For a US developer, the normal path is straightforward: create a credential in Google AI Studio, confirm that it belongs to a Cloud project you control, and enable billing only when the Free Tier no longer fits. You do not buy the key string. If your account is assigned to Prepay, the current Paid upgrade asks for at least $10 in credits; that is spendable API balance, not a key fee or Gemini consumer subscription.
Before sending production traffic, estimate one real workload and reconcile one small request against the same project's usage and billing screens. That single check catches the most expensive category error: creating a key in one project while funding, monitoring, or rate-limiting another.
The official US setup: six checks in order
- Open AI Studio and choose the project. Create or import a Cloud project owned by your organization, not a contractor or marketplace seller.
- Create the current credential type. Google's key documentation says new AI Studio keys are auth keys. Record the project ID without recording the secret itself.
- Keep the first call server-side. Put the key in a secret manager or server-only environment variable, then send a short, nonsensitive request with a fixed model and low output ceiling.
- Decide whether Free is sufficient. Check the model, RPM, TPM, and RPD shown for your account. Limits apply per project, so a second key in that project does not create a second quota pool.
- Link billing only to the intended project. If Paid is required, confirm the billing account, assigned plan, active tier, and balance in the current UI.
- Reconcile before scaling. Match the request time, returned model, input/output usage, and resulting charge. Then test how to stop, restrict, or rotate the credential.
This sequence separates three objects that many purchase guides collapse: the key authenticates, the project owns usage and ordinary request limits, and the billing account owns the payment relationship and usage tier.
What the first $10 does—and does not buy
Google's current Gemini API billing guide defines the Free-to-Paid change as linking an active billing account. New users currently default toward Prepay, although the plan assigned during the rollout can vary by account. On the Prepay path, $10 is the documented minimum credit purchase.
Those credits are deducted near real time, expire after 12 months, and are generally nonrefundable except under Google's documented Postpay-switch rule. When the balance reaches zero, every API key in projects linked to that billing account stops working; the project does not automatically fall back to Free. Billing can lag enough to permit a limited overage, so a zero balance is not a perfect hard cap. Use application-level budgets, bounded retries, and concurrency limits too.
The spend thresholds are account mechanics, not prices for better keys. Tier 1 requires active billing. Tier 2 normally requires $100 in successful paid spend plus three days from the first successful payment; Tier 3 normally requires $1,000 plus 30 days. Meeting a threshold does not guarantee approval or capacity. The focused rate-limit guide and Tier 3 guide own the detailed upgrade decision.
Price the workload before you turn on Paid
The most useful estimate is not “one million tokens.” It is a request shape multiplied by expected traffic. Consider a small US SaaS feature with 10,000 monthly requests, averaging 800 text input tokens and 200 output tokens per request. That produces 8 million input tokens and 2 million output tokens. Assume Standard service, no prompt over 200k, and no caching, grounding, media, or tools.
monthly token cost = 8 Ă— input rate + 2 Ă— output rate
Using the official prices checked July 18, 2026:
- Gemini 3.1 Flash-Lite:
8 Ă— $0.25 + 2 Ă— $1.50 = $5.00. - Gemini 3.5 Flash:
8 Ă— $1.50 + 2 Ă— $9.00 = $30.00. - Gemini 3.1 Pro Preview:
8 Ă— $2.00 + 2 Ă— $12.00 = $40.00, provided each prompt stays at or below 200k.
Eligible Batch token rows halve those token charges to $2.50, $15.00, and $20.00, but Batch only fits work that can accept its asynchronous execution model. The calculation excludes taxes, retries, cache storage, grounding, tools, audio, images, video, Flex, Priority, and any independent-provider charge. Treat it as a reproducible planning case, not a bill quote.
Read the rate with all of its labels
On the same check date, the representative Standard text-oriented rows were:
- Gemini 3.1 Flash-Lite, Stable: $0.25 input and $1.50 output per 1 million tokens; a Free Tier row is shown.
- Gemini 3.5 Flash, Stable: $1.50 input and $9.00 output; a Free Tier row is shown.
