Gemini API pricing explained

Google prices the Gemini Developer API the same way every major lab does now — per model, per million tokens, split into input and output rates — but its lineup skews cheaper at the low end than OpenAI's or Anthropic's, particularly in the Flash-Lite tier built for high-volume, low-latency work. That makes Gemini a common second-opinion model for teams doing a cost audit: worth checking even if you're not planning to switch your primary provider.

Flash, Flash-Lite and Pro

Flash-Lite is Google's cheapest current tier, aimed squarely at classification, tagging and short-turn tasks run at scale. Flash is the mid-tier default for most application traffic — good latency, solid capability, priced well under Pro. Pro is the reasoning-heavy tier for complex, multi-step tasks where accuracy matters more than per-request cost. As with every provider, picking Pro out of habit for a task Flash-Lite could handle is the fastest way to inflate a Gemini bill.

Context caching lowers the input rate on repeated context

Gemini's caching applies the same principle as every competitor's: a large context block you send on every request — a system instruction, a document, a tool schema set — can be cached once and referenced at a fraction of the standard input rate on subsequent calls. It matters most for workloads with a large, static context and a small varying query on top: document Q&A, codebase-aware agents, long-running chat sessions with a fixed persona prompt.

Multimodal input changes the token math

Gemini's multimodal models accept images, audio and video alongside text, and each of those input types is converted into tokens by its own accounting — a single image or a short audio clip can contribute a meaningful token count well beyond what the same content would cost as a text description. If you're estimating a multimodal workload, don't reuse a text-only estimate; check the model's documented handling of non-text input on its official page, and lean on the token counter for the text portion rather than guessing the combined total.

Batch pricing for asynchronous jobs

Google also offers a batch mode discount for jobs that don't need a synchronous response, applied to both input and output tokens. Combined with Flash-Lite's already low standard rate, batch-processed Flash-Lite jobs are among the cheapest way to run large-scale text classification or extraction across any provider in this dataset — worth checking against your current setup even if you're not on Gemini for anything else today.

• Multimodal input (images, audio, video) is metered differently from text and can dominate a request's token count — check the model page's context window and modality fields before assuming a text-only estimate holds.
• Preview and experimental Gemini models sometimes carry different pricing or free-tier terms than the generally available release; confirm status on the model page before committing production traffic.
• Flash-Lite's combination of low input rate and low output rate makes it worth a direct side-by-side against your current default model, not just against other Google tiers.

Compare Gemini directly against a specific Claude or GPT tier on the comparison pages — same workload, same worked cost examples, no guessing at the multiple.

Where Gemini fits against Claude and GPT

Google, Anthropic and OpenAI all ship a low-cost, mid-tier and premium-reasoning model at any given time, and the honest way to choose between them is a direct side-by-side on your actual task, not a general reputation. Gemini's Flash-Lite tier tends to undercut the equivalent entry-level tier from the other two labs on paper, but capability, context window and output quality for your specific task matter as much as the sticker price — see LLM pricing comparison, 2026 for the full lineup ranked together.

The honest caveat

Google ships new Gemini versions on a faster cadence than it retires old pricing pages, and preview models in particular can reprice or disappear without much notice. Every rate on this page is pulled live from a dataset with a per-model verified-as-of date rather than typed once and forgotten — check that date before you build a cost model around it for anything long-lived.