $X often costs $X/3 in practice. The calculator above models it honestly.

• Prices verified against vendor pricing pages on 2026-05-31. Tap any model row in the table to open its current vendor page and sanity-check before committing budget.
• API list pricing only. Enterprise discounts, committed-spend contracts, and provisioned-throughput pricing all differ.
• Long-context surcharges (some vendors charge a higher rate above 200K or 500K tokens per request) aren't modeled here. If you're routinely sending huge prompts, treat the result as a floor.
• Tool calls, embeddings for retrieval, fine-tuning, and image/audio modalities are billed separately and aren't included.
• Cache hit rate is the % of input tokens served from a prefix cache. Output is never cached.