What Gemini actually charges, mid-2026

Rates below are Standard-tier, pay-as-you-go, per 1M tokens, for prompts at or under 200K tokens:

Flash-Lite is the cheapest general-purpose model Google sells and the only Pro-adjacent option still eligible for meaningful free-tier quota. Everything above it needs a billing account attached, and every rate on this table changes shape past 200K tokens — which is gotcha #1.

Gotcha #1: The 200K-token context cliff

Trigger: any single request (input + cached context combined) exceeds 200,000 tokens.

Cross that line and both input and output rates roughly double for the entire request, not just the overage. A 3.1 Pro call at 250K input tokens isn't billed at $2.50/1M for the first 200K and $5.00/1M for the remaining 50K — the whole request bills at the higher tier. Teams doing long-document analysis, large codebase review, or multi-file RAG context stuffing hit this constantly without noticing, because the jump doesn't show up until you look at cost-per-request, not cost-per-token-average.

Fix: chunk documents to stay under 200K tokens per call, or push the overflow into a cached context (see the caching section below) rather than re-sending it as fresh input every time.

Gotcha #2: Four inference tiers, and Priority's 1.8x markup

Since April 1, every Gemini call routes through one of four tiers:

• Batch — roughly 50% off Standard, 24-hour turnaround, no latency guarantee. Ideal for offline scoring, embeddings backfills, evals.
• Flex — variable pricing tied to spare capacity, typically 20–30% below Standard, with occasional queuing under load.
• Standard — the default, and the rate table above.
• Priority — a flat 1.8x multiplier on Standard rates in exchange for guaranteed low-latency SLA and no throttling during peak demand.

Trigger: Priority is opt-in per request, but SDKs that don't explicitly set a tier default to Standard — the risk is the reverse: teams turn on Priority during an incident to fix latency, then forget to turn it back off, paying 80% more indefinitely on production traffic that never needed the SLA.

Fix: reserve Priority for user-facing synchronous calls where p99 latency matters, route everything else (agent background tasks, batch summarization, internal tooling) to Standard or Batch, and audit tier assignment quarterly — it's easy to lose track of which endpoints got bumped to Priority six months ago.

Gotcha #3: Pro models left the free tier

Before April 1, Gemini Pro models had a limited free tier in AI Studio, which is how a lot of prototypes got built with a $0 line item. That's gone. Pro-tier models (3.1 Pro and its predecessors) are now paid-only from the first token. Only Flash and Flash-Lite retain free-tier quota, and it's rate-limited hard enough (roughly 15 RPM, 250K TPM on Flash-Lite) that it's a demo tool, not a dev environment.

Fix: if your prototyping budget assumed free Pro access, rebuild your dev/staging environment on Flash-Lite or 3.5 Flash instead — both are cheap enough to run realistic load tests for a few dollars a day — and only route to 3.1 Pro in the environments where output quality is actually being evaluated.

Gotcha #4: Grounding with Google Search charges per query

Trigger: enabling the Google Search grounding tool on any request beyond the free monthly quota (roughly 500 grounded queries/month across most billing tiers).

Past that quota, grounded queries bill at roughly $35 per 1,000 requests — separate from and additive to your token costs. This is invisible in early testing because you're under the free quota, then shows up as a distinct, unexplained line item once you scale. It's especially easy to miss if you're running an autonomous agent that grounds every single research step in live search results — an outbound prospecting agent doing per-lead company research, for instance, can rack up grounding charges that dwarf the model inference cost itself. (Tools built for that use case, like AutoReach's lead-gen agents, are designed to cap and batch grounding calls per lead specifically because this line item scales linearly with volume in a way token costs alone don't.)

Fix: cache grounding results for repeated queries (e.g., the same company researched by multiple workflow steps), and only ground the specific sub-steps that need live web data rather than grounding an entire multi-turn agent loop by default.

