TokenAtlasPricing by model
Sonnet 3.5: $3/$15. Opus: $15/$75. Haiku: $0.25/$1.25. Per 1M tokens. Output is 5× input — long generations dominate cost.
Prompt caching: the 90% discount
Cache long system prompts and reference docs. Cached reads are 90% cheaper than fresh input. On a 50K-token system prompt across 1K requests/day, savings are >$1,000/mo.
Batch API
50% off everything, async with 24h SLA. Perfect for bulk classification, periodic summarization, and offline RAG ingestion.
When Opus is worth it
Almost never for production. Sonnet 3.5 matches or exceeds Opus on most benchmarks at 1/5 the cost. Reserve Opus for genuinely hard reasoning.
Worked example: RAG assistant on Sonnet
50K calls/month, 8K input tokens (docs + question) and 600 output tokens, Claude 3.5 Sonnet at $3/$15 per 1M. Uncached: input = 50K*8K=400M*$3/1M=$1,200; output=50K*600=30M*$15/1M=$450. Total ≈ $1,650/month. With cache write once and 70% of calls hitting cached reads at roughly one-tenth the input rate, input cost drops to near $450, bringing the total to about $900/month before output is even touched.
Mechanics: cache write vs cache read pricing
Anthropic prices a cache write at 1.25x the standard input rate and a cache read at roughly 0.1x standard input. That means the first call that populates the cache costs slightly more than normal, and every subsequent call reusing that prefix costs a fraction of normal — the break-even point is around two reuses of the same cached block within its TTL window.
Trade-offs and failure modes
— Cache breakpoints are manual; forgetting to set one silently reverts you to full-price input. — Batch API's 24h SLA doesn't suit anything user-facing. — Cached blocks expire (typically 5 minutes) — bursty, low-frequency traffic may never benefit. — Opus pricing ($15/$75) makes a routing mistake 5x more expensive than the same mistake on Sonnet.
What this guide does not cover
This guide does not cover Bedrock- or Vertex-specific discount structures, enterprise custom rate cards, or real-time usage monitoring — those require pulling live data from a provider console or billing API, which is outside what a token-volume scenario model like TokenAtlas is built to do. Use the figures here as planning inputs, not as a substitute for your actual invoice.
Frequently asked questions
- Is caching automatic?
- No — explicit cache breakpoints. Easy to add; TokenAtlas estimates savings.
- Bedrock vs direct API price?
- Bedrock matches direct list price; volume discounts vary.
- When to use Haiku?
- Classification, routing, simple Q&A, structured extraction.