I went through the pricing of every app on a16z's agent-economy map. Here is what they charge for, who pays, and four uncommon beliefs.

The sample. About 80 logos on a16z's "Agent Economy – Apps" map, across seven categories. I could name 73 of them and found a public price, or a clearly stated business model, for 60.
The unit. An agent can charge for access, for usage, for work, or for an outcome. Most agents still sell access. The ones with real pricing power sell work or outcomes.
The ladder. General assistants have settled on the same three steps: free, about $20, and $100 to $500. The top step exists because agents keep running when you stop typing.
Who pays. The user is only one option. Hiring, travel and shopping agents get paid by the employer, the merchant or a membership instead.
The argument. Four beliefs have this backwards: that $20 is the price of an agent, that credits are fair, that free wins, and that outcome pricing works everywhere.
What won't last. Flat unlimited plans for always-on agents, credits with no published cost per task, metered plans with no spend cap, and free generalists with no second payer.
The map comes from a16z's Welcome to the B2A Economy, where Olivia and Justine Moore argue that agents, not people, are becoming the main users of the web. It sorts about 80 apps by the job they do: horizontal assistants, work, travel, family, social, hiring and personal finance.
I went through them one at a time. If there was a pricing page, I read it. If there wasn't, I dug through FAQs, App Store paywalls and terms of service, and only then fell back on independent pricing trackers. Seven logos I never managed to put a name to. Another 13 I could name but couldn't find a price for, because they're in private beta or simply haven't published one.
That second number surprised me more than any price on the list. Almost a quarter of the apps on a map of the agent economy haven't decided what to charge yet. The rest use one of seven models.
Pricing models only make sense next to what the product does. Madhavan Ramanujam, who has advised hundreds of companies on pricing, sorts AI products on two questions: how much does it do on its own (autonomy), and how clearly can you credit it with the result (attribution). Here is every named app placed on those two axes, grouped by what it charges today.
Among apps with a published monthly price, the entry plan clusters between $15 and $30. The top plan is where they diverge, from $9.99 for a family organizer to $10,000 for a team buying Tasklet credits.
This is the most crowded part of the map and the most uniform. Nearly every assistant you text has copied the ladder ChatGPT and Claude made familiar. Poke is free, $19 or $199 a month. Hark Pro is free, then $20 for twice the free usage and $100 for ten times. Lucas is free, $20 or $100, and starts every new user on a seven-day reverse trial of the paid plan before stepping them down. folk runs $0, $9.99 and $99.99. Boski does the same thing on a weekly clock: a free allowance that refills each week, then $5 a week for double the usage plus credit packs.
Pickle has the longest ladder on the map: Plus at $20, Pro at $50, Max at $100 and Ultra at $200, all sold as App Store subscriptions. It also does something nobody else does. When you hit your daily limit, you can buy a ticket that resets it, from $1.99 for a half reset to $19.99 for a pack of ten. That's metering without the word credits. You pay only on the days you need more.
asmi has the clearest unit of any assistant on the map. It calls, texts and emails people for you, and it charges by the task: your first task is free, then $10 a month for 20 tasks or $49 for 100. That's about 50 cents per finished job, and you know what you're buying before you pay.
The labs price their agents as a reason to buy a bigger plan. OpenAI's dots come with ChatGPT Pro at $100, $200 or $500 a month and have no price of their own. xAI's Grok Bot also has no standalone price. It comes inside SuperGrok Heavy and Cursor's $120 and $200 plans, with usage past a weekly allowance billed at token cost. Meta's Muse is the exception that prices itself like a startup: free up to 100 million tokens a week, then $20 or $100 a month for more.
The best-funded generalists don't charge at all. Instinct is free during an invite-only beta and says it plans to take a cut of what users buy. Wabi, now valued at $1.25 billion, has no business model on its site. Tab is in private beta with no price. Pally used to charge $25 and $100. Its page now says it is free with near-unlimited use and sells a cloud Mac as a paid add-on.
Underdog is the one outcome price in this row. Its pricing page has no plans; it takes a percentage of the money it saves you, the way Intercom's Fin charges per resolved ticket.
