How Much Does a Personal AI Assistant Cost in 2026?

A personal AI assistant costs about $20 a month on a consumer plan, roughly $9 to $47 a month in model tokens alone if you build your own at published 2026 API rates, or a flat fee in the mid $50s on a managed open stack. The middle band is the one that surprises people, because the token bill is the part nobody budgets and the only part that grows as the assistant becomes genuinely useful.

Everything below is arithmetic on published list prices, not estimates. If you want the wider question of what you are actually buying, start with our guide to what makes a personal AI assistant yours.

The three price bands at a glance

Personal assistants fall into three pricing shapes, and the shape matters more than the headline number. Two are predictable and one is not.

Band	What you pay	What varies with use	Who it fits
Consumer subscription	About $20 a month	Usage caps, not the price	You want it working today and accept closed, metered inference
Do it yourself	A server plus a token bill	Almost everything	You enjoy operations and want total control
Managed open stack	A flat monthly fee	Nothing	You want the assistant to be yours without running it

Bands one and three are easy to plan around. Band two is where the real question lives, so that is where the math goes.

Band 1: what consumer plans actually charge

Individual plans from the mainstream assistants cluster tightly around $20 a month. Anthropic lists Claude Pro↗ at €18 a month billed monthly, or €15 with annual billing, and ChatGPT Plus and Google AI Pro sit at roughly the same figure.

Power-user tiers climb steeply from there. Claude Max starts at €90 a month for five or twenty times Pro usage, and team seats list at €18 each.

What that price buys is capability, not ownership. The model is closed, the inference runs on the vendor’s servers, and the usage cap rather than the price is what moves when you lean on it. For many people that is a perfectly good trade.

The catch appears when the assistant stops being a chat window and starts being infrastructure you depend on. At that point the question changes from “how much” to “what happens if this is withdrawn, repriced, or restricted”.

Band 2: build it yourself and rent the model

The do-it-yourself route splits into two very different bills, and people consistently misjudge which one dominates. The server is the small one.

A small cloud instance is enough to run an open chat client and the orchestration around it, because the heavy computation happens wherever the model runs. If that is a hosted API, your server is essentially a message router.

The token bill is the large one, and it is the only line that scales with how much you actually use the thing. Most cost guides stop before this point, which is why the do-it-yourself route keeps getting described as cheap.

There is a third bill in hours. Setup, updates, breakage and migrations are real costs even when no invoice arrives, and our OpenClaw setup guide is honest about what that first week looks like.

The token math, worked out

Here is the calculation nobody publishes, using Meta’s own figure for conversation size and Anthropic’s published API rates.

Meta reports that a typical assistant conversation runs 20,000 to 25,000 tokens, which is why its Business Agent costs $0.04 to $0.05 per conversation at $2.00 per million tokens. We covered that pricing change in Telegram versus WhatsApp for an AI assistant.

Take a modest habit of 10 conversations a day at 22,500 tokens each:

• 225,000 tokens a day, or 6.75 million tokens a month
• Split roughly 90% input and 10% output, because every turn resends the system prompt and the conversation so far: 6.08M input, 0.68M output

Applying Anthropic’s published rates↗:

Model tier	Input / output per million	Monthly token cost
Haiku 4.5	$1 / $5	$9.45
Sonnet 5 (introductory)	$2 / $10	$18.90
Sonnet 5 (standard)	$3 / $15	$28.35
Opus 5	$5 / $25	$47.25

That is ten conversations a day. Triple it to thirty, which is what an assistant you actually rely on looks like, and Sonnet-class usage reaches $85 a month in tokens alone, before the server and before your hours.

Does a cheaper model tier fix it?

Partly, and it is the most under-discussed lever in the whole calculation. Changing tier moves the bill further than any other single decision.

At Haiku-class rates the same ten conversations a day cost $9.45 a month, less than half a consumer subscription. At Opus-class rates the identical usage costs $47.25, five times as much for exactly the same conversations.

So “how much does a personal AI assistant cost” has no single answer even with usage held constant, because the same habits span a fivefold range depending only on which model answers.

The practical version is routing: a cheap model for most turns and an expensive one for the hard questions. It works well, and it is one more system you now own and maintain.

