AI agent cost per task
The monthly price on a pricing page doesn't tell you what a finished piece of work costs. We ran three standard tasks on Lindy, Manus and Claude Free and measured what each one used. Below are the results, followed by how per-task cost works under each pricing model.
Last tested Oct 17, 2026
Measured results
Method
3 tasks, 3 runs each per agent, a new chat for each run. Every run was screen-recorded and graded against hidden answer keys. We report medians. Cost = credits used × list price per credit. Tested Oct 16-17, 2026.
The tasks
- T1: Contract summary. Summarise a fixed 20-page synthetic contract in 400 words or fewer, then list parties, term, fees, termination rights, liability cap, auto-renewal and 3 risky clauses with page numbers.
- T2: CSV reconciliation. Reconcile bank.csv (500 rows) against ledger.csv (480 rows), which contain mixed date formats, duplicates and 12 planted mismatches, and output a reconciled CSV with reasons.
- T3: 5-email cold sequence. Write a 5-email prospecting sequence over 14 days, 120 words or fewer per email, one CTA each, with personalisation tokens, from a supplied persona and product brief.
Median cost per task and pass rate
| Agent | Price basis | T1: Contract summary | T2: CSV reconciliation | T3: 5-email cold sequence | All 3 tasks |
|---|---|---|---|---|---|
| Lindy | ~$0.01/credit list ($49.99 for 5,000 credits) | ~$0.02 (2 cr) Pass 3/3 | ~$0.59 (59 cr) Pass 3/3 | ~$0.04 (4 cr) Pass 1/3* | ~$0.65 65 credits |
| Manus | $0.005/credit list ($20/mo for 4,000 credits)† | ~$0.23 (46 cr) Pass 3/3 | ~$0.48 (95 cr) Pass 3/3 | ~$0.15 (29 cr) Pass 3/3* | ~$0.86 sum of the 3 task medians |
| Claude Free | $0 (Free plan, capacity-limited) | $0 Pass 3/3 | $0 Pass 3/3 | $0 Pass 1/3* | $0 capacity-limited |
Costs are medians of 3 runs at list price. "Pass" is runs passed out of 3. Lindy runs used the free trial, so the list rate is shown. Claude was tested on the Free plan; it has no usage meter, so no credits are shown.
* Email-test (T3) passes are provisional: they were scored by one rater, and the pass rule calls for two.
† We paid $10 for Manus's first month (promo), about $0.0025/credit. At that rate the medians would be about half the list figures shown.
Watch the test
The full test on video, 4:55. Every run was screen-recorded.
Chapters
Median time per task
| Agent | T1: Contract summary | T2: CSV reconciliation | T3: 5-email cold sequence |
|---|---|---|---|
| Lindy | ~23 s | ~5 min | 33 s |
| Manus | 47 s | 2 min 22 s | 36 s |
| Claude Free | ~17–23 s | 2 min 36 s | ~29 s |
CSV review burden
Review burden is the number of source rows a person has to check in the T2 output; 16 is the minimum for this test. Median across 3 runs: Lindy 16, Claude Free 16, Manus 22. Manus caught every planted issue but in two of three runs flagged both copies of each duplicate plus its partner row for review.
Notes
- Lindy and chat attachments: Lindy ignored .csv and .md files attached in chat (the files showed as attached, but the agent reported none). Uploading them via Memory > Files worked. PDF attachments worked. The failed attempts used extra credits that are not in the medians above.
- Claude Free capacity: after about 7-8 of these tasks in roughly 40 minutes, Claude Free locked us out for several hours. Anthropic publishes no message number for Free, so this is one observation.
- Email-test passes are provisional: T3 passes were scored by one rater. T3 fails come from hard checks (more than one CTA in an email, or vendor figures attributed to customers).
- Lindy remembers earlier sessions: in later runs Lindy referred to earlier runs, so its runs are not fully isolated from each other even in a new chat. One run that reused earlier output was discarded and rerun.
- T3 results: Lindy's T3 is 1/3; its round-3 run was ruled a fail for multiple CTAs in one email. Claude's T3 is 1/3.
Spotted an error? Corrections go to [email protected].
How per-task cost works by pricing model
Credits (Lindy, Manus, Genspark, Relevance AI)
You buy a monthly pool of credits, and each task uses some of it. Cost per task = credits used × price per credit. Credits are each vendor's own unit, so 1 credit on one product is not the same amount of work as 1 credit on another. Watch for rollover rules, top-up prices and what happens when the pool runs dry.
Plans with usage limits (ChatGPT, Claude, Gemini, Devin, Zapier Agents)
You pay a flat monthly price, and usage is capped by time windows, per-feature limits, activities or allowances. On a flat plan, the marginal cost of one more task is close to zero until you hit a limit. A fair per-task figure is therefore the plan price divided by the share of the monthly allowance a task uses, labelled "flat-rate, est." Usage past the limit often moves to API-rate billing or a higher tier.
