How are token costs calculated?

Multiply input tokens by the input rate and output tokens by the output rate, add the two amounts and divide by 1,000,000. A standard request example uses 1,000 input and 500 output tokens. Monthly budgets multiply the result by request volume; taxes, hidden reasoning tokens, tools and tier surcharges can add costs.

Which recorded model has the lowest request cost?

LFM 2.2 6b by Liquid AI costs $0.00002 for the standard request, with $0.01 input and $0.02 output per million tokens. This price ranking does not establish quality equivalence; verify task success and context requirements.

Can cached and batch discounts be combined?

Only recorded rates are used. Cached input can have cache-write or storage charges, while batch rates apply only to eligible asynchronous jobs. We do not automatically stack the discounts or treat an absent rate as zero. Check the linked provider source before budgeting.

How current is the public dataset?

The live pricing table synchronizes stored model records daily; public file snapshots update when the site is rebuilt. Synchronization is not independent provider verification. It contains 1,026 priced models across 124 providers, last updated 2026-10-31. Directory entries are counted separately. A synchronization date is not proof of independent verification.

Guide · 2026-04-14 · 6 min read

Prompt Caching: How It Works and How Much You Save

A deep dive into prompt caching across OpenAI, Anthropic, and Google — how prefix caching works, when it kicks in, and real savings benchmarks.

What Is Prompt Caching?

Every time you call an LLM API, you send a system prompt, context, and user input. If your system prompt is 3,000 tokens and you make 1M requests/month, you're paying for those same 3,000 tokens a million times. Prompt caching fixes that.

Providers detect when the beginning of your prompt (the "prefix") matches a previous request. Instead of reprocessing those tokens from scratch, they reuse cached computations and charge you a fraction of the normal input price.

Provider Comparison

Anthropic's approach gives you more control — you explicitly mark cache breakpoints in your prompt. OpenAI and Google cache automatically based on prefix matching.

How to Structure Prompts for Caching

The golden rule: put static content first, dynamic content last.

If you shuffle the order — say, putting the user question before examples — caching breaks because the prefix no longer matches.

Real Savings Example

Let's model a RAG chatbot with a 2,500-token system prompt making 500K requests/month on GPT-5:

Without caching: 2,500 × 500K = 1.25B input tokens × $2.50/M = $3,125
With caching: 2,500 tokens cached at $1.25/M = $1,562

Savings: $1,563/month — just from reordering your prompt.

When Caching Doesn't Help

Unique prefixes: If every request starts differently, there's nothing to cache
Short prompts: If your system prompt is under the minimum prefix length, caching won't activate
Low volume: Cache entries expire (typically 5-10 minutes of inactivity), so sporadic traffic won't benefit

Combining with Other Optimizations

Prompt caching stacks beautifully with tiered routing. Cache your system prompt on every tier, and you get compound savings:

1. Route simple queries to GPT-5 Nano ($0.10/M → $0.05/M cached)

2. Route complex queries to GPT-5 ($2.50/M → $1.25/M cached)

Use our Calculator to model caching savings for your specific workload.