Gemini API Cost Calculator: Does Context Caching Actually Pay for Itself?

Every cached token saves you the difference between the standard input rate and the caching rate. Every cached token also costs you the storage rate for each hour it sits there. Set those equal and the token count cancels out on both sides, which leaves a threshold measured in requests per hour rather than dollars. For a storage rate s, a standard input rate i and a cached rate c, the floor is s / (i - c) requests per hour.

The cancellation is the useful part. It means a 4,000 token cache and a 4,000,000 token cache have exactly the same break-even request rate on the same model. Cache size changes how much you win or lose, never whether you win. What changes the floor is the tier, and the spread across tiers is wide enough to flip decisions.

Gemini 2.5 Flash-Lite is the outlier worth staring at. Its token rate is so low that the saving per cached token is tiny, while storage costs the same dollar amount as it does on the Flash tiers. You need about eleven calls an hour, sustained, before an explicit cache stops losing money there. On Gemini 3.5 Flash the same arithmetic gives you well under one call an hour, so almost any live workload clears it. Two models, same feature, opposite advice. That is the number the Gemini API cost calculator above reports back as a plain yes or no for your own request volume.

The last column is where it gets awkward. Batch halves the input rate on every model, which shrinks the per-request saving, while the storage rate mostly stays put. The floor therefore goes up in batch, sometimes a lot: Gemini 2.5 Flash needs about four calls an hour in realtime and about eight in batch, and Gemini 2.5 Flash-Lite goes from about eleven to about twenty five. Moving a job to batch to save money can quietly turn a cache that was paying for itself into one that is not.

Gemini 2.5 Pro and Gemini 3.1 Pro Preview publish two sets of rates split at 200,000 prompt tokens. On 3.1 Pro Preview, input goes from $2.00 to $4.00 per 1M and output goes from $12.00 to $18.00. On 2.5 Pro it is $1.25 to $2.50 and $10.00 to $15.00. The higher rate applies to the entire prompt once you cross, not to the overflow, so the bill steps rather than slopes.

This is a quiet failure mode in retrieval pipelines. Nobody sets out to send a 210,000 token prompt. The corpus grows, the retriever returns one more chunk, and a slice of production traffic silently moves into the higher bracket. Same code, same model, larger invoice. The Gemini API cost calculator above flags the crossing and shows how far past the line you are, which is why it takes a token count rather than a vague size label.

The Flash and Flash-Lite tiers publish a single flat rate with no threshold at all, which is worth remembering when you are choosing where to put a long-context workload. Splitting one oversized request into two under-threshold requests can genuinely cost less on Pro, at the price of whatever coherence you lose by splitting.

Google prices the Batch API at 50 percent of the standard interactive cost for the equivalent model, and states that batch jobs are designed to complete within a 24 hour turnaround, with many finishing well before that. You submit the same model with the same parameters through an asynchronous path and collect the results when the job finishes.

The 50 percent summary is accurate for input and output, and only for input and output. Read the batch column on each model and the caching rows tell a different story. Google prints the words "Same as Standard" in the batch caching cell for Gemini 3.1 Pro Preview and Gemini 3 Flash Preview, and the whole 2.5 line repeats its standard caching rate unchanged, so on five of the nine models here batch buys you nothing at all on cached tokens. Gemini 3.6 Flash, 3.5 Flash and 3.1 Flash-Lite do halve theirs, and 3.5 Flash-Lite lands at $0.02 against a standard $0.03. Storage is discounted on exactly one model, Gemini 3.1 Flash-Lite, at $0.50 against $1.00. This calculator reads every one of those numbers literally instead of deriving them from the headline rule, which is why the cached batch row can come out higher than a halving would suggest.

Caching and batching stack. Google's Batch API documentation states that context caching is supported for batch requests, by naming the cached content resource in the configuration of individual requests inside the batch. That combination is usually the cheapest row this Gemini API cost calculator produces, though on the 2.5 line and on both Preview tiers the cached portion of the prompt costs the same either way, so the saving comes entirely from the uncached remainder and the output.

The honest limit of batch is latency. Google positions it for non-urgent work, and a turnaround measured in hours rules it out for anything a person is waiting on, at which point the comparison collapses back to realtime with or without a cache.