Companions are the other exception. Tomo is $19.99 a month or $119.99 a year. Robyn is $19.99 a month or $199 a year. Bo is the cheapest, at $8 a month billed annually, and it's the only one with a plan for two: Me + a Friend costs $13 a month and gets Bo for you and someone else. They charge like subscription apps because that is what they are.
Work agents are where credits took over. Manus sells 4,000, 8,000 or 40,000 credits for $20, $40 or $200, and a single task can cost anywhere from 200 to 9,000. Tasklet sells 10,000 credits for $25 and 40,000 for $100, with unlimited teammates on one shared pool. Town goes from a free tier to $239 a month and charges 3.2 to 4.4 cents per credit over the allowance. Sauna lists $29, $99, $299 and $999 and doesn't say how much usage any of them includes.
heyclicky shows what a clear ladder looks like. Free gets 25 agent messages a month, Pro at $20 gets 150, and Max at $100 gets 1,000. When you hit the limit, talking keeps working and only agent tasks pause.
The rest price against people. Skydive charges a $20 or $200 platform fee plus model costs at no markup, and its pricing page sets a $6,000-a-year agent against a $190,000 hire. Lindy's page once showed a struck-through $8,000-a-month human assistant above its $49.99 plan. Catch is one price, $99 a person a month with calls included. Howie argues a month of it costs less than an hour of a human assistant.
Joshu goes the opposite way from credits. After a 14-day trial it costs $95 a month with no token fees, no resets and no caps, and it gets its own computer, phone number and email. The only metered part is paid data and media APIs, which come out of a prepaid wallet. It's the cleanest test on the map of whether an always-on agent can survive on one flat price.
The email tools still sell seats. Fyxer is $30 or $50 per user per month. Superhuman Mail is $25 or $33 per seat on annual billing.
Travel agents mostly don't charge for advice. They make money on the booking. Lola, built by Kayak's co-founders with Booking Holdings' backing, is free to use. Members pay $5 or $25 a month on a 12-month commitment for hotel discounts of about 10% and up to 15%. Soar pitches booking at market rates with no added fee. Odessia launched free in June and was acquired by OpenAI five months later.
Arden watches your bookings and claims refunds when prices drop. Its first recovered price drop is free, then it's $15 a month or $99 a year. That's a subscription wrapped around an outcome business, which makes it one of the clearest candidates on the map to charge a share of the savings instead.
Veeza is the clearest outcome price outside hiring. It charges a fixed fee per completed visa application, EGP 1,500 to 4,000 depending on the country, with government fees billed separately.
Family assistants price like household subscriptions. Anna is $20 a month or $192 a year after a 14-day trial. Ohai's App Store listing has three tiers that scale with the household: Premium at $14.99, Premium Plus at $24.99 and Premium Group at $39.99. Orbits sells memberships at $3.99, $9.99 and $89.99. Fambot is free in beta and has said it will charge about what a streaming service costs. Google's CC is a Labs experiment behind a sign-in with no price. Nobody here charges per task, because nobody wants to meter a parent.
Social agents mostly charge nothing, and that makes sense: a social product needs everyone in the group to show up. Alfi's app is free. Texting its number instead of downloading the app costs $5 for 100 messages or 25 images. Ditto, which sets up college students on dates over iMessage, publishes no price. Hue, Anywhere and S_ had nothing public.
Hiring is the most mature category on the map, and it's where outcome pricing is normal. Jack & Jill is free for candidates. In the UK and Europe employers pay 10% of first-year salary only when a hire happens. In the US they pay $199 to $249 per hiring slot per month. Dex is reported to charge employers 20 to 30% of salary. Talentpluto is free for candidates and quotes employers privately.
Refer flips the payer. Its agent works for the candidate, and the candidate pays 20% of their first month's salary after landing the job.
The job-seeker tools sell subscriptions. Tsenta is $19, $39 or $99 a month with 25 free applications. Sonara starts with a $2.95 trial that renews at $23.95 every four weeks. Jobright charges seekers about $30 to $40 a month and recruiters $499 a seat. HiringCafe is free for seekers and charges employers.
