Prompt caching savings calculator
See how much prompt caching would save on each model, and the hit rate you need before it pays off.
Free, runs in your browser, no sign-up. Updated 2026-10-25.
Result
At a 90% hit rate, caching saves up to $3,410 a month (59%) on Claude Fable 5.1.
| Model | No cache | With cache | Saved | Break-even hit rate | Cache write |
|---|---|---|---|---|---|
| Claude Fable 5.1Anthropic · checked 2026-10-23 | $5,800 | $2,390 | $3,410 (59%) | 20% | 1.25× |
| gpt-6-astraOpenAI · checked 2026-10-23 | $5,800 | $2,660 | $3,140 (54%) | 22% | 1.25× |
| Claude Opus 5Anthropic · checked 2026-10-23 | $2,900 | $1,330 | $1,570 (54%) | 22% | 1.25× |
| Claude Opus 5.5Anthropic · checked 2026-10-23 | $2,320 | $992 | $1,328 (57%) | 21% | 1.25× |
| gpt-5.6-solOpenAI · checked 2026-10-23 | $2,320 | $1,064 | $1,256 (54%) | 22% | 1.25× |
| Gemini 4 ArgonGoogle · checked 2026-10-23 | $1,160 | $476 | $684 (59%) | 0% | none applied |
| Claude Sonnet 5.5Anthropic · checked 2026-10-23 | $1,160 | $496 | $664 (57%) | 21% | 1.25× |
| gpt-6.1-solOpenAI · checked 2026-10-23 | $1,160 | $496 | $664 (57%) | 21% | 1.25× |
| Gemini 3.1 Pro PreviewGoogle · checked 2026-10-23 | $1,220 | $572 | $648 (53%) | 0% | none applied |
| Claude Sonnet 5Anthropic · checked 2026-10-23 | $1,160 | $532 | $628 (54%) | 22% | 1.25× |
| gpt-5.6-terraOpenAI · checked 2026-10-23 | $1,220 | $592 | $628 (51%) | 22% | 1.25× |
| deepseek-v4-proDeepSeek · checked 2026-10-23 | $686 | $227 | $459 (67%) | 0% | none applied |
| Gemini 2.5 ProGoogle · checked 2026-10-23 | $838 | $433 | $405 (48%) | 0% | none applied |
| Claude Haiku 4.5Anthropic · checked 2026-10-23 | $580 | $266 | $314 (54%) | 22% | 1.25× |
| Gemini 3.8 FlashGoogle · checked 2026-10-23 | $435 | $192 | $243 (56%) | 0% | none applied |
| gpt-5.4-miniOpenAI · checked 2026-10-23 | $458 | $222 | $236 (51%) | 22% | 1.25× |
| deepseek-flashDeepSeek · checked 2026-10-23 | $165 | $59.16 | $106 (64%) | 0% | none applied |
| Gemini 2.5 FlashGoogle · checked 2026-10-23 | $204 | $107 | $97.20 (48%) | 0% | none applied |
| MiniMax M3MiniMax · checked 2026-10-23 | $165 | $78.60 | $86.40 (52%) | 0% | none applied |
| gpt-5.6-lunaOpenAI · checked 2026-10-23 | $122 | $59.20 | $62.80 (51%) | 22% | 1.25× |
| MiniMax M3.1-FlashMiniMax · checked 2026-10-23 | $82.50 | $33.90 | $48.60 (59%) | 0% | none applied |
| Claude Haiku 5.5Anthropic · checked 2026-10-23 | $58.00 | $26.60 | $31.40 (54%) | 22% | 1.25× |
| gpt-6-lunaOpenAI · checked 2026-10-23 | $58.00 | $26.60 | $31.40 (54%) | 22% | 1.25× |
Cache-write premiums are sourced for Anthropic and OpenAI only; other providers show none applied. Cache storage fees are not included.
Assumptions
Every number the result depends on, and where it comes from. Change any value and the result updates.
| Assumption | Value | Source |
|---|---|---|
| Requests per month | requests | OceanLab planning assumption. Use your own volume.Reviewed 2026-10-25 |
| Repeated prefix per request (system prompt, tools, documents) | tokens | OceanLab planning assumption. Only an identical prefix can be cached.Reviewed 2026-10-25 |
| New input per request (not cacheable) | tokens | OceanLab planning assumption: the user's message and recent turns.Reviewed 2026-10-25 |
| Output per request | tokens | OceanLab planning assumption.Reviewed 2026-10-25 |
| Cache hit rate | % | OceanLab planning assumption: depends on traffic within the cache lifetime. Low-traffic apps miss more often as entries expire.Reviewed 2026-10-25 |
Ask OceanLab to cut your LLM bill
Opens an email to OceanLab with this result and a link to your inputs. Nothing is sent until you press send, and there is no sign-up.
How this is calculated
Without caching, the repeated prefix (system prompt, tools, documents) is billed at the input price on every request.
With caching, a hit is billed at the cached input price and a miss pays the input price times the provider's cache-write multiplier. Write multipliers are sourced for Anthropic and OpenAI; for other providers no write premium is applied.
Break-even hit rate = the share of requests that must hit the cache before caching costs less than not caching. Cache storage fees are not included.
Sources
- Anthropic prompt caching pricing · checked 2026-10-25
- OpenAI prompt caching pricing · checked 2026-10-25
- Google API pricing · checked 2026-10-23
- DeepSeek API pricing · checked 2026-10-23
- MiniMax API pricing · checked 2026-10-23
Questions about this tool
How is the saving worked out?
Without caching, the repeated prefix is billed at the input price every time. With caching, hits are billed at the cached price and misses pay the cache-write price.
What is the break-even hit rate?
The share of requests that must hit the cache before caching costs less than not caching. It exists because cache writes cost more than normal input for some providers.
Are cache storage fees included?
No. Some providers charge per hour to keep a cache; check their pricing page if you hold large caches for long.
Where do the prices come from?
Each price is read from the provider's official pricing page and shows the date it was checked. Prices that could not be verified are left out, or start at 0 for you to fill in.
Other free tools
OceanLab product builds like this
Products OceanLab designed and built. Each case study covers the screens, the scope and the rules it was built for.
Kaganek - RAG Knowledge Assistant with Citations and Permissions Intact
Designed and built a clickable internal knowledge assistant for a Kraków consulting engineering firm: staff ask in Polish or English and get answers built only from the firm's own SharePoint, Confluence, Google Drive and file share documents, with every claim linked to its source passage and an honest "I don't know" when the sources are weak.
Tekmė - No-Code Workflow Automation Builder with AI Steps
Designed and built a clickable no-code automation builder for a Vilnius SaaS startup selling to EU operations teams: a node canvas with branches, error paths and AI steps that classify, extract, summarise or act as an agent, a cost estimate per run, and a run history where every step is recorded with personal data masked.
Sonde - LLM Evaluation and Observability Platform with Release Gates
Designed and built a clickable LLM evaluation and observability platform for the AI team of a French health mutual: redacted production traces span by span, release gates that block a prompt or model change when quality regresses, alerts with owners and ageing, and the monitoring record its risk and compliance teams can show a supervisor.
