Built-in LLM web search vs a SERP API: cost per answer
Claude and OpenAI charge $10 per 1,000 searches plus the result tokens. We measured the token side on 20 questions and priced 1,000 answers both ways.
As of October 2026, Anthropic and OpenAI both list their built-in web search tool at $10 per 1,000 searches and bill the retrieved content as input tokens on top. A SERP API that hands the model only titles, links and snippets injects about 360 tokens per search instead of thousands. On our worked example, 1,000 two-search answers cost $6.46 that way against an estimated $44.00 with the hosted tool (a $2-per-million-token model). The gap is mostly tokens, not the per-search fee.
What do the built-in web search tools charge?
Official list prices, read on 19 October 2026:
| Tool | Search fee | Result content |
|---|---|---|
| Claude web search | $10 per 1,000 searches | Input tokens, in the same turn and in later turns |
| OpenAI web search | $10.00 per 1,000 calls | Billed at the model’s token rates |
| OpenAI preview, non-reasoning models | $25.00 per 1,000 calls | Free |
| Gemini grounding, Gemini 3 and newer | $14 per 1,000 queries after 5,000 free a month | Per search query performed |
Anthropic prices the tool at “$10 per 1,000 searches, plus standard token costs for search-generated content” and does not bill a search that errors. OpenAI adds that on gpt-4o-mini and gpt-4.1-mini the content is billed as a fixed 8,000 input tokens per call. Neither page says how many tokens a typical search injects, and OpenAI’s guide says search_context_size “does not set an exact token count”.
How many tokens does one search put in the context?
On 19 October 2026 we ran 20 research questions through FastAF /serp: 20 of 20 returned 200 at 5 credits, with a median of 10 organic results. We built the text you would give a model and counted tokens with tiktoken (o200k_base):
import os, requests
r = requests.get("https://api.fastaf.co/v1/serp",
params={"query": "best vector database for RAG", "country": "us"},
headers={"Authorization": f"Bearer {os.environ['FASTAF_API_KEY']}"})
top5 = r.json()["organic_results"][:5]
text = "\n\n".join(f"{i}. {x['title']}\n{x['link']}\n{x.get('snippet', '')}"
for i, x in enumerate(top5, 1))| What goes to the model | Median tokens | Range |
|---|---|---|
Raw /serp JSON |
2,285 | 1,815 to 4,673 |
| Top 5 results, title + link + snippet | 362 | 277 to 462 |
| Top 10 results, title + link + snippet | 698 | 491 to 817 |
| One page behind those links, as Markdown | 4,912 | 10 to 138,718 |
The last row comes from fetching the top 5 links of each question with /scrape?formats=markdown: 96 of 100 fetches worked at 1 credit each (the 4 failures were not charged), and 16 of those pages had under 500 tokens. The five pages of one question together had a median of 25,399 tokens.
I did not call the hosted tools, so I cannot say what they inject for these questions. Documented figures are in the page-sized range: Anthropic’s example response for a one-search question shows 6,039 input tokens, and OpenAI’s mini models bill a flat 8,000. The tokenizer is OpenAI’s, so Claude counts will differ.
What does one answer cost, not one search?
An answer is a bundle: S searches, T tokens injected per search, R times those tokens are read as input. A one-shot question has R = 1. An agent loop or a chat with follow-ups re-reads them each turn, and Anthropic says results count as input tokens “in subsequent conversation turns”. Cost per answer is S x fee + S x T x R x input price; output tokens are the same on both sides, so they are left out.
Worked example: 1,000 answers, 2 searches each, Claude Sonnet 5.5 at $2 per million input tokens (Anthropic pricing). The hosted row assumes 6,000 tokens per search, in line with Anthropic’s 6,039 example.
| Setup | Search fees | Tokens, read once | Total | Read on 3 turns |
|---|---|---|---|---|
| Hosted tool, ~6,000 tokens per search (estimate) | $20.00 | $24.00 | $44.00 | $92.00 |
| FastAF snippets only, Starter pack ($2.50 per 1,000 searches) | $5.00 | $1.46 | $6.46 | $9.39 |
| FastAF snippets + 2 page fetches per answer, Starter pack | $6.00 | $21.11 | $27.11 | $69.34 |
On the Business pack ($1.33 per 1,000 searches) the snippets-only total is about $4.13. Row three matters: if the agent reads two full pages per answer, the token bill is about the same as the hosted tool’s and you save only the fee difference. The big gap appears when snippets are enough.
Two other levers. Prompt caching: the pricing page lists Sonnet 5.5 cache hits at $0.10 per million tokens, a twentieth of fresh input. Model choice: the same 12,000 injected tokens cost $0.0012 on Haiku 5.5 and $0.048 on Opus 5.5, so on a cheap model the $0.02 fee dominates.
When is the built-in tool the better choice?
Often:
- Setup time is the cost. One entry in the
toolsarray, no wrapper, no second vendor. At a few hundred answers a month, the fee is a rounding error next to an hour of engineering. - Citations are the product. Both tools return structured citations tied to the answer text. Rebuilding that on snippets is work.
- You want the model to rank and read. The provider’s retrieval picks passages for the model, which can beat five snippets when the answer sits inside a page.
A SERP API wins at volume, when you want a fixed 5 credits per search under a prepaid balance (see the spending cap pattern), when you need to choose country or language, or when you want the same search under any model. Vendor prices: SERP API cost per 1,000 searches. Pipeline: SERP APIs for RAG and grounding.
How do I get the cheaper setup?
Expose a search tool that returns the compact text above and a second tool that fetches a page only when the model asks: 5 credits per search, 1 per static page, nothing for failed requests. Over MCP the same tools reach Claude Code, Cursor or Codex, as in the MCP setup post. Two production calls on 19 October 2026 returned 200 in 4.5 s and 3.8 s, each charging 5 credits. See the Google SERP API page and the docs.
The short version
Claude and OpenAI charge $10 per 1,000 searches (Gemini $14 after a free allowance) plus the injected tokens, which Anthropic re-bills on later turns. We measured 362 tokens per search for title, link and snippet text against a median 4,912 for a page; at $2 per million tokens that made 1,000 two-search answers $6.46 against an estimated $44.00, and the gap narrows if your agent fetches full pages. Use the built-in tool for zero setup and ready citations, a SERP API for volume and predictable cost. The free 500 credits need no card: get a key.