Contents
- Verdict
- API pricing
- What Gemini costs on two real automations
- Gemini caching is priced differently, and it can lose you money
- How Gemini behaves on agent workloads
- The free tier trap, and how to avoid it
- Which businesses Gemini genuinely suits
- How to access
- Privacy and residency
- What we build
- FAQ
- Price sources
Gemini is the best-value US option at the low end and the obvious default if your business already lives in Google Workspace. Flash-Lite lists at $0.30 in and $2.50 out per 1M tokens as of August 2026, and the older 2.5 Flash-Lite is cheaper still at $0.10 and $0.40, which puts genuinely useful automation within a few dollars a month for most small businesses. The one thing to get right before you build: the free tier trains on your data and the paid tier does not. Never prototype a business workflow with real customer data on the free tier.
Best for
Google Workspace businesses, high-volume cheap classification and extraction, and anything that needs to read documents, images and video in the same pipeline.
Not for
Anyone who needs a stable, clearly documented model lineup, since the tier naming and availability shift often and capacity details are scattered across per-model pages rather than the pricing page.
Google Gemini API pricing per 1M tokens
Google prices per million tokens with three discount routes: Batch mode at 50 percent off, a Flex tier at the same rate as batch for latency-tolerant synchronous traffic, and explicit context caching, which is priced differently from every other provider on this site. Google charges a reduced per-token rate to read cached content plus an hourly storage charge of $1.00 per 1M tokens held, rising to $4.50 on 2.5 Pro. That storage clock is the thing to model, because a cache that sits idle still bills. Context windows are not published on the pricing page, but they are published on the individual model pages: gemini-3.6-flash and gemini-3.5-flash-lite are both documented at 1,048,576 input tokens and 65,536 output tokens. Where we could not confirm a figure for a specific model we say so rather than assume the family shares it. Prices below are list prices as of August 2026.
| Model | Context | Input / 1M | Output / 1M |
|---|---|---|---|
| gemini-3.6-flashBatch and Flex $0.75 / $3.75; context caching $0.15 plus $1.00 per hour storage | 1,048,576 in / 65,536 out | $1.50 | $7.50 |
| gemini-3.5-flashBatch and Flex $0.75 / $4.50; context caching $0.15 plus storage | Not published on the pricing page | $1.50 | $9.00 |
| gemini-3.5-flash-liteBatch and Flex $0.15 / $1.25; context caching $0.03 plus storage | 1,048,576 in / 65,536 out | $0.30 | $2.50 |
| gemini-3.1-flash-liteBatch and Flex $0.125 / $0.75 | Not published on the pricing page | $0.25 | $1.50 |
| gemini-2.5-proBatch $0.625 / $5.00; context caching $0.125 plus $4.50 per hour storage | Not published on the pricing page | $1.25 | $10.00 |
| gemini-2.5-flashBatch $0.15 / $1.25 | Not published on the pricing page | $0.30 | $2.50 |
| gemini-2.5-flash-liteGoogle's cheapest listed tier, though OpenAI's gpt-5-nano is half the input price; batch $0.05 / $0.20 | Not published on the pricing page | $0.10 | $0.40 |
| gemini-embedding-2Embeddings; batch $0.10. gemini-embedding-001 is cheaper at $0.15 | n/a | $0.20 | n/a |
List prices as of August 2026, taken from the provider's published pricing page on August 5, 2026. Model prices move fast - check the source links before you budget.
What Gemini costs on two real automations
We price every provider on this site against the same two workloads. Workload A is an inbox agent: 1,500 runs a month at about 3,000 input tokens each, of which 2,400 are a stable system prompt and tool schema, and 400 output tokens. Workload B is document extraction: 4,000 documents a month at 6,000 input and 300 output tokens each. That comes to 4.5M input and 0.6M output for A, and 24M input and 1.2M output for B.