Finance is the youngest category on the map. Volaren is free with starter credits, then $12 a month for 1,000 credits, and it is the only company on the map that publishes what a typical action costs: about 27 credits for an AI answer, 50 for an idea-to-trade run and 40 for a backtest. Prices, news, your portfolio and trading never use credits. AnyAlpha sells in-app purchases from $4.99 to $79.99. Scalar Field is listed from $49 a month. Supertake is free while in beta. Catalyst is a waitlist with a free research preview. Conviction had nothing public. Flip, a money manager you text in iMessage, is a plain subscription: $25 a month after a 3-day trial, or $120 a year after a 12-day trial. The yearly plan works out to $10 a month, a 60% discount, and the longer trial makes it the obvious pick. That's a clear push towards the annual plan.
Whoever scales first here will have to choose between a subscription and a share of assets, the same choice robo-advisors made a decade ago. Performance fees, the obvious outcome price, are regulated for retail investors, so this is the one category where outcome pricing is mostly off the table.
A chatbot answers when you ask. An agent keeps working after you put the phone down, and that changes three things at once.
The cost of a task varies by orders of magnitude. A question costs a fraction of a cent. Booking a flight through a browser costs a lot more. Manus's own users report tasks that cost 200 credits and tasks that cost 9,000. A flat price can't cover a 45× spread without charging light users for heavy ones.
Usage doesn't stop when the user does. In an a16z episode on the personal-agent race, Anish Acharya estimated that an ambitious agent can cost around $20 per user per day to run. That's a month of subscription revenue every day. The labs learned this first.
people use it much more than we expected
Sam Altman, on losing money on the $200 ChatGPT Pro plan · Jan 2025Read the source
The interface has no pricing page. Most horizontal agents on the map live in iMessage or WhatsApp. There's nowhere to put a three-column plan table in a text thread. Poke's response was to let the agent set the price: in beta, users haggled with it over the monthly fee, which landed between $10 and $30. It has since moved to fixed tiers. Lucas takes the opposite approach and only mentions usage limits when one actually blocks a request.
The rest of this guide argues with four things people believe about pricing in this situation.
Consumers pay about $20 a month for AI, so an agent should cost about $20.
$20 is the bottom step. Agents earn their money on a $100–$500 step, or by pricing against the human they replace.
The $20 anchor comes from chat. Price a chat assistant at ChatGPT Plus and you look reasonable next to it. But an agent that runs all day costs more than $20 to serve, and it's worth more than $20 to the person using it. The companies that have thought hardest about this put the real plan one step higher.
Pricing against labor goes further. Skydive's pricing page doesn't compare itself to other software. It compares an agent to a hire.
Acharya made the same point from the investor side. He'd rather see a consumer agent people happily pay a lot for than one whose main selling point is a subsidy. He pointed to consumers who already pay $200 to $300 a month for software.
the agent that costs $1,000 a month
Anish Acharya, a16z, on the agent he'd like to see · Sep 2026Read the source
Madhavan makes the case in numbers. On Lenny's Podcast he argued AI products can capture 25 to 50% of the value they create, compared with 10 to 20% for traditional SaaS, and that underpricing early teaches customers to expect less. A $20 agent that saves someone ten hours a month is leaving most of that on the table.
Credits are the fair answer. Users pay for what they use.
Credits are fair only when the user can predict the bill. Most pages don't say what a task costs, and some have no spending cap at all.
Credits solve the vendor's problem. They let a company charge more for heavy tasks without publishing a price for every action. Kyle Poyar has tracked how quickly credits spread, from Microsoft and Salesforce down to startups, and his 2026 survey of 230 software companies found hybrid pricing, a base fee plus usage, rose from 25% to 37% of companies in a year.
But a credit is only as fair as its definition. Town's docs say routines that run on every incoming email are the heaviest setup and that drafting costs more than labeling, but they publish no cost per action. Sauna lists four prices and no allowances. Hark doesn't say whether its "2×" means messages, tasks or minutes. An eesel reviewer who went through several of these pages concluded the usage limits matter more than the sticker price, and the pages rarely publish them.
Metering with no ceiling is worse. Usage of Grok Bot past the weekly allowance is billed at model cost, and xAI's own documentation, quoted by eesel, admits there is “no Grok Bot-specific spend cap yet.” An agent that runs unattended and bills by the token is how a user wakes up to a surprise.
Volaren is the model to copy. It doesn't just sell credits; it tells you roughly what each kind of action costs and which actions are free. Tasklet publishes an exchange rate and doesn't charge for extra teammates. Town's overage cap defaults to $0, so nothing extra is charged until you raise it.
