2026-10-14 Changes confirmed medium confidence

Google Makes BigQuery AI Token Quotas Generally Available

Four daily limits let administrators bound project- and user-level input and output tokens for LLM-backed SQL functions, while embedding functions remain outside the guardrails.

Edited by Tyronne Panaino

Google made token-based cost management for BigQuery generative AI functions generally available on October 14, 2026. The update adds four daily quotas that administrators can adjust at project and per-user levels, giving teams a preventive control over the input, output and reasoning tokens consumed when SQL calls large language models.

The change matters to data teams running generation across large tables because a single query can classify, score or transform many rows. Token use can grow with both the size of the input and the model's response, so a control that is enforced during query execution is different from receiving a billing alert after usage has accumulated.

Four daily quotas separate input from output

BigQuery now exposes project-level and per-user limits for both input and output usage. Google's published defaults are 200 billion input tokens per project per day, 150 billion input tokens per user within a project, 20 billion output tokens per project per day and 15 billion output tokens per user within a project. These are defaults that administrators can lower; they are not recommended budgets for every workload.

Usage is summed globally across regions. Output limits also include thinking tokens, which means a reasoning-heavy request can consume the output allowance even when those internal tokens are not presented as the visible answer. A team comparing query volume with quota consumption therefore needs to account for more than the rows and text it can see.

Administrators can change the limits through Google Cloud's quota controls without opening a support case. Lowering a quota takes effect immediately without approval. Google also says the controls can be managed programmatically through the Cloud Quotas API, the `gcloud` command-line tool and Terraform, which gives larger organizations a route to apply consistent settings across projects.

The guardrails cover generation, not embeddings

The quotas apply to BigQuery functions that invoke an LLM for generation or judgment: `AI.GENERATE`, `AI.GENERATE_TABLE`, `AI.GENERATE_TEXT`, `AI.CLASSIFY`, `AI.IF`, `AI.SCORE` and `AI.AGG`. They do not apply to embedding functions such as `AI.GENERATE_EMBEDDING` and `AI.EMBED`, which use a different class of model.

That boundary is operationally important. An administrator cannot assume that one AI quota governs every model-backed BigQuery feature. A project that combines embeddings with generation needs to monitor the two paths separately and verify which limits apply before treating the new controls as a complete AI-spend ceiling.

Quota exhaustion can change query results

Enforcement depends on when the limit is reached. If the required quota is already exhausted when a query starts, BigQuery rejects the request with a `QuotaExceeded` error. If consumption reaches the limit while a query is running, BigQuery stops sending new calls to the model.

For functions that support `max_error_ratio`, processing can return partial results when the unprocessed share stays inside the configured tolerance; otherwise the query fails. General-purpose functions such as `AI.GENERATE` can still complete, but skipped rows have no generated value and carry the quota error in their status field. Applications must therefore inspect row-level status instead of treating a successful query job as proof that every requested model call ran.

Token quotas and dollar caps solve different problems

Google describes the new quotas as preventive and AI-specific. Cloud Billing spend caps operate in currency across projects and services, while token limits control a particular unit of model use. Because model prices differ and output or thinking tokens can cost differently from input tokens, a token allowance is not a fixed monetary budget.

BigQuery's `AI.COUNT_TOKENS` function can estimate the text tokens in prospective input without making a billable generation call. Google cautions that this is only a lower bound: it excludes output and thinking tokens, instructions added by the AI function and non-text inputs such as images. A small representative run and observed usage are still needed before converting a desired budget into quota values.

The practical delta is clearer control, not automatic cost optimization. Teams gain a hard execution-time guardrail and separate user scope, but they still need model-aware pricing, workload measurements, error handling and coverage for excluded functions.

Status

Confirmed. Internal confidence is medium because one official Google Cloud article supports the general availability, scope, defaults and enforcement behavior, while this run did not independently test quota propagation, partial-result handling or billing outcomes.