26 October 2026 · Site Methodology & Corrections · 5 min read
I was about to write a routine "last verified" refresh on our prompt caching savings calculator -- it hadn't been touched since mid-August -- and the "Claude Sonnet" row still said $3 per million input tokens with a $0.30 cache-hit rate. That's Sonnet 4.x pricing. Current Sonnet is $2/MTok input. Small embarrassment, quick fix. But going to Anthropic's own pricing table to grab the right number turned up something I wasn't looking for: Sonnet 5.5's cache-hit price isn't just "the same $2 model, cheaper cache" -- it's a different discount rate than Sonnet 5 gets, and nothing in the headline pricing tells you that.
Sonnet 5.5's base input and output price is identical to Sonnet 5's: $2/MTok in, $10/MTok out. If you only ever look at the model-picker price list -- which is what our own calculator does, and probably what you do too -- you'd assume caching behaves the same on both. It doesn't. Every Claude model bills a cache hit at 0.1x the base input price, except two: Opus 5.5 and Sonnet 5.5, which get a steeper 0.05x multiplier. On Sonnet 5.5 that's $0.10/MTok for a cache hit. On Sonnet 5, same $2 base price, it's $0.20/MTok -- exactly double.
| Model | Input | Output | Cache-hit multiplier | Cache-hit price |
|---|---|---|---|---|
| Claude Sonnet 5.5 | $2/MTok | $10/MTok | 0.05x | $0.10/MTok |
| Claude Sonnet 5 | $2/MTok | $10/MTok | 0.1x (standard) | $0.20/MTok |
| Claude Opus 5.5 | $4/MTok | $20/MTok | 0.05x | $0.20/MTok |
| Claude Opus 5 | $5/MTok | $25/MTok | 0.1x (standard) | $0.50/MTok |
Checked directly against Anthropic's own pricing page, not a secondary source -- the footnote is explicit: "Cache hits and refreshes on Claude Opus 5.5 and Claude Sonnet 5.5 are priced at 0.05x the base input price," against the standard 0.1x everywhere else. It's a one-line footnote under a pricing table nobody re-reads once they've picked a model.
Run it through our own calculator's math: a RAG app sending a 6,000-token system+context block on 100,000 requests/month at a 90% cache-hit rate pays $54/mo for the cached portion on Sonnet 5.5, versus $108/mo on Sonnet 5 -- same model family, same $2 headline price, double the cache bill just from the multiplier most people never open the footnote to see. Scale that to a production agent doing millions of cached reads a day and the gap stops being a rounding error. The fix costs nothing: it's not a workload change, not a prompt change, just confirming which Sonnet version your API calls actually point at.
Three rows in the model dropdown were a full generation stale:
None of these were wrong the day they were written -- they were correct for whichever Claude generation was current back in roughly June. They just never got revisited across two model releases, so the "ranked by cost" table at the bottom of the page was quietly comparing today's GPT and Gemini prices against yesterday's Claude prices and still calling it a fair fight.
Updated the prompt caching savings calculator's model list to Sonnet 5.5 ($2 in / $2.50 5m-write / $0.10 cache-hit), Opus 5.5 ($4 / $5 / $0.20) and Haiku 5.5 ($0.10 / $0.125 / $0.01), refreshed the "last verified" date, and added a line to the disclaimer calling out the 95%-off cache read on Opus 5.5 and Sonnet 5.5 specifically, versus the standard 90% off everywhere else. If you're deciding between Sonnet 5 and 5.5 for a cache-heavy workload and the headline price looked like a wash -- it isn't, once caching is in the mix.
New to usage-based API pricing in general? Start with the free API-cost guides.
Deploy it yourself: DigitalOcean — $200 free credit ↗ · Hostinger VPS ↗
Checked directly against platform.claude.com/docs/en/about-claude/pricing, 2026-10-26. Reference figures, not live quotes -- confirm current terms on Anthropic's pricing page before budgeting.