Gemini's API price hike: budget the model, not the demo

The model you priced your automation on in the spring is more expensive today. Google's Gemini API pricing page now lists Gemini 2.5 Pro at $1.25 input / $10 output per million tokens and Gemini 3.1 Pro Preview at $2 / $12 (rising to $4 / $18 past 200K-token prompts). Pricing trackers logged the step-up in early July — and if your budget still carries the old numbers, the gap is real money.

What actually happened

There was no launch event; the rate card just moved. Per pricing registries that track provider changes, Gemini 2.5 Pro's output price roughly doubled (from around $5 to $10 per million), Gemini 3.1 Pro climbed from about $1.25 / $5 to $2 / $12, and Gemini 2.5 Flash — the cheap workhorse a lot of automations lean on — went from roughly $0.15 / $0.60 to $0.30 / $2.50. Google's own pricing page confirms the current numbers, including a $1.50 / $9 rate on Gemini 3.5 Flash.

Output is where it bites. Most real workloads — drafting, summarizing, structured extraction, anything agentic — generate far more output than they consume in input, so an output-side jump lands squarely on the line that scales with usage. A Flash-based flow that quadrupled its output rate didn't get a slightly higher bill; it got a different unit economics. That's the same lesson Anthropic's Sonnet 5 tokenizer taught last week from the other direction: the rate card and the invoice are not the same document.

Why it matters for your business

Provider pricing is not a fixed input you get to design around once. It's a dial someone else controls, and they turned it. If your app calls gemini-2.5-pro directly in a dozen places, a price change like this is a scavenger hunt — find every call site, re-estimate, and hope you caught them all. If the model sits behind a boundary you own — one config value, one gateway — the same change is a five-minute decision: eat it, drop to a cheaper tier, or route this workload to a different provider entirely.

That's the whole case for treating the model as a swappable component. The specific number matters less than the discipline: know your cost per task, watch it against a baseline, and keep the model layer loose enough to move. When a vendor reprices — and in 2026 they reprice constantly, up and down — you want a Tuesday config change, not a quarter-end surprise and a rewrite. Re-run the ROI on any Gemini-backed workflow now; the math you approved in Q2 may not clear anymore.

Sources: Google Gemini API pricing.