AI Model Pricing

Prompt Caching ROI Calculator

See the actual monthly savings from enabling prompt caching on Claude, GPT-5.6, or Gemini -- based on your real traffic shape, not just the headline discount.

CategoryAI Model Pricing 100% Free No Sign Up UpdatedAug 10, 2026

Interactive Tool

Prompt caching lets a provider skip recomputing the part of your prompt that repeats across requests — the system prompt, tool definitions, and few-shot examples — and bill that portion at a steep discount instead of full price every single time. This calculator shows the actual monthly savings for your specific traffic shape, not just the headline discount percentage.

Key takeaways

Claude and current-generation GPT-5.6 models both discount cached reads by 90% (0.1x input price); Gemini 3.1 Pro’s explicit caching discounts by 75%.
Caching only ever discounts input tokens. Output tokens are always billed in full, no matter which provider or model — any savings estimate that ignores this overstates the real number.
Claude charges a cache-write premium (1.25x input for the default 5-minute cache lifetime) every time the cache needs refreshing. GPT-5.6 introduced the same write-premium model in mid-2026 — cache writes now cost 1.25x input on both Claude and GPT-5.6, a change from GPT-5.6’s earlier free-write caching.
The bigger your repeated prompt relative to your unique-per-request content, the more caching saves — a large system prompt with a short user question benefits far more than the reverse.

What this tool assumes

The calculation assumes your cache is written once and stays warm for the full day of traffic, then read for every request after that — a reasonable model for a stable system prompt under steady request volume. If your actual traffic is bursty enough that the cache frequently expires and needs rewriting (Claude’s default cache lifetime is 5 minutes; GPT-5.6’s explicit caching uses a 30-minute minimum), real savings on Claude and GPT-5.6 will be somewhat lower than this estimate, since you’d pay the write premium more than once per day on either.

How to read a low or negative savings result

If your repeated-prompt token count is small relative to your unique tokens and output, caching won’t move the bill much — there’s simply not much repeated content to discount. It can occasionally show near-zero net savings for Claude specifically at low request volumes, since the 1.25x write premium on a rarely-reused cache can offset the 90% discount on the few reads that do happen. If that’s your result, caching still isn’t harmful, it’s just not where your cost is coming from — check your output token volume instead, since that’s usually the larger, undiscountable line item.

Real Discount Rates

Verified 90% (Claude, GPT-5.6) and 75% (Gemini) cached-read pricing

Output Excluded

Correctly never discounts output tokens -- a common estimation mistake

Kept Current

Updated as provider caching pricing changes

No Sign-Up

Free to use, nothing to install

How It Works

Get the most out of this tool in a few simple steps

Pick Your Model

Claude, GPT-5.6, or Gemini 3.1 Pro

→

Describe Your Prompt Shape

Repeated tokens, unique tokens, and output size

→

See Your Real Savings

Monthly cost with and without caching, side by side

Perfect For

Use this tool for any task or profession

Developers with Large System Prompts

Check if caching is worth implementing before writing the code

Teams Auditing API Spend

Quantify the exact savings before prioritizing the engineering work

Anyone Comparing Providers

See how the write-premium mechanics differ between Claude and GPT-5.6

Anyone Reconsidering

Re-check savings as traffic volume or provider pricing changes

Frequently Asked Questions

Common questions about this tool, answered directly.

Does prompt caching affect response quality?

No. Caching is a billing and latency optimization only -- the model reuses the already-computed state for the repeated prefix and generates the same output it would have generated anyway. It's not a form of compression or summarization that could change the response.

What counts as "repeated" versus "unique" tokens?

Repeated tokens are the exact same text sent at the start of every request -- typically your system prompt, tool/function definitions, and any few-shot examples that don't change. Unique tokens are whatever comes after that shared prefix and differs per request, usually the user's actual message. Caching requires an exact prefix match, so even a one-character difference in the "repeated" portion breaks the cache for that request.

Why does Gemini show a smaller discount than Claude or GPT-5.6?

This tool uses Gemini's explicit context caching figure (75% off), which is the guaranteed, predictable rate with a documented minimum cacheable length. Gemini also offers automatic implicit caching that can reach closer to 90% in practice, but since it activates automatically based on traffic patterns rather than a guaranteed setting, it's not modeled here to avoid overstating a number you can't reliably plan around.

Do I need to change my code to enable caching?

It depends on the provider and version. OpenAI's automatic caching still applies once your stable prefix exceeds their minimum token threshold, but GPT-5.6 also added explicit cache breakpoints as an option, and cache writes on GPT-5.6 now cost 1.25x input -- previously free -- so it's worth confirming current behavior for your specific model rather than assuming the older free-automatic-write model still applies. Anthropic requires explicitly marking where the cacheable portion ends with a cache_control parameter in your API call. Gemini's explicit caching requires creating a cache object via their API before referencing it in requests.

Is this the same as the RAG Pipeline Cost Calculator?

No -- they model different cost levers. The RAG Pipeline Cost Calculator breaks down generation, embedding, and vector database costs for a retrieval system. This tool is specifically about the savings from caching a repeated prompt prefix, which applies to any high-volume API use case, not just RAG.