Research-backed calculator guide

AI Research-Agent Cost per Verified Deliverable: Search, Context, Citations and Human Review

Calculate the fully loaded cost of an AI research agent only after a source-grounded deliverable passes citation checks and human review.

Why “per attempt” is the wrong unit

A research run can finish and still fail the job. It may cite a search-result snippet instead of reading the underlying page, rely on stale context, cite a claim to the wrong source, omit a material limitation, or produce a plausible narrative that a reviewer rejects. Those failure modes consume the same initial tokens as a successful run and may add another search, fetch, model pass, or review cycle.

The useful denominator is therefore an accepted deliverable, not a request, completion, or draft. Define “accepted” before you calculate. For example, a deliverable might need a specified number of distinct sources, citations attached to factual claims, a date and locale, a correct answer to the brief, and a human sign-off. A short answer that is cheap but fails one of those gates is not a low-cost deliverable; it is an incomplete attempt.

The formula makes the distinction explicit. If an attempt costs $2 and only half of attempts pass, the expected cost per accepted output is $4 before any extra repair work. If reviewer time is charged on every attempt, include it in the numerator rather than treating it as a separate success fee.

What belongs in the numerator

Search and retrieval

Count each billable search request or query under the provider’s meter. Anthropic’s Web Search documentation says API web search is charged at $10 per 1,000 searches, in addition to standard token usage for search-generated content; each search counts as one use regardless of the number of results returned. ^[1] That means three requests are $0.03 under that published rate, before token charges.

Google’s Gemini documentation uses a different meter for Gemini 3 and newer models: the project is billed for each search query the model decides to execute, with 5,000 free search requests per month and then $14 per 1,000 search requests. Older Gemini models use per-prompt billing. ^[4] Do not copy an Anthropic per-search assumption into a Gemini estimate. Record the provider, model generation, free allowance, and billing unit beside the calculation.

Fetched context is a second cost driver even when the retrieval feature has no separate line item. Anthropic describes Web Fetch as retrieving the full text of a specified URL and inserting it into the conversation; the documentation also notes that JavaScript-rendered sites need a real browser rather than this fetch tool. ^[2] Treat fetched text as input context in your token ledger, and add any separate browser, data, proxy, or third-party API fee your implementation actually pays.

Model and orchestration calls

Count planner, search-query generation, source extraction, evidence mapping, synthesis, citation repair, and final formatting calls. A retry is not free because the first answer was rejected. Track input and output tokens separately because context accumulation can make late calls much larger than early calls.

A simple estimate is:

Model cost = (input tokens ÷ 1,000,000 × input price) + (output tokens ÷ 1,000,000 × output price).

Use the current provider rate card for a production budget. The worked example below uses $2 per million input tokens and $10 per million output tokens only as an explicit illustration, not as a current universal price or a promise that every provider bills those units.

Citation and review work

Citations are more than links appended after drafting. Anthropic’s citation documentation explains that supported documents are processed into chunks and that citations identify locations in the source material, such as character ranges for plain text or page ranges for PDFs. ^[3] A verification pipeline therefore needs to check that the cited source contains the claim, that the citation points to the right passage, and that the source is appropriate for the claim.

Price automated checks as infrastructure or model work when they consume either. Then price the human gate. A reviewer may confirm source quality, inspect calculations, verify dates and units, reject unsupported claims, and request a repair. Use fully loaded reviewer cost, not just salary. If a reviewer’s loaded rate is $60 per hour and the review takes 10 minutes, the labor allocation is $10.00. State whether the reviewer checks every attempt or only escalations.

Acceptance and repair

Let A be the first-pass acceptance rate. If the attempt cost is C, the first approximation is C ÷ A. A lower acceptance rate raises effective cost even if token prices stay constant. Add a repair reserve when rejected runs create predictable extra work. A reserve can be estimated from historical repair minutes and retry tokens; until you have that history, label it as an assumption and run sensitivity cases rather than presenting it as a market benchmark.

Worked cost example

Suppose a research team wants one two-page market brief with source links and a reviewer sign-off. The workflow makes 12 model calls. Across the run, it uses 36,000 input tokens and 6,000 output tokens. It makes three Anthropic Web Search requests, folds fetched source text into the input total, allocates $0.04 of orchestration and storage overhead, spends 10 minutes in human review at a fully loaded $60/hour, and reserves $0.03 for citation repair. The model rate below is an illustrative assumption.

The largest line is reviewer time, not search. That is not a universal result; it follows from the chosen 10-minute review and $60/hour assumption. If review falls to five minutes, the numerator falls by $5. If the task is high risk and review takes 20 minutes, the accepted-output cost rises substantially even with identical model usage.

