I will say the awkward part first. Replay can read what GPT-6 Astra writes, and it knows Astra's published pricing, but I have never measured a real Astra session. So I do not have an Astra number, and I am not going to produce one.
What exists in the shipped build: the reader for the rollout format Astra writes, and the pricing rules that go with it. The path is there and the code runs.
On 123 sessions read 2026-09-17, 4.18% of my spend was tokens I paid for twice. That covers 105.2M tokens and $620.47 at list prices.
That is what a broken prompt cache does. The provider re-bills the whole conversation at write prices. Nothing errors. Nothing warns you. The bill just looks like you used the thing more.
I built Replay to find where prompt caches break and what that costs you. It can already read Astra's rollout files. The problem is I don't have any Astra traffic to point it at.
Here's the honest situation. Replay reads the same rollout format Astra writes, and it knows Astra's pricing and cache rules. But I've never actually measured an Astra session, because I can't afford to generate the traffic. When there's no data, Replay says NOT MEASURED instead of guessing. The site says the same thing.
See what your AI agent actually spent.
Replay reads the transcripts Claude Code and OpenAI Codex already keep on disk, reconstructs token usage and cache behavior, and names the turn a prompt cache broke and what it cost. It reproduces the provider's own cache reads before reporting anything, and prints NOT MEASURED where the evidence is absent instead of guessing.
Built to measure GPT-6 Astra. Not measured yet.
Single Go binary, zero third-party dependencies, no account.
