Anthropic charges 1.25x the normal input rate on the cache WRITE (the first request that populates a cached prefix) and 0.1x the normal input rate on each cache READ (every subsequent hit on the same prefix). Compute the minimum number of total cache reads N (not counting the write) after which the cumulative cost is strictly LOWER than running the same N+1 requests against an uncached endpoint at 1.0x. Show the inequality and the integer answer.
Cached cost (write + N reads) = 1.25 + 0.1N. Uncached cost (N+1 calls) = N+1. Set cached < uncached: 0.25 < 0.9N, so N > 0.28. Caching wins from the FIRST read (1.35 < 2.0).
Picture a coffee shop that charges $1.25 to engrave your favorite order on a card the first time you visit and $0.10 each time you flash that card later. Without the card, every drink is $1.00. After your first refill ($1.25 + $0.10 = $1.35 total) you've spent more than the $1 for a single drink but less than the $2 you'd have spent on two regular drinks. From your second card flash onward, you're saving by a wide margin. The card pays off remarkably fast.
Concept explanation~2 min read
Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example. Click to expand.
Concept explanation~2 min read
Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example. Click to expand.
Prompt cache break-even math is the kind of question that sorts candidates who memorized a number from candidates who can derive one. The published headline figures, 'break-even at 2 reads' is common, come from a partial analysis that doesn't compare cached against the right baseline. Writing down the inequality explicitly and solving it shows that, against the actual alternative of running the same requests uncached, the cached recipe wins from the very first read.
The pricing structure itself is the lesson. Anthropic charges 1.25x on the cache write, a 25% premium over the standard input rate, and 0.1x on each subsequent read, a 90% discount. The 1.25x write is meant to recoup some of the infrastructure cost of populating and maintaining the cache; the 0.1x read is meant to push the cost to benefit ratio sharply in favor of any prefix that gets reused. The math confirms: the write premium of 0.25 is wiped out by the first read's avoided 0.9 of uncached cost.
This deep dive does the inequality explicitly, walks the comparison against the standard alternative, explains the '2 reads' folklore and why it answers a different question, then connects the math to the production engineering, cache_control placement, TTL choice, instrumentation, and the cross-provider differences that catch teams switching between Anthropic and OpenAI.
The inequality and its solution
Define one 'input unit' as the cost of one normal-rate input token. Compare two regimes over (write + N reads) total requests:
Cached regime. The first request populates the cache: cost = 1.25 units of input. Each of the subsequent N requests reads the cache: cost = 0.1 units each. Cumulative cached cost = 1.25 + 0.1N.
Uncached regime. Each of the same N+1 requests pays the full input rate. Cumulative uncached cost = 1.0 * (N+1) = N + 1.
The break-even is where the two are equal:
Since N is a non-negative integer, the smallest N at which cached is strictly less expensive than uncached is N = 1. Verification: at N=1, cached = 1.25 + 0.1 = 1.35; uncached = 2.0; cached saves 0.65 input units, about 32.5%.
The ratio grows fast. At N=10, cached = 2.25 vs uncached = 11.0, a 79.5% saving. At N=100, cached = 11.25 vs uncached = 101.0, an 88.9% saving. As N approaches infinity, cached approaches 0.1N while uncached approaches N, the limiting ratio is 10x, exactly what you'd expect from the 0.1x read discount.
Situations where this technique stops working.
2–4 min · Everything important, quickly.
Real products, models, and research that use this idea.
- Cursor and Anysphere have publicly reported 70-90% input cost reductions on their Claude usage after enabling prompt caching on system prompts and codebase context.
- Anthropic's own Computer Use sessions cache the lengthy system prompt across all turns, making the per-turn input cost manageable despite the verbose tool definitions.
What an interviewer would ask next. Try answering before peeking at the approach.
QHow does Anthropic's extended 1-hour TTL change the break-even math?
Extended TTL has a higher write multiplier (2x instead of 1.25x as of 2026 pricing). Inequality becomes 2.0 + 0.1N < N+1, giving N > 1.11, so break-even at N=2. Worth it only for workloads where 5-minute TTL would have caused frequent re-writes.
Red flags & common mistakes
The phrases that signal junior thinking. Click to expand.
Red flags & common mistakes
The phrases that signal junior thinking. Click to expand.
Quoting 'two reads to break even' without showing the inequality. That number compares cached vs cached without the premium, not cached vs uncached. The cached recipe wins versus uncached on the very first read.
60 second bullets to scan on the way to the call.
How Anthropic's 1.25x write / 0.1x read pricing works
The exact inequality and its solution
Primary sources. Browse if you want the original framing.
Same topic, related formats. Practice these next.