What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
iTechGuides is reader-supported. When you buy through links on our site, we may earn an affiliate commission. As an Amazon Associate I earn from qualifying purchases. Learn more
Google Cloud made token-based cost management generally available for BigQuery generative-AI functions on October 14, 2026. Administrators can now set daily project-wide and per-user limits for input, output, and model “thinking” tokens, giving teams a way to cap unexpectedly large bills before workloads scale across a dataset.
What BigQuery token quotas control
BigQuery’s SQL generative-AI functions can send large numbers of rows to a large language model and receive generated results. Because those requests are billed by tokens, a broad query can consume far more input or output than a user expects. The new controls add four independent daily quota dimensions. Google’s current cost-control guide lists these defaults:
| Quota | Default allowance | What it limits |
|---|---|---|
| Project input tokens | 200,000,000,000 per project per day | Total prompt/input tokens consumed by the project |
| Per-user input tokens | 150,000,000,000 per project, per user, per day | Prompt/input tokens attributable to one user in the project |
| Project output and thinking tokens | 20,000,000,000 per project per day | Generated output plus model reasoning (“thinking”) tokens for the project |
| Per-user output and thinking tokens | 15,000,000,000 per project, per user, per day | Generated output and thinking tokens attributable to one user |
These are daily defaults, not guaranteed prices or capacity for every model. Google can change quota values, so check the quota entry in your Cloud project before relying on a number for a budget or launch plan.
How administrators configure the limits
Quota management is a Google Cloud console function controlled by permissions in Identity and Access Management (IAM). An administrator can leave the defaults in place, request more capacity for a workload, or create a stricter override to enforce an internal spending ceiling.
#1 Best Overall
- Open the Google Cloud console and go to IAM & Admin.
- Select Quotas & System Limits.
- Filter the service list for the BigQuery API.
- Find the input or output/thinking token quota and choose the option to edit it.
- Enter the project-wide or per-user value your governance policy allows, then submit the change. Use a quota increase when a legitimate workload needs more capacity; use an override when you need a tighter cap.
The limits are aggregated globally across regions. Running jobs in multiple locations therefore does not create separate regional allowances that can bypass the project’s daily view.
What happens when a query reaches a quota
Before execution
BigQuery performs a quota check before starting a query that contains a covered generative-AI function. If the applicable quota is already exhausted, the query is rejected with a QuotaExceeded error instead of beginning execution.
Rank #2
During execution
BigQuery also monitors consumption while a query runs. If a quota is exhausted after execution has started, subsequent large-language-model calls are rejected. Rows that depend on those calls can fail, so a query may return only earlier results or may fail overall rather than stopping cleanly at a row boundary.
Controlling partial results
For functions that support the max_error_ratio option, that setting determines how much row-level failure the query can tolerate. Depending on the configured ratio and the number of failed calls, BigQuery can return partial results or report total query failure. Treat this as an execution-policy decision: a small ratio is appropriate when every generated value is required, while a more permissive ratio can preserve useful results from a large batch.
Rank #3
Which BigQuery functions are covered
The controls apply to BigQuery SQL functions that perform generative-AI inference with Gemini models. The announcement names AI.GENERATE, AI.IF, AI.CLASSIFY, and AI.SCORE; AI.GENERATE_TEXT is another example of the covered generative workload. Exact throughput and model limits can vary by function and model, so the token quota should be treated as a spending guardrail rather than a promise that every function can process the same volume.
Embedding and semantic-search exclusion
Token quotas do not apply to embedding-generation and semantic-search functions such as AI.EMBED. Google says those functions use different billing structures. Do not estimate an embedding pipeline’s cost or protection level from the four generative-token quotas; check the applicable pricing and quota documentation for that service instead.
Important billing details and edge cases
Cached tokens
Cached tokens do not count toward these quotas. A workload that reuses cached context can therefore consume less quota than an otherwise similar workload that sends fresh input, although its total bill still depends on the billing rules for the function and model.
Provisioned-throughput models
Customers using Gemini Enterprise Agent Platform models with provisioned throughput should set the BigQuery token quotas high. Provisioned-throughput billing is not token based, so a low token cap could block work even though the associated capacity is being paid for through a different billing model.
Best Value
How to use the controls in a practical cost policy
Set a project ceiling first
Use the project input and project output/thinking limits to bound the total daily exposure of production and development jobs combined. This catches a runaway scheduled query even when it is launched by an account that has not exceeded its individual allowance.
Add per-user guardrails
Per-user limits prevent one analyst, service account, or automation identity from consuming the entire project allowance. Decide how service accounts map to owners before setting the value; otherwise a shared identity can concentrate many teams’ workloads under one per-user bucket.
Match the policy to query behavior
Input limits protect against wide tables, long prompts, and accidental scans. Output/thinking limits protect against verbose generation and models that spend substantial reasoning tokens. Set both categories deliberately; controlling only input does not cap generated output.
Do these 3 things before closing this tab:
1Scan for outdated or missing drivers - takes under a minute2Repair Windows errors before they cause bigger problems3Fix the driver behind crashes, sound loss and screen glitchesTest failure handling
Run a controlled batch that approaches the chosen limits and verify the application’s response to QuotaExceeded and row-level failures. If downstream systems cannot accept partial data, require a strict max_error_ratio policy or add a retry and reconciliation process outside the query.
What this changes for BigQuery teams
The feature does not turn generative-AI queries into fixed-price operations. It gives administrators a self-service boundary around token consumption, with visibility that remains consistent across regions and enforcement both before and during execution. Teams still need to monitor query shape, model choice, prompt length, output volume, and the separate billing rules for embeddings or provisioned throughput. The practical result is a more predictable maximum exposure: a query can be refused before it starts, or its later model calls can be stopped, rather than consuming unlimited token-based usage until the dataset is finished.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