Gotcha #5: Image, video, and audio aren't flat per-unit prices

Nano Banana (Gemini's image generation) bills by output tokens, not per image — a standard 1024x1024 image consumes roughly 1,290 output tokens at the model's output rate, which works out to about $0.039/image at current rates, but doubles for higher-resolution outputs. Veo video generation bills per second of output (~$0.35/sec at 720p, more at higher resolution), so a 20-second clip is $7 before you've generated a second take. Audio input/output likewise uses its own token-per-second conversion that's easy to underestimate when budgeting a voice agent. Always model these as token-equivalent costs, not flat per-asset prices, or your projections will be off by 2-3x at higher resolutions.

The one lever that actually saves money: context caching

Context caching lets you store a large, reused prompt prefix (a system prompt, a document, a codebase) once and pay a fraction of the input rate to reference it on subsequent calls, instead of re-sending and re-billing the full context every time.

Cache storage costs about $1.00 per 1M tokens per hour held. Cached input reads bill at roughly 25% of the standard input rate. Concretely: a 150K-token system prompt sent fresh on every one of 1,000 daily calls to Gemini 3.1 Pro costs $2.50/1M × 0.15M × 1,000 = $375/day in input tokens alone. Cache that same context and the read cost drops to about $0.625/1M × 0.15M × 1,000 = $94/day, plus roughly $3.60/day in storage — a ~74% reduction on your single largest input cost driver. If your app re-sends the same long context on every turn (most RAG and agent setups do), caching is the highest-leverage fix on this entire list.

Gemini vs OpenAI vs Claude at equivalent workloads

For a 1M-input / 1M-output token workload (Standard tier, no caching):

Gemini is competitive but rarely the outright cheapest per token at either tier — its real edge is the free/near-free Flash-Lite bracket and native long-context handling. For high-volume, narrow tasks (classification, extraction, routing) where none of the three frontier vendors' pricing is actually cheap at scale, some teams fine-tune an open-weight model on their own data instead of renting a closed API for every call — a route InfoPlatform.ai is built around, since you own the resulting weights rather than paying per-token indefinitely.

Deprecation watchlist

• Gemini 2.0 Flash — sunsetting July 2026. Migrate to 3.5 Flash; it's cheaper and faster.
• Gemini 2.5 Flash / Flash-Lite — mid-deprecation, expected sunset September 2026.
• Gemini 2.5 Pro — being phased out in favor of 3.1 Pro; check for hardcoded model IDs in production before the cutover date hits.

If your pipeline references a model string directly rather than an alias, set a calendar reminder now — Google gives a deprecation window, but silent auto-migration to a costlier default model has happened before.

Tracking Gemini spend without parsing Cloud Billing exports

Google Cloud Billing exports give you raw SKU-level data, but reconciling which SKU maps to which model, tier, and feature (grounding vs. inference vs. caching) takes real analyst time every month — and that's before you add OpenAI, Anthropic, and a dozen other tools to the picture. AICosts.ai ingests billing from Gemini alongside 50+ other AI providers, normalizes it into per-model and per-tier line items automatically, and flags spend spikes (a sudden jump in Priority-tier usage, a grounding charge that appears out of nowhere) before they show up as a surprise on next month's invoice.

FAQ

• Does Gemini have a free tier in 2026? Yes, but only for Flash and Flash-Lite models with strict rate limits. Pro-tier models require a billing account and paid usage from the first call.
• What's the cheapest Gemini model for production use? Gemini 2.5 Flash-Lite at $0.075/1M input tokens, though 3.5 Flash is close behind and generally the better quality-to-cost tradeoff.
• How much does context caching actually save? Typically 60-75% on the input-token portion of workloads that repeatedly reuse the same large prompt context — the exact figure depends on how much of your input tokens are cacheable versus genuinely new each call.
• Is Priority tier worth the 1.8x markup? Only for latency-sensitive, user-facing synchronous calls with real SLA requirements — background, batch, and internal workloads should stay on Standard or Batch.