Free calculator · Fable 5.1 cache reads $0.25/MTok since 2026-09-01 · prices verified 2026-09-19
Anthropic prompt caching pricing calculator: break-even and monthly savings
Anthropic prompt caching pricing comes down to two questions: how many reads it takes to repay the cache write, and what the monthly bill looks like once your stable prefix is served from cache. On Anthropic models a 5-minute cache write costs 1.25x base and pays back after one read; the 1-hour write costs 2x and pays back after two. Cache reads cost 10% of base on most Claude models and 2.5% on Claude Fable 5.1, $0.25 per million tokens against $10 uncached. Enter your prefix size, calls per window and volume to see the curve.
Table view
| Period | Uncached | Cached |
|---|---|---|
| 1 | 0 | 0 |
| 2 | 0 | 0 |
| 3 | 0 | 0 |
| 4 | 0 | 0 |
| 5 | 0 | 0 |
| 6 | 0 | 0 |
| 7 | 0 | 0 |
| 8 | 0 | 0 |
| 9 | 0 | 0 |
| 10 | 0 | 0 |
| 11 | 0 | 0 |
| 12 | 0 | 0 |
| 13 | 0 | 0 |
| 14 | 0 | 0 |
| 15 | 0 | 0 |
- Cache read price
- $0.2/MTok (10.0% of base)
- Cache write price
- $2.5/MTok (1.25x base)
- Break-even reads
- 1
- Cache windows per month
- 20,000
- Annual savings
- $71,136
On Claude Sonnet 5, caching a 12,000-token prefix across 15 calls saves 60%: $9,900 becomes $3,972 a month.
Assumptions and sources (15)
| Constant | Value | Basis |
|---|---|---|
| Claude Fable 5.1 price per MTok | $10 in / $50 out, cache read $0.25, cache write $12.5 (cache read 0.025x base) | sourcedAnthropic pricing page, fetched 2026-09-19 |
| Claude Opus 5 price per MTok | $5 in / $25 out, cache read $0.5, cache write $6.25 | sourcedAnthropic pricing page, fetched 2026-09-19 |
| Claude Sonnet 5 price per MTok | $2 in / $10 out, cache read $0.2, cache write $2.5 (introductory price made permanent) | sourcedAnthropic pricing page, fetched 2026-09-19 |
| Claude Haiku 4.5 price per MTok | $1 in / $5 out, cache read $0.1, cache write $1.25 | sourcedAnthropic pricing page, fetched 2026-09-19 |
| GPT-5.6 Sol price per MTok | $5 in / $30 out, cache read $0.5, cache write $6.25 | sourcedSECONDARY: CloudZero GPT-5.6 pricing tracker (updated 2026-09-10). openai.com/api/pricing returned HTTP 403 on 2026-09-19 |
| GPT-5.6 Terra price per MTok | $2 in / $12 out, cache read $0.2, cache write $2.5 | sourcedSECONDARY: CloudZero GPT-5.6 pricing tracker (updated 2026-09-10). openai.com/api/pricing returned HTTP 403 on 2026-09-19 |
| GPT-5.6 Luna price per MTok | $0.2 in / $1.2 out, cache read $0.02, cache write $0.25 | sourcedSECONDARY: CloudZero GPT-5.6 pricing tracker (updated 2026-09-10). openai.com/api/pricing returned HTTP 403 on 2026-09-19 |
| Gemini 3.6 Flash price per MTok | $0.75 in / $3.75 out, cache read $0.075 ($1.50 / $7.50 / $0.15 after 2026-12-31) | sourcedGoogle Gemini API pricing page, fetched 2026-09-19 |
| Gemini 3.5 Flash price per MTok | $1.5 in / $9 out, cache read $0.15 | sourcedGoogle Gemini API pricing page, fetched 2026-09-19 |
| Anthropic write premium | 5-minute 1.25x base; 1-hour 2x base | sourcedAnthropic pricing page, 2026-09-19 |
| Anthropic read discount | 0.1x base; 0.025x on Fable 5.1 / Mythos 5.1 | sourcedAnthropic pricing page, 2026-09-19 |
| OpenAI cache pricing | cached input ~90% off; write 1.25x | sourcedSECONDARY: CloudZero tracker 2026-09-10 (openai.com returned 403) |
| Google cache storage | billed per hour, omitted here | estimateGemini API pricing, 2026-09-19; storage fee not modelled |
| Minimum cacheable prefix | 1,024 tokens on most Claude models | sourcedAnthropic prompt caching docs |
| Window model | one write then (calls - 1) reads per window; prefix byte-identical | estimateStudio simplification; real hit rates depend on traffic gaps |
Key takeaways
- A 5-minute cache write pays back after one read. A 1-hour write after two.