The agent race goes to whoever gets the most users, so the winner will be free.
Free works when a second party pays. Without one, free is a bet that inference costs fall before the money runs out.
Plenty of agents on the map are free. In the a16z episode, the hosts noted that of 122 assistants David Pawlan catalogued, the best-funded generalists are the free ones. They also called that a subsidy, and a fragile one. Tom Shaughnessy of Delphi Ventures has made the same argument about the labs: consumer subscriptions are priced far below what businesses pay for the same models through the API, a gap he puts at 40×. A free agent built on top of those prices is subsidized twice.
The free agents that hold up all have someone else paying.
Taking a cut of transactions has a catch, and the user should know it. Instinct's users reportedly spend over $1,300 a month through it. A cut of that is worth more than any subscription, but it also means the agent earns more when you spend more. A shopping agent paid by the transaction has a reason to recommend buying. Underdog's version points the other way: it earns more when you spend less.
The a16z hosts saw durable value elsewhere, in narrow agents that charge a lot to a small group. Acharya noted that a business can now reach a $100 million run rate with tens of thousands of customers. Pawlan's example was an assistant only for single mothers of children under three. Nothing on the map is that narrow yet. Veeza, a visa agent for people with weak passports, comes closest.
Every agent will eventually charge per outcome.
Outcome pricing shows up only where the outcome is an event with a clear moment and a clear value: a hire, a visa, a refund.
The strongest advocate for outcomes is Bret Taylor, whose company Sierra charges customer-service clients when the agent resolves an issue.
It's just so obviously the correct way to build and sell software.
Bret Taylor, Sierra, on Lenny's PodcastRead the source
Madhavan's framework says where that works. Outcome pricing belongs in the corner where autonomy and attribution are both high. Each of the other corners has its own best fit. Earlier I placed every app where it sits today. Below is my reading of his framework applied to the map: what each corner should charge, who is already there, and who I think should move.
Hiring is in the hot corner because a hire is a single event with a salary attached. That's why Jack & Jill can charge 10% and Refer can charge the candidate. Arden and Underdog get there too, because a refund has a dollar amount. An emptier inbox has no such moment. Nobody can say which reply was worth what.
Even in business software, outcome deals can drift back to fixed fees. Jason Lemkin of SaaStr wrote that one outcome-based contract in his fund crossed $1 million a year and the customer quickly moved it to a fixed contract. His warning is short: “a pricing model is not a product.” Kyle Poyar's test for outcome pricing asks whether outcomes are consistent, attributable, measurable and predictable. Most agents fail at least two of those.
Flat, unlimited plans for always-on agents. If $200 a month didn't cover heavy chat users, it won't cover agents that run all night. Almost every assistant on the map that started unlimited has added a cap, a usage multiple or a credit pool. Joshu is betting the other way, and it's the one to watch.
Credits without a cost per task. If a user can't estimate a bill before running a task, they stop running tasks. Publish a rate, show the cost before an expensive run, and default to a cap.
Metering with no spend cap. An unattended agent that bills by the token needs a ceiling the user sets. Without one, the first surprise bill is also the last.
Haggling as a business model. Poke's negotiated pricing was a great launch story. It moved to three fixed tiers once it had to serve millions of messages.
Free generalists with no second payer. They depend on costs falling faster than funding runs out. Several have already been acquired: Poke by Cognition, Odessia by OpenAI.
Standalone versions of a feature the platforms will bundle. Calendly launched an assistant you CC on email, the same thing Howie sells. The labs put agents inside $100-plus plans. Pawlan expects at most one independent travel agent to survive on its own.
Pricing has a shorter shelf life. Poyar's latest data says pricing used to last about 18 months and now lasts about six. The map shows it. In 2026 Lindy dropped its free plan, then moved its entry price from $49.99 to $29.99. Pally went from $25 and $100 plans to free. Poke went from haggling to tiers. Plan for your pricing to change, and tell users when it does.
The labs set the ceiling. When ChatGPT Pro includes an always-on agent for $100 and Grok Bot comes with Cursor, a standalone generalist has to be cheaper or better at one job. Specialists don't compete with those plans. Generalists do.
