Using caching effectively

What is caching?

In a conversation or AI Agent, many parts of the input may be sent again across many turns:

• System Prompt;
• Context;
• long documents;
• tool definitions;
• previous history.

Prompt or context caching lets a model provider reuse previously processed input when a new request meets that provider’s matching and eligibility rules.

Caching is handled on the model/API provider’s infrastructure.

F68 AI does not build its own content cache store for this feature.

Caching is not reusing answers

Cache does not mean F68 AI takes an old answer and returns it to you.

The model still generates new output for the new request.

Caching mainly reuses the repeated input/context.

When is caching useful?

Caching is especially useful when:

• using an AI Agent with a long System Prompt;
• using the same context across many turns;
• working repeatedly on the same set of documents;
• keeping a long conversation going;
• running many steps that share a stable part of the context.

Caching offers the greatest benefit when a stable block of context is reused across many requests.

When is caching less valuable?

Caching is less valuable when:

• you are only doing one short task;
• the context changes almost entirely in every turn;
• older content is constantly edited or moved;
• the repeated part does not meet the model’s cache conditions.

Turn caching on / off in F68 AI

In Settings, users can choose:

Prompt Caching: On / Off

On
F68 AI uses the model/API’s caching mechanism when the provider supports it and lets the application control it.

Off
F68 AI does not actively request caching on models/APIs that let the application control this mechanism.

Some providers apply implicit / automatic caching directly at the API layer. For those models, caching may still be applied automatically by the provider under its own policy.

Policies differ by vendor

OpenAI / GPT

GPT models support prompt caching on a matching prefix/context.

With GPT-5.6 and newer models:

• eligible input can be cached;
• cache has a minimum lifetime under the model’s policy;
• cache write and cache read are charged separately;
• cache read costs less than regular input on supporting models.

The specific policy and rates come from the model/backend configuration at the time of use.

Anthropic / Claude

Claude supports prompt caching with different cache durations.

Common configurations include:

• a cache lifetime of about 5 minutes;
• a longer cache of about 1 hour.

Cache write and cache read are charged separately.

Google / Gemini

Gemini supports context caching.

Some models/APIs use automatic implicit caching.

Some APIs support explicit caching.

F68 AI applies whatever mechanism the corresponding API/model allows.

xAI / Grok

Grok supports automatic prompt caching on the prefix that is identical between requests.

A cache hit is recorded by the provider in usage and has its own cached-token pricing per model.

DeepSeek

DeepSeek uses automatic context caching.

The API records separately:

• cache hit;
• cache miss.

The cache hit portion is charged under DeepSeek’s pricing policy at the time of use.

Other models

Caching information is shown or applied only when the corresponding model/API supports it.

A single cache policy is not used to describe all models.

How to increase the chance of a cache hit?

1. Keep the Agent configuration stable while working on the same workflow.
2. Keep the System Prompt and context stable at the start of the context when no change is needed.
3. Continue in the same conversation while working on the same task.
4. Don’t edit or rearrange older context unless necessary.
5. Use Branch chat when you want to try a new direction but still keep the original thread.

View Cache usage

In Usage history, if the API/model returns the corresponding data, F68 AI shows usage components such as:

Input · Output · Cache Read / Cache Hit · Cache Write · Credit

Not every model returns the same usage structure.

The final credit charge is based on the actual usage reported for that model or API.

Easy-to-remember principles

Caching doesn’t make the model smarter.
Caching helps the repeated part of the context be processed more efficiently.

Context control helps you decide what the AI needs to carry.
Caching helps reduce cost when that context is reused by the provider.