Tool-enabled AI often has two bills: using the language model and invoking a metered capability. Web search illustrates why a token-only forecast can underestimate retrieval-heavy agents.
Pricing snapshot: Vendor first-party documentation checked on 25 October 2026. Original price effective date unknown unless separately declared. Amounts are published list figures, subject to tier, region, currency and eligibility.
The search call is not the model completion
OpenAI’s price table lists Web search at $10 per thousand calls for the general listing, in addition to model-rate search content tokens. A separate non-reasoning Web search preview line quotes $25 per thousand calls. These labels identify distinct tool/product combinations; do not choose the cheapest row without confirming which tool variant and target model are supported. The model’s ordinary input and output charges remain separate.
The price page also states that Responses, Chat Completions, Realtime, Batch and Assistants endpoints are not themselves priced separately: it is the selected model and used tools that determine metered charges. That does not make tool calls free.
A workable search ledger
For every application request, retain model ID, tool type, observed number of chargeable calls, search-returned tokens, other input/output tokens, model tier and region. Count repeated tool invocations across retries and multi-step answers. A workflow asking one visible user question can trigger several searches; request count is not a substitute for tool-call count.
A simple example: 2,500 qualifying general Web search calls at $10 per thousand yield a $25 tool-call component. Add search content tokens at the applicable model rate and model generation cost. Because response length and amount of indexed content vary, do not invent an all-in per-question price from the call fee alone.
Why agent architecture affects cost
A model that repeatedly searches for similar evidence can incur repeated call charges even when its final answer is brief. Add routing checks, stop conditions and deduplicated queries before launching web-dependent agents at scale. Conversely, an aggressive cache of old search results can cause stale answers. The design choice is a balance between freshness, accuracy and cost, not merely fewer API calls.
Internal document search, hosted shell execution and browser actions have different usage semantics. Do not assume the Web search unit price applies to those tools or to an external search API with a separate subscription.
Budget a large editorial job
Suppose an editorial pipeline handles 10,000 source-verification requests. The forecast needs expected tool calls per request, failed and repeated calls, number of source tokens returned, model input/output, and human escalation. A one-search-per-request assumption is plausible only if the actual orchestration enforces it. Use test logs from your own deployment when available and label hypothetical scenarios clearly.
Monitor total cost per independently verified article, not just cost per query. If additional retrieval prevents a costly factual correction, a higher tool spend may produce lower total editorial cost, subject to verification.
Vendor distinctions and audit trail
The source is the official OpenAI API pricing page accessed for the 25 October 2026 cutoff. It is a mutable commercial page with no verified price-effective date in the captured material. Store the tool version and actual invoice separately; a customer or cloud marketplace may have contract-specific charges. All amounts here are USD before region-specific tax or conversion.
Alternatives include running search less frequently, searching a known document corpus or using a different provider whose call unit may differ. None of those alternatives is a free substitute without checking their own usage and licensing terms.
Pricing evidence boundary
This cost discussion concerns OpenAI (web-search; web-search-preview). Reproduce a quote only with its exact model or plan identifier, chargeable unit, billing region, processing mode and timestamp; a changing vendor rate card is not a permanent price guarantee. If any one of these dimensions is missing, keep the associated rate and worked scenario under editorial review instead of filling the gap with an estimate.
Sources & Further Reading
- developers.openai.com — pricing (official documentation; accessed 2026-10-25)