What would save the most
Where the money goes
The same session on other models
Each on its default cache setting, cheapest first. Select one to load it above. Tokens are not equal across providers, and a cheaper model may need more turns, so treat this as a starting point.
Every turn, as a table
| Turn | Minute | Prompt | Read | Written | Full price | With caching | Without | Cache |
|---|
How this is worked out
Every turn sends the whole conversation again. The prompt on turn k is your starting context, plus k lots of new input, plus k minus 1 lots of output. That growth is why long sessions get expensive, and why caching matters.
Anthropic and OpenAI GPT-5.6 or later. Each turn reads the prefix the last turn left, and writes what is new at the cache-write price. If the gap is longer than the lifetime, nothing is read and the whole prompt is written again. On Claude, reads refresh the lifetime for free, and a request counts from when it starts.
Older OpenAI models. There is no write premium. OpenAI gives typical lifetimes rather than fixed ones, so we use the low end and say so next to the setting.
Gemini implicit caching. No write charge and no storage charge, with the published cached-token price on a hit. Gemini 4 Argon reprocesses the last 2,050 tokens of the previous request when a turn adds fewer than 2,050.
Gemini explicit caching. Your starting context is cached once at the input price, then read at the cached price. Storage is charged per token per hour for the whole TTL. New turns are never added to it.
What this leaves out
- Token counts are what you type in. We do not tokenize anything, and the same text counts differently between models.
- Google publishes no lifetime for implicit caching and does not guarantee hits, so those results are a best case.
- Google's Gemini API page does not state how cache creation is billed. We use the standard input price, from Google Cloud's caching overview for the same models. Minimums for some Flash-Lite models are not published, so none is applied.
- OpenAI's minimum for models before GPT-5.6 varies by request, so none is applied. Their breakpoint spacing is also not modelled, except GPT-5.5 at 2,048 tokens.
- Clearing the context writes the whole prompt again, as if nothing were cached. With a separate breakpoint on your system prompt, part of it could still be a hit.
- Not covered: batch and fast-mode prices, thinking tokens billed beyond what you enter as output, tool and search fees, images, regional uplifts, negotiated discounts, rate limits and context-window limits.
- Anything that changes the prefix, such as editing tool definitions, breaks the cache in real life. Here the prefix never changes.
This is an estimate from published list prices, and your bill may differ. Cachemeter is independent and is not affiliated with or endorsed by Anthropic, OpenAI or Google. See the prices and their sources.