On gemini-2.5-flash-lite at $0.10 in and $0.40 out per 1M tokens, Workload A costs about $0.69 a month and Workload B about $2.88, dropping to about $1.44 in batch mode. That is the cheapest list pricing of any US major and it is not close. On gemini-3.5-flash-lite at $0.30 and $2.50, the same two workloads cost about $2.85 and $10.20, or $5.10 batched. On gemini-3.6-flash at $1.50 and $7.50 they are about $11.25 and $45.00, or $22.50 batched. On gemini-2.5-pro at $1.25 and $10.00, about $11.63 and $42.00.
The practical read is that at small-business volumes Gemini removes cost as a decision input almost entirely at the Flash-Lite end. When an automation costs under three dollars a month to run, the only thing worth arguing about is whether it works, which is a much healthier conversation than a per-token comparison. It also means you can afford to run a second model over the same inputs as a cross-check, which is a genuinely useful pattern for extraction work where a silent error is expensive.
Two structural notes before you budget. The Flex tier prices synchronous requests at the same rate as batch, so latency-tolerant live traffic gets the discount without going asynchronous, which is a real advantage over providers where batch is the only route to half price. And the pricing page publishes no context windows at all, so capacity planning means opening the individual model page: gemini-3.6-flash and gemini-3.5-flash-lite are both documented at 1,048,576 input tokens and 65,536 output tokens, and we did not confirm the figure for every older model on the list.
Gemini caching is priced differently, and it can lose you money
Every other major provider prices prompt caching purely per token. Google does not. Explicit context caching charges a reduced per-token rate to read cached content, typically a tenth of the input rate, plus an hourly storage charge for holding the cache: $1.00 per 1M tokens per hour on the Flash and Flash-Lite tiers, and $4.50 on 2.5 Pro. That storage clock runs whether or not anyone calls the model, and it changes the arithmetic completely.
Work it through on Workload A with gemini-3.5-flash-lite. The 2,400-token system prefix cached continuously costs 0.0024M tokens times $1.00 an hour times roughly 730 hours, which is about $1.75 a month in storage alone. The saving from caching those 3.6M monthly input tokens is the difference between $0.30 and $0.03 per 1M, which is about $0.97. You have paid $1.75 to save $0.97. Explicit caching on this workload makes the bill worse.
The break-even is a clean number and it is worth memorising, because it does not depend on how big your prompt is. Dividing the hourly storage rate by the per-token saving gives roughly 2,700 calls a month on gemini-3.5-flash-lite, about 540 a month on gemini-3.6-flash where the per-token saving is larger, and about 2,900 a month on gemini-2.5-pro where storage costs four and a half times more. Below those volumes, on a continuously held cache, do not use explicit caching. Above them, it pays and keeps paying.
The alternative is free. Implicit in-memory caching applies automatically with a 24-hour TTL, costs nothing to store, and Google documents it as compatible with a zero-retention arrangement because it is RAM-only and project-isolated. For most small-business automations that is the right answer: structure the prompt so the stable material sits at the front, let implicit caching do what it can, and only reach for explicit caching when you have measured a workload above the break-even volume.
Batch mode is the simpler lever and it is a flat 50 percent off. The Flex tier gets you the same rate synchronously if your workflow can tolerate variable latency, which suits a background enrichment job that runs on a cron trigger but still wants an answer in the same request.
How Gemini behaves on agent workloads
Gemini's distinctive strength for automation is not reasoning, it is intake. Real business inputs are photographed invoices, scanned delivery notes, voicemail recordings and screen-recorded walkthroughs, and Gemini takes all of those natively in the same call as text. Collapsing a transcription tool, an OCR tool and a model into a single call removes two integration points and two failure modes. On maintenance burden alone that is often worth more than a per-token difference.
The published research on long-horizon agents applies here as everywhere. LOCA-bench, released in early 2026, grows agent context under controlled conditions while holding the task fixed and reports that most models including frontier ones degrade sharply as context grows. A 1,048,576-token input limit is a ceiling to design away from, not a target. AgentCE-Bench, from April 2026, found that injecting tool-call failures at just a 10 percent rate visibly degraded agent performance across every horizon tested, which is an argument for retry and backoff and idempotency on writes regardless of which model you pick.
