BigQuery adds daily token quotas to rein in generative AI costs

Four daily quotas, managed through IAM

The generative AI functions in BigQuery, including AI.GENERATE, AI.IF, AI.CLASSIFY and AI.SCORE, put large language models directly into SQL. According to Google Cloud, a single query can classify millions of support tickets or score each product review for sentiment. Because these functions are billed by the token, running them across large datasets can drive up costs quickly.

The new controls consist of four daily quotas managed through the Google Cloud Identity and Access Management (IAM) quota system. The defaults, as listed by Google Cloud, are:

• AI input tokens per day (per project): 200 billion
• AI input tokens per day per user (per project, per user): 150 billion
• AI output tokens per day (per project): 20 billion
• AI output tokens per day per user (per project, per user): 15 billion

Per-user limits apply to each user within a project, not across an organization. The quotas are global, so usage is summed across all regions. Output quotas also count "thinking" tokens, in the same way reasoning tokens count toward the Gemini platform bill.

The quotas cover AI.GENERATE, AI.GENERATE_TABLE, AI.GENERATE_TEXT (and its alias ML.GENERATE_TEXT), AI.CLASSIFY, AI.IF, AI.SCORE and AI.AGG. Embedding functions such as AI.GENERATE_EMBEDDING and AI.EMBED are not covered, because they run on different, substantially cheaper models, according to the company.

Adjusting limits and handling enforcement

Administrators can view and change the quotas in the Google Cloud console under IAM & Admin, Quotas & System Limits, without opening a support case. Lowering a quota takes effect immediately and needs no approval. Raising one is done by requesting an increase on the same page. Google Cloud says many requests are auto-approved, while larger increases may go through a quick review.

Changing a quota requires one of several roles: Owner, Editor, Quota Administrator or Service Usage Admin. The quotas can also be managed through the Cloud Quotas API, the gcloud CLI and Terraform, which allows consistent guardrails across projects. The Quotas & System Limits page also shows historical token consumption, with selectable 7-day, 14-day, 30-day and custom ranges, to help spot spikes.

Enforcement works by stopping work. A query is rejected up front with a QuotaExceeded error only if a quota it needs is already exhausted. Otherwise it starts, and if a quota runs out mid-query, BigQuery stops sending new LLM calls. Functions that support max_error_ratio, such as AI.IF, return partial results if unprocessed rows stay within the ratio, and otherwise fail. General-purpose functions such as AI.GENERATE still succeed, but skipped rows have no generated value and their status field contains the quota-exceeded error.

Google Cloud also contrasts the quotas with Cloud Billing spend caps. Token quotas are preventive and AI-specific, while spend caps work in dollars at the project and service level. Because spend cap enforcement is not instant, some overages are incurred, according to the company. It recommends token quotas as the first line of defense and a spend cap as a project-wide safety net.

Sizing quotas with AI.COUNT_TOKENS and a worked example

To choose quota values, Google Cloud points to the AI.COUNT_TOKENS function. It returns the number of tokens a model's tokenizer produces for a text string without making a billable generation call. The count is a lower bound on input tokens: it leaves out output and thinking tokens, the instructions each AI function adds around the input, and non-text inputs such as images.

The post includes a budget example for keeping a project's daily AI spend around $100 with the Gemini 3.1 Flash-Lite model. Using published pricing at the time of writing, which reflects a 50% discount for BigQuery's batch processing, input tokens cost $0.125 per 1 million and output tokens (response and reasoning) cost $0.75 per 1 million. With a sample-run ratio of 100:1 input to output tokens, Google Cloud arrives at about 754 million input tokens and 7.54 million output tokens per day.

The company's best practices are to start with per-user quotas so one user or runaway experiment cannot drain a project's budget, to measure with AI.COUNT_TOKENS and a small sample run before scaling, to set max_output_tokens where a function accepts it because output tokens cost more, and to adjust limits as usage and budgets change.

What it means for BigQuery teams

For organizations running LLM-backed SQL at scale, the feature gives administrators a self-serve way to cap token use before costs build up, rather than relying only on dollar-based limits. Google Cloud says token-based cost management is available now, and the documentation on controlling generative AI costs covers the details.