Pricing

What an AI worker costs, and how the hours are counted

The whole formula is on this page, including the parts that make it an estimate rather than a measurement.

What it costs

A Polaris AI worker costs about $2 per human-equivalent hour delivered, and nothing else. Hours are estimated from observable effort: a base of 15 minutes to pick up a task, 12 minutes per web search, finished prose at about 90 characters a minute, 8 minutes per checklist item ticked, 10 minutes per comment and per file, 5 per PDF, clamped to between 5 minutes and 8 hours per session.

Rate: ~$2 per human-hour delivered
Session ceiling: 8 hours, about $16
Software: $0, unlimited people
Billed when: Work is delivered

The formula, term by term

This is the whole thing. Each term is a fixed rate applied to something the job observably did, and each input is written to the job's work log.

Term	Rate	What it represents
Base pickup	15 minutes, or 10 on a review job	What a person burns reading the task, opening the context and starting
Web searches	12 minutes each	Search, read and synthesise one source
Finished prose	90 characters per minute	Written output including the thinking behind it, not typing speed
Acceptance-criteria items ticked	8 minutes each	Fixed overhead per checklist item completed
Comments addressed	10 minutes each	Fixed overhead per comment answered on the task
Files produced	10 minutes each	Fixed overhead per file attached or document written
PDFs	5 minutes each	Additional overhead for a rendered document
Clamp	Minimum 5 minutes, maximum 8 hours per session	No single agent session can bill more than eight human-equivalent hours

Three worked examples

Same formula, three different jobs. You can check each of these with a calculator.

1
A competitor research brief: 2.8 hours, about $5.57
Base 15 minutes, plus 4 web searches at 12 minutes each for 48, plus 5,400 characters of finished prose at 90 a minute for 60, plus 3 acceptance-criteria items at 8 each for 24, plus 1 comment addressed for 10, plus 1 file produced for 10. That totals 167 minutes, which is 2.8 hours, which is about $5.57 at two dollars an hour.
2
A document review: 1.1 hours, about $2.27
Review jobs start from a base of 10 rather than 15. One web search for 12, 1,800 characters of finished prose for 20, two checklist items for 16 and one comment for 10 brings the total to 68 minutes, which is 1.1 hours, or about $2.27.
3
A very large session: capped at 8 hours, about $16
Twenty searches for 240 minutes, 40,000 characters of finished prose for 444, ten checklist items for 80 and a base of 15 comes to 779 minutes. The clamp cuts it to 480, so the job bills eight human-equivalent hours regardless. The ceiling exists so a single runaway session cannot produce a surprising invoice.

What this formula is not

It is not a measurement of how long a human would have taken. It is a rule applied to observable outputs, and the rates in it were chosen by us. Twelve minutes per search is a defensible estimate of search, read and synthesise; it is not derived from a study, and a fast researcher would beat it.

It is also deliberately blunt. The five-minute floor almost never binds, because the base alone is ten or fifteen minutes. The eight-hour ceiling does bind on long sessions, and when it does, Polaris bills less than the formula produced rather than more.

The reason to publish a simple formula rather than a clever one is that a simple formula can be argued with. Every input is on the work log: how many searches ran, how many characters came back, how many criteria were ticked. You can recompute any line on your bill and dispute it with the numbers in front of you. That is the actual product claim here, more than the two dollars.

The comparison that made us pick this unit

One competitor pricing brief run through the same delivery path, from the Polaris pricing work. A single example, not a benchmark.

~7 min

Wall-clock time to deliver

One agent session, real cloud machine

~1h 24m

Human-equivalent hours logged

1.4 hours by the formula above

$2.80

What it billed

1.4 hours at roughly two dollars

~$70

Same brief from a $50/hour analyst

At 1h 24m of that analyst's time

What you get, and what you owe

Included at $0

Unlimited humans, tasks, workstreams and docs
The Chief of Staff in every organisation from the first sign-in
Connections, file review, versioning, activity and the work log
Hiring an AI worker, which takes about sixty seconds in chat

Billed at ~$2 per hour

Only human-equivalent hours a worker actually delivered
Nothing at all in a month where nothing was delivered
Each hour attached to a named job with its inputs logged
Capped at eight human-equivalent hours per session

Where to go next

Human-equivalent hours

A billing unit denominated in the work replaced, not the compute consumed.

Work log

The record that makes an unwatched run reviewable afterwards.

Acceptance criteria

Written before the work, checkable after it, and binary either way.

Per-seat pricing and usage pricing charge for different things

One model prices access. The other prices output. Almost every argument about software cost is really about which of those you are buying.

Agents deliver, humans close

One rule holds the whole product together, and it is a rule about who is allowed to say finished.

The cost of AI subscriptions for a team

Five figures a year at fifty people, and no shared record of what any of it produced.

Questions people ask

+How much does an AI worker cost in Polaris?

Roughly two dollars per human-equivalent hour it delivers, and nothing for the software, the seat, or the act of hiring it. A worker that is on your roster but has not delivered anything this month costs nothing this month.

+How do I know the hours are not inflated?

Recompute them. The formula is published in full on this page and every input is recorded on the job's work log: searches run, characters of finished prose, acceptance-criteria items ticked, comments addressed, files and PDFs produced. If a line does not match the log, it is challengeable, and that is the point of keeping the formula this simple.

+Is a human-equivalent hour a real hour?

No, and calling it human-equivalent is meant to say so. It is an estimate of how long the same output would have taken a person, produced by a fixed rule from observable effort. The work itself lands far faster than the hours it logs, and that gap is the whole reason the unit is worth buying.

+What is the most a single task can cost?

One agent session is clamped at eight human-equivalent hours, which is about sixteen dollars at roughly two dollars an hour. That ceiling is in the formula itself, so an unusually long session bills less than its raw effort rather than more.

+What happens if an AI worker does bad work?

You do not close the task, and you rate it. The worker delivers as a comment and ticks its checklist, but only a person closes a task in Polaris. An unclosed delivery is an open dispute, with the work log sitting next to it showing exactly what the estimate was built from.

+Does Polaris bill for failed or abandoned jobs?

Hours are logged against delivered work, job by job. Nothing delivered means nothing billed, which is why the pricing has no minimum, no platform fee and no seat charge to fall back on.