Do prompt cache reads count toward the Claude Code and Codex usage limits?
Updated
Yes. Cache reads still count, at a reduced rate. Claude Code re-reads history at the cached token rate, so 'a one-line question in a session that has been open all day still draws usage for the whole conversation'; cache reads cost 0.1 times base input, and a break longer than the one-hour subscription cache lifetime triggers a full re-read. OpenAI lists caching among the factors that affect Codex usage, with cached input at one tenth of the input rate.
What the official docs say
Cache read tokens are 0.1 times the base input tokens price
Quoted from Claude Platform docs: Prompt caching a one-line question in a session that has been open all day still draws usage for the whole conversation
Quoted from Claude Code docs: Manage costs effectively Cache misses: your first message after a break longer than the cache lifetime misses the cache and reprocesses your full context.
Quoted from Claude Code docs: Manage costs effectively The lifetime is an hour on a subscription
Quoted from Claude Code docs: Manage costs effectively Model choice, context, reasoning, tool use, retrieval, and caching all affect usage
Quoted from Codex docs: Pricing (learn.chatgpt.com) we root caused it to an optimization that we rolled back that had an impact on cache hit rates
Quoted from Tibo Sottiaux on X, 2026-05-23 (cache hit rate optimization rolled back)
How it is counted
On Claude, a cache hit is cheaper input, not free input. The Platform docs say 'Cache read tokens are 0.1 times the base input tokens price' (0.025x on Claude Fable 5.1 and Mythos 5.1), while cache writes cost 1.25x base input for a 5-minute cache and 2x for a 1-hour cache.
Claude Code applies that to your whole conversation. 'With prompt caching, Claude Code re-reads that history at the cached token rate', which is why 'a one-line question in a session that has been open all day still draws usage for the whole conversation'. Misses are worse: 'Cache misses: your first message after a break longer than the cache lifetime misses the cache and reprocesses your full context', and 'The lifetime is an hour on a subscription'. The /usage breakdown flags long context or cache misses when they account for 10% or more of recent usage.
Codex says the same in fewer words: 'Model choice, context, reasoning, tool use, retrieval, and caching all affect usage'. Its credit price list bills cached input at one tenth of input (for GPT-6 Astra, 25 credits per million cached tokens against 250 for input). On 2026-05-23 Tibo Sottiaux traced a run of fast-draining limits to an optimization 'that had an impact on cache hit rates', which was rolled back.
| Provider | Item | Counted? | Source |
|---|---|---|---|
| Claude Code | Cache read | Yes. 0.1x base input price (0.025x on Fable 5.1 and Mythos 5.1) | [1] |
| Claude Code | Cache write | Yes. 1.25x base input for 5-minute caches, 2x for 1-hour caches | [1] |
| Claude Code | Cache miss after a break | Yes. Full context reprocessed; the lifetime is an hour on a subscription | [2] |
| Codex | Caching | Yes. Listed among the factors that all affect usage; cached input is one tenth of input on the credit price list | [3] |
FAQ
Why does a one-line question in a long Claude Code session use so much?
Because the whole history is re-read. The Claude Code docs say that with prompt caching it 're-reads that history at the cached token rate, so a one-line question in a session that has been open all day still draws usage for the whole conversation'.
How long does the Claude Code cache last on a subscription?
An hour. The docs say 'The lifetime is an hour on a subscription', and that 'your first message after a break longer than the cache lifetime misses the cache and reprocesses your full context'. Starting a fresh session after a long break avoids paying for a stale history.
Does Codex charge for cached input?
Yes, at a reduced rate. The pricing page lists caching among the factors that 'all affect usage', and its credit price list bills cached input at one tenth of the input rate (25 versus 250 credits per million tokens for GPT-6 Astra).
Has caching ever made limits drain faster?
On 2026-05-23 Tibo Sottiaux wrote that Codex had 'root caused it to an optimization that we rolled back that had an impact on cache hit rates'. Lower hit rates mean more full-price input, which is what users saw as limits draining faster.
Sources
- Claude Platform docs: Prompt cachingverified 2026-09-17
- Claude Code docs: Manage costs effectivelyverified 2026-09-17
- Codex docs: Pricing (learn.chatgpt.com)verified 2026-09-17
- Tibo Sottiaux on X, 2026-05-23 (cache hit rate optimization rolled back)verified 2026-09-17
Want the next reset on your phone?