A log study of Claude Code and Codex sessions found a sharp rise in cold cache reads after an hour of inactivity in Claude Code, while Codex’s cold reads increased more gradually. The results offer developers a useful look at what may happen when they return to long coding-agent sessions, but they are observations from one operator’s logs, not a controlled test of either provider’s systems.
Watch Desk analysis
What happened
The study’s author analysed local session logs from 500 recent sessions, mostly Claude Code, and a separate probe of 2,359 Codex sessions. In the Claude Code sample, 6 of 7 calls made after a 60-to-62-minute pause were classed as cold, compared with none of 10 after a 58-to-60-minute pause. The author says this pattern matches Claude Code’s documented one-hour cache lifetime for subscription use.
The Codex results showed no equivalent cliff: 16.7% of calls after one to two hours were cold, rising to 46.2% after two to six hours and 97.9% after 12 to 24 hours. The author notes that OpenAI describes the relevant cache lifetime as a minimum, not a deadline, and says the data cannot establish why a particular call had a low cache read.
Read the log study and its methods.
Key findings
- Claude Code shows a sharp change around an hour
In this sample, cold reads rose from 0 of 10 calls just before an hour to 6 of 7 just after it. - Codex’s cold reads rise over longer pauses
The share was 16.7% after one to two hours and 97.9% after 12 to 24 hours. - The cost figures are a model, not subscription measurements
The study uses API list-price weights and cannot establish how cache reads affect plan usage. - The samples have real limits
They come from one operator’s usage, use different definitions of “cold”, and are not a controlled cache-lifetime experiment.
Why it matters
Coding agents can resend long conversation histories with each call. When a return is cold, more of that input may need to be processed at the regular rate, which can affect efficiency. The study’s figures help explain why a long break might matter differently in Claude Code and Codex, while stopping short of showing what any individual user will pay or consume under a subscription.
The author estimates that human returns after more than an hour accounted for a quarter of the sample’s price-weighted input index. That is the total weight of those calls, not the extra cost caused by cold reads. A warm return would still have had a cost; the study does not measure the difference in subscription usage.
Our read
This is a worthwhile field observation, not a verdict on either provider’s cache design. The Claude Code pattern is striking, and the Codex curve is useful context, but the small samples near the one-hour mark and the lack of a controlled probe matter. Treat the numbers as a reason to ask better questions about long-session workflows, not as a billing calculator.
What to watch
- Whether controlled tests with fixed prompts and settings reproduce either pattern.
- Whether the providers clarify how cache reads affect subscription limits and usage.
- Whether further studies separate idle expiry from other causes of low cache reads.
Discussion spark: For long coding-agent sessions, would you rather the tool preserve the full conversation after a break, or start fresh with a short summary to avoid paying to carry old context?
Sources and evidence
Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.