- Fable 5.1 cache reads cost $0.25 per million, 2.5% of base. That is the agentic cost story of September 2026.
- Put the stable prefix first and keep it byte-identical. One changed character invalidates the cache.
- Caching multiplies with loop length. Long tool loops are where it saves most.
How prompt caching savings are calculated
Per window: the first call writes the prefix at the write price and the rest read it at the read price. Uncached input, output and everything after the prefix are billed as normal. Compare that to every call paying full input price for the prefix, and multiply by windows per month.
The break-even is the write premium divided by the per-read saving. On Anthropic's 5-minute cache the premium is 0.25x and the saving is 0.9x, so one read repays it (Anthropic prompt caching docs). TheoSym builds the prefix layout into the harness it ships with every production agent, because the saving only exists if the prefix is stable.
- Cached per window = write + (calls - 1) x read + dynamic + output
- Uncached per window = calls x (prefix + dynamic + output)
- Break-even reads = write premium / per-read saving
Anthropic prompt caching pricing vs OpenAI and Google: cache read and write cost
Anthropic: read at 0.1x base, or 0.025x on Fable 5.1 and Mythos 5.1; 5-minute write 1.25x, 1-hour write 2x. OpenAI: cached input about 90% off with a 1.25x write, per the CloudZero tracker (secondary source; openai.com returned 403 on 2026-09-19). Google: Gemini 3.6 Flash context caching at $0.075 per million plus an hourly storage fee this tool omits.
The 2026-09-01 Fable 5.1 change is the one to watch. Anthropic cut cache reads 75% and said highly agentic workloads can fall by up to about 45% (VentureBeat). A frontier model that was too expensive for a loop in August may not be in September. The eval decides; see the AI Factory workproofs.
What is a good prompt cache hit rate for an agent?
Above 80% of input tokens for a tool-using agent with a large system prompt and tool definitions. Below 50% usually means the prefix is being invalidated: a timestamp in the system prompt, tool definitions that reorder, or a conversation trimmed from the front instead of the back.
The fix is structural. Stable content first, in a fixed order, byte-identical. Dynamic content last. Anthropic's automatic caching manages breakpoints as a conversation grows; you still own the ordering. TheoSym's AI consulting reviews trace cost per loop step, which is where the missed hits show up.
When do you need a caching review?
When the agent bill grows faster than usage, when you move to a frontier model, or when calls per task pass about ten. At that loop length the prefix is being re-sent enough times that the read discount dominates the bill.
Also when a model retires: GPT-5.5 leaves ChatGPT and Codex on 2026-10-14, and the replacement model has different cache economics. Price it here, then test it with evals before switching. Regulated finance workloads route to QGI, where reproducibility, not cost, sets the model choice.
Anthropic prompt caching pricing: questions people ask
How does Anthropic prompt caching pricing work on Claude?+
Cache reads cost 10% of base input price on most Claude models and 2.5% on Claude Fable 5.1, where a cached token costs $0.25 per million against $10 uncached. A 5-minute cache write costs 1.25x base and pays back after a single read, so repeated system prompts and documents typically save 50% to 90% of input cost.
What is the break-even point for prompt caching?+
On Anthropic's 5-minute cache, one read repays the 1.25x write premium; on the 1-hour cache (2x write) two reads do. OpenAI applies a similar 1.25x write and 90% read discount. If your prefix is reused fewer times than that within the window, caching costs more than it saves.
Why did Anthropic cut Fable 5.1 cache read pricing by 75%?+
Agentic workloads re-read the same system prompt, tool definitions and context on every loop step, so cache reads dominate the bill. Anthropic priced Fable 5.1 cache hits at $0.25 per million tokens from 1 September 2026 and says highly agentic workloads can fall by up to about 45%.
Does prompt caching work for multi-turn agents?+
Yes, and that is where it pays most. Place the stable prefix first (system prompt, tools, reference documents), keep it byte-identical, and let the conversation grow after it. Anthropic's automatic caching manages breakpoints as the conversation grows; the calculator's calls-per-window input models the loop length.
Does this calculator include Google cache storage fees?+
No. Gemini bills a per-hour storage fee for cached context on top of the $0.075 per million read price for 3.6 Flash, and that storage line is omitted here. For short windows it is small; for long-lived caches it is not. Treat the Gemini figures as a lower bound and check the pricing page for your window length.
Next step
Turn the estimate into a scoped plan
Send the calculator result. You get one reply from a human within one business day, with an eval plan for the workflow and a real scope.
Prefer to talk? +1 657-888-0688 or contact@theosym.com