Quick verdict
Gemini pricing is a stack of meters on two surfaces: the Gemini Developer API (ai.google.dev/pricing) and Vertex AI / Gemini on Google Cloud (Vertex generative AI pricing). If you only compare “$/1M tokens,” you will under-budget production. The short version:
- High-volume / simple turns: prefer Flash or Flash-Lite. They are the cheap path for classification, routing, extraction, and most chat when eval quality still passes.
- Hard reasoning / frontier quality: use Pro, but watch the long-context tier. Prompts above ~200K tokens jump to a higher input/output band on Pro-class models.
- Repeated large prefixes: use context caching (discounted cache reads + hourly storage). Stable system prompts, tool schemas, and document prefixes repay the cache when hit rate is high.
- Offline jobs: use Batch where published (often about 50% off standard token rates). Do not put interactive chat on Batch.
- Grounding with Google Search: Gemini 3-style free tier is roughly 5,000 prompts/month, then about $14 per 1,000 search queries. Grounding can exceed model tokens on search-heavy assistants.
- AU/SG cutover: pick Region + grounding policy before you fall in love with a model sticker. Wrong residency or always-on Search grounding is more expensive to unwind than picking Flash vs Pro wrong once.
Cipher’s first-hand cutover: treat Gemini cost as model tokens + cache storage + (optional) grounding queries + platform choice (API vs Vertex). Teams that only track token dashboards are usually surprised by Search grounding once every turn hits the web.
This post is the cost companion to Amazon Bedrock vs Google Vertex AI (2026). Platform footprint first; pricing shape second.
Last updated: August 2026. Rates change; re-verify ai.google.dev and Vertex generative AI pricing before you publish a budget externally.
Gemini API vs Vertex: what actually differs
Per-token model rates are often close across the Developer API and Vertex for the same Gemini family. The bill and the operating model still diverge:
| Dimension | Gemini Developer API | Vertex AI (Gemini on GCP) |
|---|---|---|
| Who buys it | API key / Google AI billing | GCP project, Cloud Billing, often enterprise agreements |
| Identity & network | API keys; simpler startup | IAM, VPC-SC, private endpoints, org policy |
| Residency controls | Limited vs full GCP Region design | Region / location selection, data residency patterns on GCP |
| Ops fit | Fast prototypes, product APIs, solo and startup stacks | Enterprise IAM, audit, BigQuery gravity, shared GCP estate |
| Cost shape | Model + cache + grounding (+ Batch) | Same family of meters, plus GCP project/SKU packaging and enterprise discounting |
Cipher rule: if the customer already lives on GCP (BigQuery, IAM, VPC), default to Vertex for production. If they are AWS-centric, do not force Vertex for a sticker win; compare against Bedrock pricing and the platform post above. Token parity is not a reason to split the security review across two clouds.
Directional model rates (re-verify)
Order-of-magnitude paid-tier figures from Google’s published Gemini API pricing as of August 2026. Always confirm the live tables.
| Model class | Input / 1M | Output / 1M | Planning note |
|---|---|---|---|
| Flash-Lite (e.g. Gemini 3.1 Flash-Lite) | ~$0.25 | ~$1.50 | Cheapest high-volume path for simple agentic tasks, translation, light extraction |
| Flash (e.g. Gemini 3.5 Flash) | ~$1.50 | ~$9.00 | Speed + quality middle; common production default when Lite fails evals |
| Pro ≤200K (e.g. Gemini 3.1 Pro) | ~$2.00 | ~$12.00 | Frontier reasoning / hard multimodal turns |
| Pro >200K context | ~$4.00 | ~$18.00 | Long-context rate jump: if input context exceeds ~200K, all tokens on that request often bill at the long-context band (confirm live docs) |
Batch: Google publishes Batch rates at roughly half of standard input/output for supported models (e.g. Flash ~$0.75 / $4.50; Pro ≤200K ~$1.00 / $6.00). Context cache reads are a steep discount to standard input (often ~90% off on recent Gemini families for explicit cache hits), plus storage about $1 / 1M tokens / hour on Flash-class and about $4.50 / 1M tokens / hour on Pro-class. Treat every figure as directional until you re-check.
Context caching, Batch, and grounding
Context caching
Cache large, stable prefixes (system instructions, tool schemas, fixed document sets). You pay to create/store the cache and a discounted rate when requests hit it. Implicit caching may help automatically; explicit caching gives control and predictable discounts. Delete or shorten TTL when the prefix goes stale so you do not pay storage for dead context.
