How does context caching work?

Caching lets you store a large input once and reference it in later calls at a reduced rate. You send the big document or codebase to the cache, get a cache handle, then every follow-up question reuses it. Cached input tokens cost a fraction of fresh ones. For a document you plan to query more than twice, caching pays for itself immediately.

How do you route between Ultra and Flash?

Not every call needs the flagship. A simple router decides per request:

• Routine summarization, classification, extraction → Gemini Flash.
• Cross-document reasoning, multimodal analysis, hard synthesis → Gemini 3 Ultra.
• First attempt cheap; escalate to Ultra only on failure or explicit need.

In practice, 80% of calls are fine on Flash. Routing those away from Ultra is the single biggest saving, bigger than any prompt tweak.

How much can you actually save?

Take a pipeline of 1,000 document queries against a 200k-token corpus. Sending fresh context each time on Ultra costs thousands of dollars. Enable caching and the input cost drops by roughly 75% on repeat queries. Route 80% of calls to Flash and the blended bill falls by another large step. Combined, a 70-85% reduction is realistic without changing answer quality on routine tasks.

What prompt practices reduce tokens?

Ask for the output format up front so the model does not ramble. Request JSON when you will parse it. Set a max output token cap — long outputs are billed and often wasteful. And split multi-part questions into targeted calls rather than one mega-prompt that drags in irrelevant context.

When is caching not worth it?

If each request uses a unique document, there is nothing to cache and you pay full price. Caching also adds complexity: cache handles expire, so a long-idle workflow re-uploads. For one-shot jobs, skip it. For a batch run overnight against a fixed corpus, it is almost always right.

Quick Summary

• Context caching reuses large inputs at a steeply reduced token rate.
• Route 80% of calls to Flash, reserve Ultra for hard tasks.
• Combined caching plus routing can cut a bill 70-85%.
• Cap output tokens and request structured JSON to save more.
• Caching only helps when the same context repeats across calls.

Last updated September 2026.

What is Gemini 3 Ultra?

Gemini 3 Ultra is the top tier of Google’s Gemini 3 family, sitting above the Pro and Flash variants. It accepts text, images, audio, and video, and supports very large inputs — enough to hold multiple books or a mid-sized repository. It is a hosted model: no weights, accessed through the API or AI Studio.

How do you feed it a long document?

Three steps cover most cases:

1. Upload the file through the File API, which returns a handle you reference in the prompt.
2. Ask a scoped question — “list every indemnity clause and the page it appears on.”
3. Request structured output, like JSON, so the answer is machine-usable.

Uploading once and asking many questions is far cheaper than re-sending the document each turn. The File API keeps it resident for the session.

Is a million-token context really usable?

Partly. Retrieval quality degrades in the middle of very long inputs — models tend to weight the start and end more heavily. Practical guidance: if you only need one section, retrieve that section rather than sending everything. Send the whole thing when the task genuinely needs cross-document reasoning, such as reconciling figures across a dozen reports.

How much does Gemini 3 Ultra cost?

Input tokens are billed at a premium tier, higher than Flash by roughly an order of magnitude. A single 500k-token document query can cost several dollars. That is fine for a one-off audit and wasteful in a loop. Use Flash for routine summarization and reserve Ultra for the analysis where the extra capability pays for itself.

Where does it beat other models?

Multimodal input is the differentiator. Drop in a screen recording plus the source code and ask where the UI diverges from the spec — Gemini 3 Ultra can watch and read in one pass. For pure code edits, tools like gpt-5-codex are more focused. For long-document and multimodal reasoning, Ultra has few peers.

Quick Summary

• Gemini 3 Ultra is Google’s flagship model with million-token context.
• Accepts text, images, audio, and video through the File API.
• Upload once, ask many scoped questions to control cost.
• Long-context retrieval weakens in the middle — retrieve when you can.
• Premium pricing; keep Ultra for hard analysis, use Flash for volume.

Last updated September 2026.