Hosted web search: token-only billing
Native Responses requests declaring web_search, web_search_2025_08_26, web_search_preview, or web_search_preview_2025_03_11 use token-only soft budgets. No supplier certification profile or search price is required. Existing authentication, model access restrictions and available-balance admission remain active.
The gateway preserves the search tools and caller-provided limits. It does not add max_tool_calls or max_output_tokens, remove search, or create a hard-budget reservation for these requests. Complete measured input/output tokens, including reported reasoning and cache tokens, use the deployment's normal token pricing. Search/tool charges are excluded even when upstream reports them or the deployment has a search price. Raw upstream usage and search results remain in responses and logs.
A request can spend more than the balance remaining when it started; concurrent search requests do not have a hard spending guarantee. Completed token charges count toward subsequent admission. Missing or contradictory token usage remains unpriced rather than being estimated. Search output exceeding a caller limit is charged using actual reported tokens and does not create a new uncertain reservation.
Observable response headers:
X-XHub-Budget-State: token_onlyX-XHub-Billing-Scope: tokens_onlyX-XHub-Budget-Guarantee: softX-XHub-Tool-Costs: excluded
The usage price snapshot stores the same scope and guarantee under budget. Streaming clients receive the policy headers before generation. Tool usage is retained as upstream evidence, not charged. Supported public search history can be continued through native Responses; cross-protocol conversion of search history remains unsupported.
Soft budgets now apply to all model requests. The budget layer no longer rejects hosted tools using an allowlist; protocol conversion and upstream capability restrictions still apply. Existing uncertain reservations are not automatically released or modified.
Regression uses local HTTP suppliers and a private PostgreSQL schema; browser tests create and clean up a model and limited key in both supported UI languages. These prove gateway policy, not upstream search availability, quality or compliance with client-provided limits.