Where we would be cautious with Gemini specifically is long tool-calling chains under a single model. The lineup is optimised around fast, cheap, high-throughput execution, and the generally available Pro-class option as of August 2026 is still 2.5 Pro while Gemini 3.1 Pro sits in preview. If your workflow genuinely needs a dozen dependent tool calls with recovery from partial failure, run your own twenty-case eval against a flagship from another vendor before committing rather than taking either vendor's word for it.
We publish no benchmark scores. Agent leaderboards turn over monthly, vendors quote the ones that flatter them, and none of them measure whether a model can triage your inbox against your business rules. Twenty real cases from your own business, graded by someone who knows the right answers, settles the question in an afternoon and stays valid longer.
The free tier trap, and how to avoid it
Google's free Gemini API tier is genuinely generous and genuinely unsuitable for business data. Google's own API terms say that on Unpaid Services it uses the content you submit and the responses generated to provide, improve and develop Google products and services, and that human reviewers may examine API input and output, with the data disconnected from your account before review. On Paid Services, Google states it does not use prompts or responses to improve its products. This is not a subtle distinction and it is the most common mistake we see in small business AI projects.
The failure mode is predictable: someone prototypes an automation on the free tier using real customer emails, it works, and it quietly goes into production without anyone revisiting the billing setup. By the time it is noticed, months of customer data have been processed under the wrong terms. If you take one thing from this page, make it this: attach billing and confirm you are on the paid tier before a single real record goes through.
Beyond that, the retention picture has some edges worth knowing, and to Google's credit they are documented rather than buried. Search grounding and Maps grounding data is retained for 30 days with no opt-out, so an automation that grounds against web search cannot claim a zero-retention posture on those calls. Files uploaded through the File API persist until deleted, which makes your cleanup routine part of your compliance story rather than an optimisation. Explicit context caching retains content for the TTL you set. And two APIs are stateful by default: the Interactions API stores state unless you pass store as false, and the Live API keeps conversation state for up to 24 hours unless you disable session resumption.
Zero data retention is available on paid services by approval rather than as a self-serve toggle, and when approved Google states that all user content and identifiable metadata are cleared before logging. If you need it, start that conversation before you design the workflow, not after, because two of the exemptions above are features you might otherwise build on.
On residency we are going to say plainly what we could not confirm. The zero-retention documentation does not mention regional options or data residency at all and points to the enterprise agent platform for enterprise-grade controls. A Developer API data residency documentation path returned a 404 when we checked in August 2026. If keeping data in a specific country is a contractual requirement, treat it as an open question to put to Google in writing rather than an assumed capability.
Which businesses Gemini genuinely suits
If your business runs on Google Workspace, Gemini has an advantage that has nothing to do with model quality: the data is already there. An automation that reads Gmail, writes to Sheets and files documents in Drive is dramatically simpler when the model, the identity layer and the storage are the same vendor. We have built the same workflow both ways, and the Google-native version has fewer moving parts and fewer things to break at 3am.
If your inputs are messy and physical - photographed invoices, scanned delivery notes, phone-recorded voicemails - Gemini's multimodal handling collapses what would otherwise be a three-tool pipeline into one call. That is a real reduction in maintenance burden, not just a convenience.
Where we would steer you elsewhere is when you need a stable, well-documented lineup you can build against for years. The naming has churned, the current-line Pro tier was still preview as of August 2026, and capacity limits are underdocumented. None of that is fatal, but it means more of your engineering attention goes to keeping up with the vendor than it would on a more conservative provider.
As always, the practical recommendation is a two-model stack. Gemini Flash-Lite is an excellent cheap first pass on almost any workload, whatever you use for the judgement steps. Wiring it in as the volume tier costs little and often halves a bill.