Per-seat licenses (Microsoft 365 Copilot, Notion)
You pay per user per month, whatever the volume. Per-task cost falls the more each person uses it. Extra metered items, such as custom agents or Notion credits, add on top.
Entry price and what you get
| Agent | Model | Headline price | What the plan gives you | Free tier | Measured cost per task |
|---|---|---|---|---|---|
| Lindy | Credits | $49.99/mo | Credits: everyday asks 2-250, deep work 250-1,000, big builds 1,000-2,500; top-up $10/1,000 credits | $50 free credits for 7 days, no card | ~$0.65 for 3 tasks (list) |
| Manus partial | Credits | $20/mo | Credits consumed by LLM tokens, VM time, third-party APIs; refund on platform failures | Daily refresh credits (Manus 2.0 lite only) | ~$0.86 for 3 tasks (list) |
| Genspark | Credits | $24.99/mo | Monthly credits for AI tasks; weekly Free Quota for Standard mode; core AI Chat/Image models included, flagship models use credits | 100 credits/day up to a lifetime limit | Not tested yet |
| Relevance AI | Credits | Custom / sales | Custom vendor credits | None listed | Not tested yet |
| Zapier Agents | Plan with usage limits | $33.33/mo | Activities: each trigger, knowledge answer, action, browse/scrape, search = 1 activity; separate from Zap tasks | 400 activities/mo | Not tested yet |
| Devin | Plan with usage limits | $20/mo | Usage allowance per plan that auto-refreshes; extra usage at API pricing | Light quota | Not tested yet |
| Claude | Plan with usage limits | $20/mo | Usage limits on rolling 5-hour window plus weekly limits; optional usage credits at API rates | Yes | $0 on Free (capacity-limited) |
| ChatGPT | Plan with usage limits | $8/mo | Usage limits per plan; Business/Enterprise can buy workspace credits beyond included limits | Yes | Not tested yet |
| Gemini | Plan with usage limits | $4.99/mo | Compute-based limits refresh every 5 h up to weekly limit; extendable with AI credits; Flow credits 200/1,000/10,000-25,000 | Yes | Not tested yet |
| M365 Copilot partial | Per-seat | $23.5/mo | Seat license; custom agents metered for Copilot Chat users | Copilot Chat free with eligible M365 plans | Not tested yet |
| Notion AI | Per-seat | $20/mo | Notion credits for Custom Agents ($10/1,000/mo) | AI trial | Not tested yet |
"Headline price" is the figure we use for comparison. It is often per user and sometimes an annual-billing equivalent; see each agent's page for caveats. Last checked 2026-10-16.
The 3 standard test tasks
Every agent runs the same three tasks in a new chat, on the plan being tested, with the default model. We record the usage meter before and after, wall-clock time, human interventions and retries (max 2). Each task runs 3 times, and we report the median. Full rules are in the methodology.
T1: Contract summary
Summarise a fixed 20-page synthetic contract (MSA) in 400 words or fewer, then list parties, term, fees, termination rights, liability cap, auto-renewal and 3 risky clauses with page numbers. Pass: all 7 fields correct and at least 2 of 3 risky clauses match the answer key.
T2: Reconciling two messy CSVs
Reconcile bank.csv (500 rows) against ledger.csv (480 rows). The files contain mixed date formats, duplicates, "$" vs plain numbers and 12 planted mismatches. Pass: at least 11 of 12 mismatches found, no more than 2 false positives, and a downloadable reconciled CSV.
T3: 5-email prospecting sequence
Write a 5-email cold sequence over 14 days for a fictional SaaS product selling to HR directors at 200-person firms, with subject lines, 120 words or fewer per email, one CTA each, and personalisation tokens. Pass: 5 emails, word limits met, and a rubric score of at least 7/10 from two raters. The emails are never sent.
Cost per task FAQ
What does an AI agent cost per task?
It depends on the pricing model: credits used times the price per credit, or for flat-rate plans a share of the monthly price. In our Oct 16-17, 2026 tests (median of 3 runs per task, at list prices), Lindy used about $0.65 and Manus about $0.86 for all three standard tasks combined; Claude Free cost $0 but is capacity-limited.
Why can't I compare credits across vendors?
Each vendor defines its own credit or activity unit, and the same task can use very different amounts. Only running the same task on each agent gives a fair per-task cost.
Which tasks does AgentBill test?
A 20-page contract summary, reconciling two messy CSV files, and writing a 5-email prospecting sequence. Each runs 3 times per agent in a new chat, and we report the median.
See also: pricing comparisons · pricing changelog