The rules side by side
What each provider documents, as the simulator models it. Each model page has the rules for that model.
| Rule | Anthropic | OpenAI | |
|---|---|---|---|
| Turning it on | You mark cache breakpoints in the request, or turn on automatic caching, which moves the breakpoint forward as the conversation grows. | Automatic. Any repeated prefix can be read from cache, with nothing to set. | Implicit caching is automatic on Gemini 2.5 and newer. Explicit caching means creating a cache object and referring to it. |
| How long it lasts | 5 minutes by default, or 1 hour. Each read refreshes the lifetime for free, timed from the start of the request. | GPT-5.6 and later: at least 30 minutes, or at least 12 hours on the Responses API. Older models give typical lifetimes only: about 5 to 10 minutes in memory, about 30 minutes with 24h retention. | Implicit: Google publishes no lifetime and no guarantee of a hit. Explicit: the TTL you set, 1 hour by default. |
| Price to write | 1.25x the input price for 5 minutes, 2x for 1 hour. | GPT-5.6 and later: 1.25x for 30 minutes, 2.5x for 12 hours, as total rates. Older models: no write charge. | Implicit: no charge. Explicit: the input price to create it, plus storage. |
| Price to read | 0.025x to 0.1x the input price, depending on the model. | 0.05x to 0.1x the input price. OpenAI calls it cached input. | 0.05x to 0.1x the input price. |
| Storage | None. | None. | Explicit only: $0.50 to $4.50 per million tokens per hour, for the whole TTL. |
| Shortest prompt cached | 512 to 4,096 tokens, depending on the model. Shorter prompts run uncached, with no error. | 1,024 tokens on GPT-5.6 and later. On older models OpenAI says it varies by request. | 2,048 to 4,096 tokens for implicit caching. Not published for Gemini 3.5 Flash-Lite, Gemini 3.1 Flash-Lite, Gemini 3 Flash Preview and Gemini 2.5 Flash-Lite. |
| Long prompts | No long-context premium on any Claude model here, except Claude Haiku 5.5, priced as a whole above 100K tokens. | Long-context prices above 272K input tokens, on every model here except GPT-5.4 mini and GPT-5.4 nano. | Gemini 3.1 Pro Preview and Gemini 2.5 Pro cost more above 200K tokens. |
Sources: Anthropic's pricing and caching pages, OpenAI's pricing and caching pages, Google's pricing and caching pages, and Google Cloud's caching overview for how explicit caches are created. See every price and source.
How many reads until caching pays off
For one prefix sent again and again inside the cache lifetime, counted by the simulator's cost model. The answer does not depend on how long the prefix is, as long as it clears the minimum.
Claude, 5 minute cache
Pays for itself after 1 read
A write costs 1.25x the input price, so one read inside 5 minutes covers it.
Claude, 1 hour cache
Pays for itself after 2 reads
A write costs 2x. Worth it when your gaps are longer than 5 minutes but under an hour.
GPT-5.6 and later, 30 minutes
Pays for itself after 1 read
A write costs 1.25x. OpenAI may keep the entry longer than 30 minutes.
GPT-5.6 and later, 12 hours
Pays for itself after 2 reads
A write costs 2.5x, for a cache that outlives a lunch break or a night.
No write charge
Pays for itself after 1 read
Older GPT models and Gemini implicit caching. Caching can never cost more than not caching, but the lifetime is not guaranteed.
Gemini explicit caches: uses needed to break even
An explicit cache is billed at the input price to create, then for storage over its whole TTL, read or not. The cheaper the model, the smaller the saving per read, so a long TTL can take hundreds of uses to earn back. Counts include the request that creates the cache. Prices per million tokens.
| Model | Input | Storage per hour | 5 minute TTL | 1 hour TTL | 24 hour TTL |
|---|---|---|---|---|---|
| Gemini 3.8 Flash | $0.75 | $0.50 | 2 | 2 | 19 |
| Gemini 3.6 Flash | $0.75 | $0.50 | 2 | 2 | 19 |
| Gemini 3.5 Flash-Lite | $0.30 | $1.00 | 2 | 5 | 91 |
| Gemini 3.1 Pro Preview | $2.00 | $4.50 | 2 | 4 | 62 |
| Gemini 3.1 Flash-Lite | $0.25 | $1.00 | 2 | 6 | 108 |
| Gemini 3 Flash Preview | $0.50 | $1.00 | 2 | 4 | 55 |
| Gemini 2.5 Pro | $1.25 | $4.50 | 2 | 6 | 98 |
| Gemini 2.5 Flash | $0.30 | $1.00 | 2 | 5 | 91 |
| Gemini 2.5 Flash-Lite | $0.10 | $1.00 | 3 | 13 | 268 |
What the lifetime does to a real session
The simulator's coding agent: 40 turns, 20K tokens of starting context, 3K in and 600 out per turn, 1.5 minutes apart, three long breaks of 8 minutes. Only the cache setting changes.
