Skip to main content

Comparison

OpenAI vs Anthropic Web Search Pricing: Same $10 Sticker, Two Different Bills

By AInews Desk
OpenAI vs Anthropic Web Search Pricing: Same $10 Sticker, Two Different Bills
7 min read
1227 words

On a pricing page, the OpenAI and Anthropic web search tools look almost identical. Anthropic lists web search on the Claude API at $10 per 1,000 searches. OpenAI's pricing page lists its web search tool at $10.00 per 1k calls. The search fee is where the similarity ends. On both APIs, the retrieved content is billed again as tokens, and the two vendors count those tokens differently, so the same search fee can lead to two very different invoices.

What does one search cost on each API?

Anthropic has one rate. It charges $10 per 1,000 searches, plus standard token costs for search-generated content. Two billing rules matter in practice. Each search counts as one use regardless of how many results come back, and a search that fails with an error is not billed.

OpenAI has several rates, depending on which tool and model family you call:

  • Web search (incl. images): $10.00 / 1k calls, with search content tokens billed at model rates.
  • Web search preview (reasoning models): $10.00 / 1k calls, with content tokens billed at model rates.
  • Web search preview (non-reasoning models): $25.00 / 1k calls, with search content tokens free.

That last line is easy to miss. The per-call fee is two and a half times higher, but you pay nothing for the retrieved content. If your queries pull in a lot of text, the higher fee can cost less overall. The catch is that the Responses API still accepts web_search_preview for legacy integrations, but OpenAI says to use web_search for new integrations, and web_search_preview does not support filters. OpenAI also notes that search actions incur a tool call cost. Model responses can contain other action types, such as open_page and find_in_page, which are supported in reasoning models.

Where the token bills split

Content tokens are the bigger cost, and the two vendors handle them very differently.

On Anthropic's side, web search results retrieved throughout a conversation count as input tokens, both in search iterations during a single turn and in subsequent conversation turns. Results stay in the conversation. A search done early in a long chat is billed again as input on every later turn. In a multi-turn assistant, this repeat charge can cost more than the one-cent search fee.

Anthropic offers a way to reduce it. With basic web search, every result is loaded into Claude's context window. With web_search_20260209 or later, Claude writes and runs code that filters results first, so only relevant content reaches the context window, which reduces token use on search-heavy requests. If you are still on web_search_20250305, check this before you scale.

OpenAI is more predictable for some models. For gpt-4o-mini and gpt-4.1-mini with the non-preview web search tool, search content tokens are billed as a fixed block of 8,000 input tokens per call. That makes budgeting simple, though you pay for the full block even when a search returns very little. Retrieved context is also capped: the web search context window is limited to 128k, even when the underlying model has a larger context window.

A hard cap versus a volume dial

The two vendors give you different settings for limiting spend. Anthropic gives you a counter. You can steer how often Claude searches through the system prompt, but to enforce a limit, use max_uses to cap the number of searches in a turn. To meter usage, the response's usage block reports server_tool_use.web_search_requests. Claude Managed Agents work differently. There, web search is on by default in the agent toolset, without max_uses, so you can't set a per-turn cap in that setting.

The max_uses cap is stricter than it first appears: failed searches count against the limit, so an erroring search still uses up part of that per-turn budget. Go over it and the search comes back as an error with the max_uses_exceeded error code.

OpenAI controls the amount of content instead of the number of searches. Its search_context_size setting controls how much context from web search results is made available to the model: low for a lookup that needs only a few details, medium as a balanced default, high when the answer may need more. It does not set an exact token count or guarantee a specific number of sources or citations. To keep searches off the live web, set external_web_access: false on the web_search tool to run in offline/cache-only mode.

Domains, location and citations

DimensionAnthropic (Claude API)OpenAI
Search fee$10 / 1,000 searches$10 / 1k calls; $25 for non-reasoning preview
Content tokensInput tokens, carried into later turnsModel rates; free on non-reasoning preview
Failed searchesNot billed—
Volume controlmax_uses per turnsearch_context_size low/medium/high
Domain filterAllowed or blocked list, never bothUp to 100 allowed or 100 blocked
Domain formatBare domain, optional path, no schemeNo http/https prefix; covers subdomains
Filter scopeSet on the web_search toolResponses API with web_search only
Locationuser_location on the toolCountry, city, region, timezone
CitationsIncluded for search-sourced contentURL annotations in message output

Anthropic is strict about domain lists. Send both allowed_domains and blocked_domains and the API returns a 400 error. Entries are bare domains with an optional path, such as example.com/blog. OpenAI's filters parameter accepts up to 100 allowed_domains or up to 100 blocked_domains; you omit the HTTP or HTTPS prefix, and subdomains are included. One limitation is easy to miss: domain filtering is only available in the Responses API with the web_search tool. Chat Completions users can't use it.

For location, Anthropic's user_location goes on the tool alongside max_uses and the domain lists. OpenAI accepts an approximate location using country, city, region, and/or timezone.

Both vendors cite their sources. Anthropic's response includes citations for sources drawn from search results. OpenAI returns annotations for the cited URLs in the message output, and it sets a rule for your app: when your interface displays web results or information from them, you must show inline citations as prominent, clickable links.

Choosing by workload

The bill depends on how your app uses search.

  • Long multi-turn chats with Claude: Retrieved results are billed again on later turns. Dynamic filtering and a tight max_uses will cut more cost than the one-cent search fee ever will.
  • Quick, content-heavy lookups on OpenAI: consider the $25.00 / 1k calls non-reasoning preview, whose search content tokens are free, as long as you accept the legacy web_search_preview tool and its lack of domain filters.
  • Strict per-request budgets: Anthropic's per-turn search cap and per-search metering are easy to reason about. OpenAI's fixed 8,000-token block on the mini models does the same for token costs.
  • Allowlisted research tools: If you need domain filtering on OpenAI, you need the Responses API. On Anthropic, choose either an allowlist or a blocklist per request.

Before you budget for either API, run a sample of real queries. Look at the search count in the response metadata and the input tokens on later turns, not just the price per thousand.

Sources

7 min read
•
1227 words

Tags

OpenAI vs Anthropic web search pricing comparisonOpenAI search API cost per 1000 callsAnthropic Claude web search tool pricingAI web search tool pricing models

Related Articles