A Gemini bill is not one number. Input, output, cached input and batch work are metered on separate lines, and the same request can cost several times more depending on which line it lands on. Rates follow the model version; token counts follow your request shape.
Google prices four metered lines, and they are not interchangeable:
So monthly cost = (input × input rate) + (output × output rate) + (cached × cached rate) + (cache storage × hours) + (batch × batch rate). Every optimisation changes one of those terms. We do not copy Google's rates: they differ per model and per tier. Read them from Google's Gemini API pricing page for the version you call.
Generation costs more per token than reading on every frontier model, and two multipliers stack: the output rate is higher, and most applications emit more output tokens than they consume.
A chat workload sending 3M input and 6M output tokens a month, with output at 4× input, comes to (3 × 1) + (6 × 4) = 27 cost units, of which output is 24 — 89% of the bill from 67% of the tokens. Flip the shape to summarisation: 12M input against 0.6M output is 14.4 units, and output is 17%. Same rates, opposite advice.
Gemini's large windows invite one mistake: paste the whole corpus into every request and let the model sort it out. It works, and it loads the input line faster than almost anything else.
Rule of thumb: above roughly 60% input share, work the input term — retrieve less, trim the system prompt, cache the stable prefix. If output dominates, cap generation length and check the model version.
Batch runs asynchronously at a lower rate. Split work on latency first: nightly classification, backfills and evaluation runs belong in batch; anything a user is waiting on does not.
A discount on twice the volume you needed is still more spend than not running the job.
Caching meters reads, priced below standard input, and storage, billed by time. That storage charge makes an always-on cache read once an hour a different proposition from one read twenty times a minute.
Break-even turns on reuse: one use is a loss, many uses across a stable prefix is the point. The prefix must be byte-stable, and your traffic must return inside the cache lifetime.
Cost units, output at 4× input, so the ratios hold at any dollar rate. Token counts are illustrative.
| Workload | Input tokens | Output tokens | Cost units (input + 4 × output) | Output share of bill | First lever |
|---|---|---|---|---|---|
| Chat, short prompts and long replies | 3M | 6M | 27 | 89% | cap max_tokens, trim the system prompt |
| Long-context document Q&A | 12M | 0.6M | 14.4 | 17% | retrieve less, cache the stable prefix |
| Agent loop with history | 6M | 1.5M | 12 | 50% | reset context, limit steps |
| Batch summarisation | 20M | 2M | 28 | 29% | batch path, then a smaller model |
The long-context row spends 12M tokens to get a 0.6M token answer — an input problem in output's clothing. The batch row shrinks twice: at the batch rate, then on a smaller model.
Five steps, using numbers you already own.
Re-run the sheet when rates change, when you swap versions, or when the bill moves more than 20% with no matching traffic change.
Tapix AI is a gateway, not a model vendor. We carry Gemini models below Google's list price — gemini-3-flash at $0.35 input and $2.10 output per million tokens against $0.50 and $3.00 on our pricing page, checked 2026-11-19. Same model identifier, per-token billing from prepaid credit, no monthly fee, failed requests not billed. That is a rate lever, not a substitute for the usage work above.
Q: How much does the Gemini API cost per million tokens?
It depends on model version and line: input, output, cached input and batch are priced separately, and some models carry a long-prompt tier. Read the current figures on Google's page.
Q: Why did my Gemini bill jump without more traffic?
Usually context growth, a prompt that crossed a long-prompt threshold, or a retry storm. Check output length per call, not the rate card.
Q: Is input or output more expensive?
Output, on every frontier model. Control output volume for chat and content work, input volume for long-context and retrieval work.
Q: Does prompt caching always save money?
No. Cached reads are cheaper than standard input, but storage is billed by time. A prefix used once loses; one reused many times inside the lifetime saves.
Q: How can I verify the prices on this page?
Compare our pricing page with Google's page for the same model identifier on the same day. If a provider will not show both numbers, treat its discount as unverified.
Price the output token first, compute your own input/output ratio in cost units, then fix whichever term dominates: context size, generation length, caching or batching. Or stop overpaying per token — the pricing page lists both rates per model.