The second payer is the moat. The categories with the clearest pricing (hiring, travel, visas) are the ones where money already changes hands and the agent can take a share of it. Look for the transaction in your category before designing a plan table.
Trials are moving inside the product. Lucas starts everyone on the paid plan for a week and steps them down. Arden gives away the first refund. Both let the user feel the paid version before they're asked to pay, without a separate trial screen.
Your next customer may be an agent. The a16z piece describes an agent that upsold the authors on a Descript plan to finish a video. If agents choose which products to subscribe to, a pricing page that only a person can read loses sales. Publish prices an agent can read and compare.
Find the moment of value. Write down the single event your user would pay for: a meeting booked, a hire made, a refund recovered. If you can't name one, you're selling access.
Place yourself on the two axes. How much does your agent do alone, and how clearly can you credit it with the result? The "What each app should price as" chart tells you which model fits.
Measure your cost per task, not per user. Run 50 real tasks and record the cost of each. The spread between the cheapest and most expensive decides your model more than any competitor does.
Look for a second payer. If an employer, merchant or partner already pays for the outcome, charge them and keep the agent free for the user.
If you sell credits, publish the rate. Show credits per dollar, typical cost per task, and an estimate before expensive runs. Default the overage cap to zero.
Price the top step for heavy users. Your heaviest 5% decide your margin. Give them a $100–$200 plan with a multiple of the allowance rather than unlimited.
Re-check every quarter. Model costs and competitors move every few months. So should your prices.
One row per named logo. Prices are monthly in USD unless stated and were checked between September and November 2026 against company pages, FAQs, App Store listings and the sources linked below. "Not public" means private beta, waitlist, or no price on the site. Seven unnamed logos are left out.
It's a structure users already know from ChatGPT and Claude, so it needs no explanation in a text thread. The middle step catches casual users. The top step is where the margin comes from.
Only where the outcome is a countable event with a clear value, like a hire, a visa or a refund. For general assistants the result is spread across many small actions, and no one can price each one.
Launch free if someone other than the user will pay later, such as employers, merchants or members buying perks. If the user is the only possible payer, charge early. Madhavan's point is that underpricing trains users to expect less.
No. Credits are a reasonable way to handle tasks whose cost varies widely. They become a problem when the cost per task is hidden and there's no spending cap. Volaren shows how to do it well.
Many are in private beta, and the category is young. A lot of founders are also waiting to see whether the labs give agents away inside existing plans before they commit to a price.
Only compare where the agent finishes the job without supervision. Comparing a $6,000 agent to a $190,000 hire is persuasive only if the agent actually replaces that hire's work.
Welcome to the B2A Economy, Olivia Moore & Justine Moore, a16z (source of the market map and the logos used in the figures)
The Personal Agent Race Is Here, Anish Acharya & David Pawlan, The a16z Show (summary)
Pricing your AI product, Madhavan Ramanujam on Lenny's Podcast
Bret Taylor on Lenny's Podcast (transcript)
Why everyone's switching to AI credits and Growth Unhinged pricing archive, Kyle Poyar
Why the AI business model is cracking, Tom Shaughnessy (Delphi) on Unchained
The real question for outcome-based pricing, Jason Lemkin, SaaStr
OpenAI is losing money on ChatGPT Pro, TechCrunch
Company pricing pages and FAQs: Poke, Tasklet, Sauna, Muse (eesel), Pally, heyclicky, Volaren, Arden, Anna, Tab, asmi, Joshu, Flip, Underdog, CC (Google Labs), Lucas and Boski FAQs; App Store listings for Ohai and AnyAlpha
Independent pricing write-ups: eesel on dots, Grok Bot, Town, folk, Fyxer, Instinct; Carly on Hark, Manus, Catch
Travel: Lola tiers, OpenAI acquires Odessia (Skift), Veeza fees
Hiring: Jack & Jill US/EU pricing, Dex, Refer, Sonara
Some figures come from independent trackers rather than a company's own page, and agent pricing changes often. Treat every number as of the date checked. Logos belong to their companies and are reproduced from the a16z map for identification.
A practical guide to the six principles under the prompt box

How to turn individual workflows into a shared workspace, especially inside design teams

Everything you need to get started with an agentic design system.


From ChatGPT to Figma AI, explore the best AI UX patterns from leading products.