The 85% acceptance rate also matters. At a 70% rate, the same $10.232 attempt becomes about $14.62 per accepted deliverable. At 95%, it becomes about $10.77. These are estimates calculated from the stated assumptions, not observed vendor performance. Keep the acceptance rate segmented by task type because a literature brief, a regulated memo, and a lightweight fact sheet should not share one blended quality rate.

For monthly planning, multiply the accepted-output cost by accepted deliverables, then add fixed platform or staffing costs that are not already allocated per run. For 500 accepted briefs at the 85% scenario, the variable budget is about $6,019. A separate fixed platform subscription, data contract, or reviewer management cost must be added once, not silently folded into the per-run number twice.

How to improve the unit economics without weakening verification

Start with the acceptance rule, then reduce avoidable work. Limit search breadth when additional sources do not change the conclusion, but do not skip required source checks. Cap fetched text to the passages needed for the question and preserve source identifiers so a reviewer can reproduce the evidence. Anthropic’s fetch documentation warns that dynamically rendered pages need browser handling, so budget the fallback path rather than assuming every URL has the same retrieval cost. ^[2]

Separate cheap mechanical checks from expensive judgment. A script can detect missing URLs, duplicate citations, empty evidence fields, and mismatched dates. A model can help locate unsupported sentences, but a human may still need to verify whether the source actually supports the claim. Citation-enabled workflows can expose source locations, but the presence of a citation is not proof that the claim is correct. ^[3]

Control the loop with caps. Log maximum searches, maximum fetched pages, maximum model calls, retry count, and reviewer minutes. The AI agent budget guardrail and P95 calculator is useful when occasional long runs create a tail risk that an average cost hides. Use a P95 or percentile budget when a few difficult briefs consume many more sources or repair passes than normal.

Reconcile estimates with invoices. Compare the run ledger with provider usage exports, search requests, storage, and review tickets. The AI API invoice reconciliation calculator helps separate an estimate from actual billed usage. For a new system, use a pilot sample to replace assumptions with measured distributions; do not report a made-up “typical” cost before you have enough comparable accepted outputs.

Build cost and run cost should remain separate. Engineering time for connectors, evaluation, prompt maintenance, and deployment belongs in an implementation or amortization view. The AI agent development cost estimator covers that adjacent question. This guide’s unit is the recurring cost of producing one accepted research deliverable after launch.

FAQs

What is the cheapest way to calculate AI research-agent cost?

Start with a per-attempt ledger containing model tokens, search requests, fetched-context tokens, tool fees, workflow overhead, repair work, and reviewer time. Then divide by the number of accepted deliverables. A token-only estimate may be useful for model selection, but it is not a verified-deliverable cost when review or rework is material.

Should human review be included in cost per verified deliverable?

Yes, when human review is part of the acceptance definition. Allocate reviewer minutes using a fully loaded hourly rate and state whether every attempt is reviewed or only flagged attempts. If review is optional for some deliverables, publish separate unit costs for reviewed and unreviewed paths rather than blending them without an explanation.

Do search and fetch have the same billing unit?

No. Billing depends on the provider and product. Anthropic documents Web Search at a per-search-request rate plus token usage, while Google documents Gemini 3 and newer Search grounding as billed per query after a monthly free allowance. Web Fetch can add context that increases token usage, and a browser or third-party retrieval fallback may add its own fee. Check the current provider documentation before finalizing a budget. ^{[1] [2] [4]}

Is a cited answer automatically a verified deliverable?

No. A citation can point to a real source while still being incomplete, poorly matched to the claim, stale for the required date, or attached to a conclusion the source does not support. Verification should test claim-to-passage support, source quality, date and unit handling, required coverage, and the acceptance rubric. Citation tooling helps expose evidence locations; it does not replace the rubric or reviewer. ^[3]

How many accepted outputs do I need before using a measured acceptance rate?

Use a pilot large enough to represent the task mix, then report the sample size, task definition, and acceptance rule. Do not transfer a rate from one research domain to another without checking whether source volume, ambiguity, and reviewer strictness changed. Until the sample is credible, show a sensitivity range such as 70%, 85%, and 95% instead of one falsely precise rate.

What costs are excluded from the worked example?

The example excludes fixed subscriptions, data licensing, engineering build cost, taxes, currency conversion, provider minimums, browser or proxy fees, and the opportunity cost of a reviewer waiting on a queue. It also assumes fetched context is already included in the input-token total and that the listed search price applies. Replace those assumptions with your contract and usage records before using the result for pricing or capacity decisions.