Is Poe worth paying for? Do the points math first, then judge a bot by what it charges

It's 11:40pm and you have 3,000 points left. You have now asked the same question three times — the first answer was too short, the second too long, and the third you sent without changing a single condition, hoping for different wording. The notification that your points are gone arrives the next morning.

This piece does not tell you to subscribe, and it does not tell you not to. It does two things: it lays out how Poe's points actually work, using the official help centre text; and it gives you a set of standards you can check yourself, to decide whether a particular bot deserves the points it charges.

1. Points are not an allowance. They are a per-message bill.

Poe's official FAQ is refreshingly plain about it: "Bots consume points when they generate responses. Each user has a set of points that can be used flexibly across all bots on Poe. Each bot shows the points required (for example, image generation costs a fixed number of points, or varying amounts depending on text length)."

Three things are easy to skim past there. First, points are one pool, not a per-bot allowance — every point you spend on bot A is a point bot B does not get. Second, the billing unit is a message (one response), not a month of access: "I subscribed, so I can use it freely" is a misunderstanding. A subscription gives you more points, not unlimited ones. Third, the unit price differs per bot and can float: image bots commonly charge a flat rate, while chat bots charge by how much they read and write.

As for what a point is worth, the official help centre offers a ruler: Poe shows a reference purchase price of about $30 per 1,000,000 points, and the purchases FAQ describes add-on points as starting at $30 per 1,000,000 points, with no transaction fee. Note the phrase "reference price": it is not the per-call price, because each bot has its own rate card. But used as a ruler it means one point is roughly $0.00003, so you can translate "this costs 2,000 points" into "about six cents" and compare it with the alternatives.

Free users and subscribers differ in the rhythm of issuance

• Non-paying users: receive points daily automatically; they reset at the end of each 24-hour cycle and do not accumulate. When they run out you either wait for the next day or buy more immediately.
• Subscribers: points are granted daily or monthly depending on the plan; unused points do not roll over unless the plan says otherwise; add-on points can be purchased.
• Add-on points: valid for one year, non-refundable, non-transferable, not convertible to cash; bought on the Poe API key page. Your balance is viewable as plan points and add-on points separately.

On the price of the subscription itself, the official help centre gives exactly one figure: from $4.99/month ("access to all the leading models and more, from just $4.99/month"). How many tiers there are, how many points each grants, and how annual billing works are not listed officially — I am not going to guess. Section 5 says where to look at the live numbers.

One more thing matters if you share a subscription: group members' point usage is not individually capped (once added, they can use 100% of the subscriber's points), and only the subscriber can buy add-on points. Sharing means "we all drain one pool together", not "each of us gets a slice".

2. Before you spend, there are three official places to see the price

1) Before: the rate button at the top of the chat. The official description: whether you start a new chat or open an existing one, the top of the chat shows the bot description including the approximate points for the first message and a "Rate" button; opening it shows how the point price is calculated. This is the only pricing statement you can see before pressing send.

2) During: budgets. Poe supports setting a budget per message: if a message would exceed it, Poe notifies you and you decide whether to raise the budget for that chat; a single message will not exceed the budget unless you confirm it. The global budget lives in Settings; a per-chat budget lives in that chat's settings.

3) After: the breakdown. For a single message, hover over it and choose Info from the three-dot menu (long-press on mobile) to see the points that message cost and how that number was composed. For the whole picture: open Poe.com in a browser → Settings → Points → View point history.

3. Six checks you can run yourself: is this bot worth the points?

① Is the rate flat, or does it float by volume?
Flat rates (common for image and video) are easy to budget. Volume-based rates (chat) mean the longer the context, the higher the unit price — the same bot can differ several-fold between a short prompt and a long conversation.

② Does it let you look before you pay?
Some bots publish examples or recorded runs, so you can see real output before spending your own points. Something that is willing to show its work is usually more reliable than something that only says it is "the most powerful".

③ How does it bill when it fails?
Timeouts, errors, off-target answers — whether points are charged in those three cases is the clearest signal of whether a bot respects its users. An honest implementation answers "was I charged?" per case, rather than leaving you to guess from your point history. You can test this directly: send a request you expect to fail, then open Info and look at the breakdown.

④ Does it give a conclusion, or also how certain it is?
"Pick A" and "Pick A, confidence 0.93" are different products: the first leaves you all-in or all-out, the second lets you decide whether spending another message to confirm is worth it. And if a bot is equally confident about everything, its confidence carries no information.

⑤ Do its numbers have a source?
When you see "95% accuracy", ask three questions: measured on which benchmark, over how many items, and how many failed? A percentage that cannot answer those three is not a percentage.

⑥ Is its context policy saving your points or spending them?
Poe has an automatic context management toggle. When it is on, Poe controls how much history is sent each time, prioritising recent exchanges. The official wording is blunt: with it on you can chat longer without consuming as many points, and in most cases you will not even notice a difference in the responses; turn it off and the bot remembers earlier content, but the total points per message go up. The official note adds that retaining everything usually does not improve results. So "I want it all" carries a price tag under this billing model.

