If your Gemini API bill jumped and you're trying to pin down why, it's almost always one of a handful of explainable mechanics — several of them specific to how Gemini bills thinking and caching, not the usual reasoning-token or usage-tier story you'd get from another vendor. Here's what to check first.
Thinking tokens are billed as output tokens — and the field that reports them isn't always populated
Google's own documentation states that response pricing is the sum of output tokens and thinking tokens: a thinking-enabled model's internal reasoning is billed at the same rate as the visible answer, tracked separately in the API response as a thoughts_token_count field alongside the normal output count. Developers on Google's own AI developer forum have reported that field isn't always populated even when thinking is active, which makes it easy to underestimate how much of a response's token count was invisible reasoning rather than the answer you actually see.
In a multi-turn conversation, resending a thought signature costs you again
Gemini preserves a model's reasoning context across turns using a "thought signature" that your app passes back on the next call — and per Google's own docs, that signature counts toward your input tokens, and therefore your cost, on every turn you resend it. An agentic loop or chat integration that keeps full conversation history — thought signatures included — in every subsequent call can accumulate this compounding cost without any single turn looking unusual on its own.
Context caching bills an ongoing storage fee, not just a cheaper read
Gemini's explicit context caching discounts cached input tokens well below the standard input rate, but it adds a cost mechanic other vendors' caching doesn't have: a separate, ongoing storage charge billed per token, per hour, for however long the cached content stays alive — charged whether or not anything actually hits the cache during that window. A cache set up for a burst of traffic and never explicitly expired keeps accruing that storage fee quietly, on top of whatever you're saving on reads.
A burst of traffic can trip a ten-minute spend ceiling before it touches your RPM
Beyond the usual requests-per-minute, tokens-per-minute, and requests-per-day limits, Gemini's usage tiers each carry a spend-based rate limit measured in rolling ten-minute windows — $10 per ten minutes at Tier 1, $50 at Tier 2, $200 at Tier 3 — regardless of how far you are from your monthly budget. A short, legitimate burst of higher-cost requests can trip that ceiling and return a 429 well before it comes anywhere near your account's actual spend limit, which is easy to mistake for ordinary rate limiting rather than this separate spend-pacing constraint.
Google changed a price and gave it a future effective date, not a headline
Google does publish price changes with a specific future effective date rather than an immediate jump — Gemini 3.8 Flash's context-caching price, for instance, is listed at one rate through the end of 2026 and a higher rate starting the following January, on Google's own current pricing page. That's easy to miss if a workload was priced once and never checked again before the date actually arrived. Check the Google calculator above for the latest verified rate and price history, and add a free watch so the next change reaches you before your invoice does.