Batch
Use Batch for evals, enrichment, and overnight rewrites. Same quality class at published discount; latency can stretch (often toward ~24h-class batch SLAs; confirm current docs). Interactive assistants stay on standard (or Priority/Flex if you explicitly buy those modes).
Grounding with Google Search
This is the silent second meter. For Gemini 3-style grounding, Google has published roughly 5,000 free prompts per month (shared across Gemini 3), then about $14 per 1,000 search queries. One user prompt can trigger multiple Search queries. On a search-heavy product assistant, grounding can exceed model token spend. Cipher’s rule: turn grounding on only for intents that need the live web; do not default “ground everything.”
Worked shape: three monthly scenarios (directional)
Planning shapes, not quotes. Plug your mix into the live calculators.
| Scenario | Traffic shape | Dominant meters | Planning note |
|---|---|---|---|
| Internal pilot | Tens of thousands of turns; Flash + Flash-Lite mix; no Search grounding | On-demand tokens | Stay on Developer API or a single Vertex project; optimize prompt size before Pro |
| Customer assistant + RAG + cache | Hundreds of thousands of turns; large stable system prompt; retrieved chunks vary | Tokens + context cache storage/reads | Measure cache hit rate weekly; Pro only for escalations. Compare total to Bedrock KB path in Bedrock vs Vertex |
| Search-grounded research assistant | Many turns call Grounding with Google Search; Pro for synthesis | Grounding queries + Pro tokens (+ long-context if docs are huge) | Model the $14/1k search band after the free tier. Grounding often dominates before token stickers matter |
AU/SG cutover vs Bedrock
For Australian and Singapore teams, Cipher’s order is deliberate:
- Region / residency: Confirm where prompts, responses, and logs may live. Vertex Region choice and GCP org policy beat a cheap Flash sticker that forces the wrong geography.
- Grounding policy: Decide which intents may call Search (or Maps). Always-on grounding is a product decision with a hard meter, not a default checkbox.
- Platform footprint: AWS-heavy estates usually stay on Bedrock unless Gemini-specific quality or BigQuery gravity wins. GCP-heavy estates usually stay on Vertex. See the full matrix in Bedrock vs Vertex AI.
- Model tier: Only after the three above, pick Flash-Lite / Flash / Pro and whether Batch or cache applies.
For product-level monthly bands beyond Google meters alone, see What it actually costs to run an AI product in 2026.
Choose Gemini / Vertex if…
- Choose Flash-Lite if volume is high, tasks are simple, and your eval set still passes.
- Choose Flash if you need a stronger default without Pro’s sticker and long-context jump.
- Choose Pro if hard reasoning or multimodal quality fails on Flash, and you have capped when >200K context is allowed.
- Choose context caching if a large prefix is stable across many turns and you will measure hit rate.
- Choose Batch if the job is offline and the model is on the Batch discount list.
- Choose Vertex over the Developer API if you need enterprise IAM, VPC, residency, or GCP billing alignment.
- Choose Bedrock instead if your data, IAM, and ops already live on AWS and Gemini is not a hard requirement. Re-price on Bedrock pricing.
Related: Bedrock vs Vertex AI (2026) · AWS Bedrock pricing · Cost breakdown: running AI products · Applied AI Engineering
FAQ
Is Vertex more expensive than the Gemini API for the same model? Often not on raw $/1M tokens. Vertex’s value is IAM, VPC, residency, and GCP billing. Compare TCO and compliance fit, not only stickers.
When does the long-context Pro rate kick in? Google publishes a higher band when prompt input context exceeds about 200K tokens on Pro-class models (e.g. ~$4 / $18 per 1M vs ~$2 / $12). Confirm whether the whole request bills at the long-context rate on the live page.
Can grounding cost more than the model? Yes. After the free monthly grounding allowance, Search queries at about $14 per 1,000 can dominate on web-heavy assistants. Gate grounding by intent.
How should Australian teams choose between Vertex and Bedrock? Lock Region and cloud footprint first, then grounding policy, then model tier. Platform follows data; see Bedrock vs Vertex.
Who can help model a realistic Gemini / Vertex budget? Cipher Projects is an Australian-led AI engineering studio that designs Gemini and multi-cloud AI architectures for AU/SG teams, including cache, grounding policy, and residency under accounts you control.
Conclusion
Gemini and Vertex pricing rewards teams that budget in layers: model tier, long-context band, cache, Batch, and grounding queries, with API vs Vertex chosen for IAM and residency rather than a one-cent sticker gap. Pick Region and grounding policy before Pro, measure cache hits before you scale Flash traffic, and only then compare the bill to Bedrock. That is how you avoid both under-powered pilots and search-grounding surprises.