One thing we build happens to sit on ④ and ⑤

By the standards above, we are being checked too. What we make is Decider, a judge bot on Poe: give it a question and two candidate answers, and it returns which one is stronger, with a calibrated confidence value and a written rationale.

Let me draw the boundary first: on raw pick rate we claim no advantage. On JudgeBench we measured 92.5% under the official protocol (620 judgments, 6 failures, disclosed), against 92.2% for a plain direct baseline — that is a tie, not a win. Our only differentiation is whether the confidence can be trusted: in the same recorded runs, judgments reported at 90%+ confidence were right 99.6% of the time, and those in the 80–90% band were right 94.0% (same 620 judgments and 6 failures; the benchmark, the method and the failure count are all disclosed). In other words, it does not sell "I am right"; it sells "I know how likely I am to be wrong this time" — and that is what helps you decide whether to spend another message.

Before spending any of your own points, you can read its 27 recorded cases in the open, or open it directly here: open Decider on Poe.

4. Seven situations where the points you spend are probably wasted

1) Re-asking the same question without changing a single condition.
Unless what you want is precisely "let me see if it is stable", you are buying random numbers. Note the nuance: if you genuinely resend identical input, with unchanged context, within about five minutes, Poe's cache discount may kick in automatically; the waste is mainly when you change the wording but not the information — no new information, and you lose the cache discount.

2) Switching topics inside a long conversation.
The first official optimisation tip is: when you change topic, clear the context or start a new chat. Chat bots bill by what they read, so dragging a long history about A along to ask about B is paying for irrelevant content.

3) Leaving automatic context management off as your default.
The official position is explicit: retaining everything consumes more points and usually does not improve results. Unless you genuinely need it to remember something from long ago, this is the easiest ongoing leak to miss.

4) Spending points on work that needs no judgement.
Classification, extraction, format conversion — tasks with a definite answer do not need the most expensive tier. Using it there is paying for capability you do not use.

5) Using a chat bot for exact arithmetic, counting, and date comparison.
Poe's own FAQ says it plainly: bots are usually correct but can also make mistakes; before relying on one for an important decision, it is worth verifying elsewhere; and do not treat its output as medical, legal or investment advice.

6) Doing work you will need to revisit inside a temporary chat.
Temporary chats consume points all the same; their content is permanently deleted within 24 hours of creation and does not enter Memory, cannot be searched, cannot be shared. Good for a one-off question, not for something you need to look up tomorrow.

7) Forgetting the conditions on the cache discount.
Officially: when you send messages consecutively (generally no more than five minutes apart), when previous messages have not been edited, and when enough of the repeated input overlaps, some points are discounted automatically; the common discount rates are 50% and 90% (each applying to some models — check the rate card for that bot). "Come back every ten minutes to ask one thing" and "ask everything in one sitting" can differ twofold in cost.

5. So is the subscription worth it?

Compute it rather than taking anyone's word:

1. Look at how many calls you actually needed last month and roughly how many points each took — the Info breakdown and point history already have this.
2. Multiply out your monthly consumption and compare it with what your plan issues: plan points + add-on points.
3. Use $30 per 1,000,000 points as the ruler, translate the shortfall into money, and compare that with the alternatives.

If you only use one vendor's models a few times a week, a subscription is probably not worth it. Poe's own comparison makes the value proposition clear: other vendors' $20/month subscriptions cover only their own models, whereas Poe's claim is one subscription spanning several vendors, starting at $4.99/month. So the deciding variable is not the price — it is whether you actually move between multiple models and capabilities.

Several things I could not verify, so I am not writing numbers for them:

• Tiers and the points each grants: not listed in the official help centre. The live numbers are in Poe's Settings and "Get more points" pages.
• Regional price differences: the help centre has no regional price list, only a list of which countries and regions can subscribe; I did not find mainland China on that official list (the list is official text, and it can change).
• Promotions: no promotional information in the official help centre. Any discount anyone quotes is less reliable than the number on the checkout page.
• Refunds: except as the Refund Policy provides or Quora decides otherwise, subscription charges and other final charges are non-refundable; contact Poe for web subscriptions, and Apple/Google for their stores. After cancelling, access lasts to the end of the current billing period, and you can only upgrade within the same platform.
• How many points a given bot charges: that lives in the in-product rate card, not in the help pages; trust the Rate button at the top of the chat.

To close

The honest thing about a points system is that it turns "using AI" from an all-you-can-eat monthly fee into a per-question price. That is uncomfortable in the short run and good in the long run — it forces you to ask better questions.

So next time, before you press send, look at two things for ten seconds: the Rate button at the top, and what you actually want out of this question. The first decides whether it is worth it; the second decides whether you need to press send a second time. And when you pick a bot, check three things: does it show you the price before you spend, does it tell you whether you were charged when it fails, and does its conclusion come with how certain it is.