Google adds token quotas for BigQuery AI functions

Google has made token-based cost controls generally available for BigQuery generative AI functions, adding daily token quotas at project and user level.

The controls apply to BigQuery functions that use large language models through SQL, including AI.GENERATE, AI.GENERATE_TABLE, AI.GENERATE_TEXT, AI.CLASSIFY, AI.IF, AI.SCORE and AI.AGG.

The change addresses a growing cost issue as companies run generative AI tasks across large datasets in BigQuery. Because these functions are billed by the token, spending can rise quickly when teams apply classification, scoring or generation tasks to millions of rows.

Administrators can now set four daily quotas through the Cloud Identity and Access Management quota system: input tokens per day at project level, input tokens per day per user, output tokens per day at project level and output tokens per day per user.

Default limits are 200 billion input tokens per day per project and 150 billion input tokens per day per user. For output and thinking tokens, the defaults are 20 billion per day per project and 15 billion per day per user.

The quotas are global, so usage is counted across regions rather than tracked separately by geography. Output quotas also include thinking tokens, the reasoning tokens used by the model.

What is covered

The quota system applies only to LLM-backed generation functions in BigQuery. Embedding functions such as AI.GENERATE_EMBEDDING and AI.EMBED are excluded because they run on different, cheaper models.

Administrators can change quotas directly in the Google Cloud console without filing a support request. Lower limits take effect immediately, while requests for higher limits can be submitted through the same interface and may be automatically approved or sent for review, depending on the size of the increase.

Quotas can also be managed through the Cloud Quotas API, the gcloud command-line tool and Terraform, giving organisations a way to apply the same controls across several projects.

Monitoring use

The quotas page also shows historical token consumption, allowing administrators to spot spikes and compare trends over seven-day, 14-day, 30-day or custom ranges. That data can help teams judge how new AI workloads affect token consumption before budgets are exceeded.

Google drew a distinction between token quotas and billing spend caps in Google Cloud. Token quotas are a preventive control tied specifically to AI usage, while billing spend caps operate in dollar terms across projects and services and may still allow some overage because enforcement is not immediate.

Under the enforcement model, BigQuery rejects a query at the outset only if the required quota has already been exhausted. If a query starts and the limit is reached during execution, BigQuery stops sending new LLM calls.

For functions that support max_error_ratio, including AI.IF, a job may return partial results if the number of unprocessed rows remains within the allowed ratio. General-purpose functions such as AI.GENERATE can still complete, but skipped rows will not contain a generated value and their status field will show a quota-exceeded error.

Budget planning

To help users estimate likely consumption before running large jobs, Google pointed to the AI.COUNT_TOKENS function. It returns the number of tokens a model tokenizer would produce for a text string without making a billable generation call.

That allows teams to assess the size of input data before scaling up a workload. The figure is only a lower bound because it excludes output tokens, thinking tokens, instructions added by AI functions and non-text inputs such as images.

Google also outlined an example of how users could work backwards from a daily budget. Using Gemini 3.1 Flash-Lite in batch mode, published pricing equates to USD $0.125 per 1 million input tokens and USD $0.75 per 1 million output tokens after a 50% batch discount.

In the example, a project aiming to keep daily AI spending near USD $100 with an average 100:1 input-to-output token ratio would set limits at roughly 754 million input tokens and 7.54 million output tokens per day. Projects using several models should either budget by model or blend the relevant rates.

A practical approach is to begin with per-user quotas so one experiment or one user does not consume the entire project budget. Google also advised users to measure token counts before scaling, pay close attention to output tokens because they cost more, and set maximum output lengths where functions allow it.

The release adds a new layer of cost control as Google pushes generative AI deeper into BigQuery, where analysts and data teams can invoke models from within SQL queries across large structured datasets.