| Model | Cache setting | With caching | Without | Saved | Goes cold | Simulator |
|---|---|---|---|---|---|---|
| Claude Opus 5.5 | 5 minutes | $3.53 | $15.39 | 77% | 3 | Open |
| Claude Opus 5.5 | 1 hour | $2.50 | $15.39 | 84% | 0 | Open |
| GPT-6.1 Sol | 30 minutes | $1.00 | $7.70 | 87% | 0 | Open |
| GPT-6.1 Sol | 12 hours | $1.41 | $7.70 | 82% | 0 | Open |
| Gemini 3.1 Pro Preview | Implicit (automatic) | $1.33 | $7.74 | 83% | Never, best case | Open |
| Gemini 3.1 Pro Preview | Explicit, 1 hour TTL | $6.56 | $7.74 | 15% | 1 | Open |
On Claude Opus 5.5, the 40-turn coding agent costs $3.53 on the 5 minute cache, which goes cold 3 times at the long breaks, and $2.50 on the 1 hour cache, which stays warm. Each cold restart writes the whole prompt again at the write price.
What sends a cache cold
- A gap between requests longer than the lifetime. On Claude and GPT-5.6 and later, the next request pays the write price again for the whole prompt.
- Any change early in the prompt, such as a timestamp in the system prompt or an edited tool definition. Everything after the change is a miss.
- Clearing or compacting the conversation, which starts a new prefix.
- A prompt under the minimum length, which is never cached at all.
Try your own session in the simulator. It shows each cold restart on the chart and which change would save the most.
Questions
Which provider has the cheapest prompt caching?
As a share of the input price, the deepest discount is 0.025x, on Claude Fable 5.1. In dollars, the cheapest cache read is $0.01 per million tokens, on Claude Haiku 5.5, GPT-6 Luna and Gemini 2.5 Flash-Lite. On the 40-turn coding agent the cheapest run with caching is Gemini 2.5 Flash-Lite at $0.062, a best case because Google does not guarantee hits. A cheaper model may need more turns for the same work.
How long does the prompt cache last on Claude, GPT and Gemini?
Claude: 5 minutes, or 1 hour at a higher write price, and every read resets the clock. GPT-5.6 and later: at least 30 minutes, or at least 12 hours on the Responses API. Older GPT models: typically 5 to 10 minutes in memory, or about 30 minutes with 24h retention. Gemini: implicit caching has no published lifetime, and an explicit cache lasts for the TTL you set.
Does writing to the cache cost extra?
On Claude, yes: 1.25x the input price for 5 minutes and 2x for 1 hour. On GPT-5.6 and later, 1.25x for 30 minutes and 2.5x for 12 hours. Older GPT models and Gemini implicit caching charge nothing to write. A Gemini explicit cache is created at the input price and then charged for storage every hour.
When does prompt caching cost more than it saves?
When the cache expires before it is read again. Every cold start pays the write price for the whole prompt, so a Claude or GPT-5.6 session with gaps longer than the lifetime can cost more than sending everything uncached. Gemini explicit caches pay storage whether they are read or not: a 24 hour cache on Gemini 2.5 Flash-Lite needs 268 uses to pay for itself.
What is the minimum prompt length for prompt caching?
Claude: 512 to 4,096 tokens, depending on the model. GPT-5.6 and later: 1,024 tokens. Gemini implicit caching: 2,048 to 4,096 tokens, with no published minimum on Gemini 3.5 Flash-Lite, Gemini 3.1 Flash-Lite, Gemini 3 Flash Preview and Gemini 2.5 Flash-Lite. Shorter prompts are processed without caching.
Every Anthropic model Every OpenAI model Every Google model All 35 models compared
An estimate from published list prices, and your bill may differ. Cachemeter is independent and is not affiliated with Anthropic, OpenAI or Google.