This is the single biggest cost cliff in the Gemini catalogue and the reason most production systems are designed around the 200K threshold. For long-document RAG workflows, it is cheaper to chunk and run two separate requests than to combine the full document into one that trips the threshold. Flash-Lite does not have a long-context surcharge — it stays at $0.25/$1.50 regardless of context length up to its 1M-token limit (Google AI pricing page, September 2026).
Prompt caching at 90% off
Gemini context caching is the cheapest in the family at $0.025-$0.125 per million cached tokens, depending on model and context length.
Gemini 2.5 Pro cached input is $0.125 per million tokens at short context and $0.25 at long context — a 90% discount off the standard input rate. Gemini 2.5 Flash cached input is $0.03 per million tokens for text, image, and video workloads and $0.10 for audio. Gemini 3.1 Flash-Lite is the cheapest at $0.025 per million cached tokens (Google AI pricing page, September 2026).
Storage is billed separately at $4.50 per million tokens per hour for Pro and $1 per million tokens per hour for Flash. For a 100,000-token system prompt cached for an hour, that adds $0.45 (Pro) or $0.10 (Flash) to the bill regardless of cache hit volume. The economics work best for high-reuse prefixes (10x+ per hour); below that, the storage cost dominates the cache-read savings (Google AI pricing page, September 2026).
Grounding: free tier and per-prompt costs
Grounding with Google Search and Maps has a generous free tier, then charges per grounded prompt above it.
Google Search grounding on Gemini API is free for the first 1,500 grounded prompts per day on paid Gemini API access, then $35 per 1,000 additional grounded prompts. Google Maps grounding is free for the first 10,000 grounded prompts per day, then $25 per 1,000 additional grounded prompts. The free tier is shared between Flash and Flash-Lite and resets daily (Google AI pricing page, September 2026).
Vertex AI charges $2.50 per 1,000 requests for custom-data grounding — significantly more than the $35 per 1,000 standard Search grounding. Custom-data grounding is required when the model needs to query a private corpus (a Vertex AI Search index, BigQuery table, or other Google Cloud data source). For pure Google Search grounding, the Gemini Developer API is cheaper than Vertex AI (Google AI pricing page, September 2026).
What enterprise buyers should do next
Three actions for organisations evaluating Gemini API pricing in 2026.
- Set up a guardrail at the 200K threshold. Either chunk requests under the threshold or accept the long-context rate. Most production systems need explicit code to enforce this — Gemini 2.5 Pro's 2M context window tempts teams to combine documents that would have been cheaper split.
- Audit prompt-caching strategy. A 100K-token system prompt reused 10x per hour costs $0.45/hour on Pro and $0.10/hour on Flash in storage. Below 10x reuse, the storage cost outweighs the cache-read savings. Tune storage TTL to actual reuse patterns.
- Plan grounding volume. Search grounding is essentially free for low-volume apps (1,500 RPD covers most production RAG pipelines). At higher volumes, $35 per 1K grounded prompts adds up — for a 100,000-prompt/day RAG system, grounding adds $3,500/month.
What to watch next
Three near-term datapoints. First, the Gemini 3.0 Pro general availability — Google published Gemini 3.0 Pro pricing at $1.25/MTok in Q2 2026 but has not yet confirmed GA for the third-generation model. Second, Vertex AI's introduction of separate fine-tuning pricing for Gemini 3.x — fine-tuning is currently bundled into per-token training costs that are not always published. Third, the rate of change on the Flash-Lite line — Google typically refreshes Flash-Lite every 6-9 months; a Flash-Lite 3.5 or 4.0 is plausible for Q1 2027 with a new price point.