- Gemini 3.1 Pro Preview: $2.00 input and $12.00 output at prompts up to 200k; above 200k, $4.00 and $18.00; no Free Tier row is shown.
Audio, images, Live, TTS, grounding, caching storage, and other serving lanes use different units or rows. A search snippet that displays another number may have surfaced Priority, another modality, or another context band. The current model index and pricing page must be read together immediately before deployment.
Stay Free, enable Paid, or move to a governed Cloud route?
| Decision signal | Appropriate next step | Proof required before production |
|---|---|---|
| Prototype traffic fits the displayed Free limits and contains no data that requires Paid terms | Stay Free and measure actual tokens | Project ID, active model, account-specific limits, usage record |
| Production needs a paid-only model, higher operating range, or Paid data terms | Link billing to the intended Developer API project | Assigned plan, tier, balance, first reconciled charge, stop procedure |
| IAM, regional controls, procurement, audit, or dedicated enterprise governance is the real requirement | Design the applicable Google Cloud or enterprise contract | Product surface, identity method, region, quota, pricing, data and support terms |
For the current no-cost model and quota boundaries, use the separate Gemini API Free Tier guide rather than copying a universal RPM or RPD number into a production plan.
A consumer Gemini subscription does not establish that a Gemini Developer API project has credits. Likewise, a Cloud or enterprise identity can use a different authentication and pricing contract. Do not merge those receipts into one “Gemini subscription” budget.
The direct route still has a short eligibility check: confirm the current available-regions list, age 18+ requirement, account, and payment eligibility. Do not falsify location, identity, or billing information. This check is part of account setup, not a reason to shop for someone else's raw key.
An independent provider is a different purchase
An API provider may be useful when an OpenAI-compatible client, one multi-model account, or a separately negotiated billing and support arrangement solves a concrete operational problem. It should issue its own token for its own endpoint. Its model mapping, balance, quota, logs, data use, invoice, refund, and support rules belong to the provider contract—not Google's.
For example, LaoZhang AI's current getting-started documentation publishes the OpenAI-compatible base URL https://api.laozhang.ai/v1 and uses a LaoZhang-issued credential. Its Models API is the runtime check for available routes. The token is not a Google Gemini API key, and this page makes no promise that its current price, availability, or uptime will match Google direct.
Before depositing more than canary money, answer all of these in writing:
- Who issues, restricts, rotates, and revokes the token?
- Which exact base URL and protocol accept it, and can an alias change upstreams?
- How are errors, timeouts, stream interruptions, and retries billed?
- What are the live model list, concurrency limits, log retention, data use, and deletion controls?
- Who owns invoices, refunds, incidents, escalation, and the migration path out?
If any owner is unknown, stop. A marketplace seller handing over a raw Google key does not become a legitimate provider merely because the key returns a response.
Failure modes that another key will not fix
- 401/403: diagnose key type, restriction, project, API permission, and IAM. Use the permission guide; adding credit is not a permissions fix.
- 429: inspect the active project's RPM, TPM, RPD, model dimension, and retry behavior. More keys in that project share the same limits.
- Balance zero: top up only if intended; otherwise keep traffic stopped. Do not assume Free fallback.
- Unexpected charge: stop retries, record request IDs and usage metadata, and reconcile the model, lane, project, and account before resuming.
- Possible leak: revoke or rotate at the actual issuer, inspect usage and billing, then remove the secret from code, logs, screenshots, and client bundles.
Google also says unrestricted standard keys are already rejected and plans to reject all standard keys in September 2026. That deadline is volatile and applies to the Gemini Developer API transition, not every Google Cloud credential. Recheck the official key page before publication and migration.
Data terms are part of the cost decision
Under the current Gemini API Additional Terms, unpaid-service content may be used to improve Google products and may be human reviewed after protective processing. Prompts and responses submitted to Paid Services are not used to improve Google's products, although limited safety, security, and operational logging remains. This is not legal advice; review the terms applicable to your organization and data.
The purchase decision is complete only when you can name the credential issuer, project/account owner, endpoint, billing owner, quota owner, model, and stop path. For a US developer using Google direct, that begins with a free project credential and ends with a reconciled usage charge—not a key purchase.