Strengths for agent workloads
- The cheapest list prices among the US majors, with 2.5 Flash-Lite at a fifth the cost of most mainstream budget models
- Native multimodal handling of documents, images, audio and video in one pipeline, which suits real-world business inputs
- Batch mode at 50 percent off, plus a Flex tier at the same discount for synchronous traffic that can tolerate variable latency
- Just over a million tokens of input context on the current Flash and Flash-Lite models, with 65,536 tokens of output
- Direct integration paths into Google Workspace, Sheets, Drive and Apps Script, which is where a lot of small business data already sits
- Zero data retention is documented in detail on paid services, including which features are exempt and why
- Implicit in-memory caching runs at a 24-hour TTL and is documented as compatible with ZDR because it is RAM-only and project-isolated
Weaknesses
- The free tier uses your content to develop Google products and human reviewers may examine input and output, which makes it unsuitable for business data
- The pricing page publishes no context windows at all, so capacity planning means opening a page per model
- Explicit context caching bills hourly storage on top of token rates, which can cost more than no caching on a low-traffic workflow
- Model naming and tier availability change frequently, and Gemini 3.1 Pro was still preview as of August 2026 with 2.5 Pro the generally available Pro-class option
- Search and Maps grounding retain data for 30 days with no opt-out, even under a ZDR arrangement
- The ZDR documentation does not address data residency, which is handled elsewhere on the enterprise platform
How to access Google Gemini
- Google AI Studio and the Gemini API for the direct path. Fastest way to test, and where the free tier lives - use it for experiments only, never with real customer data.
- Vertex AI on Google Cloud for production, where you get region selection, enterprise terms and IAM alongside the model.
- OpenRouter and similar aggregators for cross-provider comparison without a separate integration.
- Apps Script and Workspace add-ons for automations that live directly inside Sheets, Docs and Gmail, which is often the shortest path for a small business already on Google.
Data privacy and residency
- This is the single most important distinction on this page. Google's API terms state that on Unpaid Services it uses the content you submit and the generated responses to provide, improve and develop Google products and services, and that human reviewers may examine API input and output with the data disconnected from your account first. On Paid Services, Google states it does not use your prompts or responses to improve its products. A business workflow must run on the paid tier, and there is no ambiguity about it.
- On paid services, prompts and responses are logged temporarily for policy-violation detection and service security. Zero data retention is available on approval and covers prompts, responses, system instructions, cached content and files: when approved, Google states all user content and identifiable metadata are cleared before logging.
- Four things are documented as exempt from ZDR and are worth checking against your own workflow. Google Search grounding and Google Maps grounding both retain data for 30 days with no opt-out. Files uploaded through the File API are stored until you delete them. Explicit context caching retains content for whatever TTL you set. Two more need a configuration change rather than being exempt outright: the Interactions API stores state unless you set store to false, and the Live API retains conversation state for up to 24 hours unless you disable session resumption.
- Implicit in-memory caching runs on a 24-hour TTL and Google documents it as not violating ZDR, on the basis that it is RAM-only and isolated to your project. That is a useful detail, because it means the free automatic caching path is compatible with a zero-retention posture while the paid explicit caching path is not.
- On data residency we have to be honest about what we could not verify. Google's ZDR documentation does not address regional options or residency at all, and directs readers to the enterprise agent platform documentation for enterprise-grade controls. A documentation path for Developer API data residency returned a 404 when we checked in August 2026, and the enterprise locations page we reached was an API index that did not state the guarantees. If residency is contractual for you, get the specific commitment from Google in writing rather than relying on a summary, including ours.
- For anything covered by health, financial or legal confidentiality rules, minimise what you send and redact identifiers the automation does not need. Confirm billing is attached and you are on Paid Services before the first real record goes through, because the training distinction turns entirely on that.
What we build on Google Gemini
- High-volume document and image extraction where inputs arrive as scans, photos and PDFs rather than clean text
- Workspace-native automations: Gmail triage, Sheets enrichment, Drive document classification
- Meeting and call processing where audio goes in directly rather than through a separate transcription step
- Cheap first-pass classification in a two-model stack, escalating only the ambiguous cases
- Overnight batch processing of backlogs at roughly half price
Model choice is a line item, not a strategy. If you want the workflow designed, built and handed over with a plain-language report of what it ran and what it saved, start with a free AI automation audit or see the full done-for-you automation services.
Google Gemini FAQ
How much does the Gemini API cost?
As of August 2026, list prices run from $0.10 in and $0.40 out per 1M tokens on gemini-2.5-flash-lite, through $0.30 and $2.50 on gemini-3.5-flash-lite, to $1.50 and $7.50 on gemini-3.6-flash and $1.25 and $10.00 on gemini-2.5-pro. Batch mode is 50 percent off, and a Flex tier offers the same discount for latency-tolerant synchronous traffic. That low end is the cheapest list pricing among the US majors.
What would a real Gemini automation cost me per month?
A 1,500-run-a-month inbox agent using about 3,000 input and 400 output tokens per run costs roughly $0.69 a month on gemini-2.5-flash-lite, $2.85 on gemini-3.5-flash-lite and $11.25 on gemini-3.6-flash. A 4,000-document extraction job at 6,000 input and 300 output tokens each is about $2.88, $10.20 and $45.00 on the same three, halving in batch mode. At those numbers the build cost dwarfs the running cost.
Should I use Gemini context caching?
Only above a certain volume, because Google prices caching differently from everyone else: a reduced per-token read rate plus hourly storage of $1.00 per 1M tokens held, or $4.50 on 2.5 Pro. Held continuously, explicit caching breaks even at roughly 2,700 calls a month on gemini-3.5-flash-lite, 540 on gemini-3.6-flash and 2,900 on gemini-2.5-pro. Below that it costs more than it saves. Implicit in-memory caching is free, automatic, has a 24-hour TTL and is documented as ZDR-compatible.
What is the Flex tier?
Flex prices synchronous requests at the same rate as batch, currently 50 percent off standard, in exchange for tolerating variable latency. It is genuinely useful for background enrichment work that runs on a schedule but still wants an answer inside the request rather than through an asynchronous job. Most providers make you choose between full price and asynchronous batching, so this middle option is a real advantage.
Does Google train on Gemini API data?
It depends entirely on the tier. On the free tier, Google states that prompts and responses are used to improve products and human reviewers may see content. On the paid tier, they are not used for training. Any business automation must run on the paid tier, and you should confirm billing is attached before sending real customer data.
Is Gemini good enough for business automation?
For classification, extraction, summarisation and document processing, comfortably yes, and the cheap tiers make high-volume versions of those jobs affordable. Its particular strength is handling messy real-world inputs like scans, photos and audio natively in one call. For very long multi-step agent chains, test it against a flagship before committing.
Can I keep Gemini data in a specific country?
We could not verify that for the Gemini Developer API as of August 2026, and we are not going to assert it. Google's zero-retention documentation does not mention regional options or residency, and points to the enterprise agent platform for enterprise-grade controls. A Developer API data residency documentation path returned a 404 when we checked. If residency is contractual for you, ask Google for the specific commitment in writing before you build.
What is the Gemini context window?
The pricing page publishes no context windows at all, but the individual model pages do. As of August 2026, gemini-3.6-flash and gemini-3.5-flash-lite are both documented at 1,048,576 input tokens and 65,536 output tokens. We did not confirm figures for every older model on the pricing list, so check the specific model page for the one you intend to use rather than assuming the family shares a limit.
Which parts of Gemini are excluded from zero data retention?
Four are documented as exempt: Google Search grounding and Google Maps grounding, both retained 30 days with no opt-out; files in the File API, stored until you delete them; and explicit context caching, retained for the TTL you set. Two more are stateful by default rather than exempt: the Interactions API stores state unless you set store to false, and the Live API keeps conversation state up to 24 hours unless you disable session resumption.
Gemini or OpenAI for a small business?
If you already run on Google Workspace, Gemini usually wins on integration simplicity alone. If you use a wide range of no-code automation tools, OpenAI has the broader native support. On price, Gemini's low end is cheaper. On lineup stability and documentation, OpenAI is currently the more predictable vendor to build against.
Is there a Gemini 3 Pro model?
As of August 2026 the stable Gemini 3 line is Flash-only, with a Pro-tier preview available and Gemini 2.5 Pro remaining the generally available Pro-class option. If you need a Pro-tier model under production terms, 2.5 Pro is the one to build against for now.