Why prompt caching does not rescue the number

Prompt caching genuinely cuts repeated input, and it is the first objection people raise. It just does not fit the way a personal assistant is used.

The saving is real where it applies. Sonnet 5 cache reads cost $0.20 per million tokens against $2 for fresh input, a tenfold difference on the largest part of the bill.

The limit is the cache window, which defaults to five minutes. A personal assistant is used in short bursts spread across a whole day, so caching helps inside a single burst and does nothing between them.

Budget as though most turns miss the cache, then treat any savings as upside. Assuming the opposite is how people end up with a bill three times their estimate.

The price change on September 1, 2026

Here is the structural problem with a metered assistant, and it has a date on it. Sonnet 5’s $2 and $10 per million tokens is introductory pricing through August 31, 2026, and the standard rate from September 1 is $3 and $15.

Same assistant, same habits, same conversations. The monthly token bill in the example above goes from $18.90 to $28.35, a 50% rise you did not choose and cannot decline.

That is not a criticism of the pricing, which is public and clearly flagged. It is the point: on a metered plan, the price of the thing you depend on is someone else’s decision, exactly like the unpredictable API bills we have written about before.

Flat pricing is not cheaper by definition. It moves that particular risk off your side of the table.

Band 3: what a flat managed plan costs

Tessera’s Founder plan is $55 a month, or 55 € in Europe, with no token meter and a pace limit of 20 requests per minute. Annual billing is $47 a month, which is $564 a year, and the founding price is locked for the first 25 seats.

Included at that price: the install, updates and monitoring, an isolated instance on servers in Europe with an encrypted disk, and all five open models. Qwen3.6-35B-A3B for chat, Whisper large-v3 for transcription, Kokoro for speech, plus embeddings and a reranker.

The models matter to the cost question specifically. They are Apache 2.0 and MIT licensed, so the thing you depend on cannot be withdrawn or repriced out from under you the way a closed model can.

For what companies pay at larger sizes, our private AI pricing breakdown covers the rest of the ladder.

The costs that never make the spreadsheet

Four line items are missing from almost every comparison, and together they usually exceed the ones people do count.

• Voice. Transcription and text to speech are separate meters on most stacks. If you want to send a voice note and hear a reply, price those too.
• Memory and storage. Long-running assistants accumulate history, and retrieval over it is another set of calls.
• Your hours. A weekend of setup plus an hour a month of maintenance is real money at any sensible rate for your time.
• Migration. When a closed model is deprecated, you rewrite prompts and retest behaviour. That cost arrives on the vendor’s schedule.

The build versus buy breakdown puts numbers on the last two for larger deployments, and the logic scales down cleanly to one person.

How much does a personal AI assistant cost per year?

Annually the three bands separate clearly, and the do-it-yourself range is wide because it depends entirely on how much you use it.

Option	Yearly cost	Notes
Consumer subscription	About $240	Closed model, vendor servers, usage caps
DIY, 10 conversations a day	About $340 in tokens	At Sonnet-class standard rates, plus server and hours
DIY, 30 conversations a day	About $1,020 in tokens	Same rates, plus server and hours
Managed open stack, annual billing	$564	Flat, no token meter, open weights

The honest reading is that a managed open plan costs more than a consumer subscription and less than heavy do-it-yourself use. What you buy for that difference is open weights and a bill that does not move.

Which band fits you

Pick by how you expect to use it, not by the smallest number on the page. The three cases are genuinely distinct.

Take a consumer plan if you want the strongest closed model with zero setup, your use is bursty, and it does not matter to you whose servers it runs on. This is the right answer for most casual use.

Build it yourself if operations are something you enjoy, you want to swap models freely, and your usage is low enough that the token bill stays under a consumer subscription. Be honest about the hours.

Take a flat managed plan if the assistant is becoming infrastructure, you want open weights so it cannot be withdrawn, and you would rather pay a known number than watch a meter. That is what Founder is for.

FAQ

Is there a free personal AI assistant?

Yes, in two forms. Vendors offer capped free tiers on closed models, and fully local stacks cost nothing but electricity if you already own hardware that fits the model. Neither is free once your time is priced.

Why is my AI assistant bill higher than I expected?

Almost always because every turn resends the system prompt and the whole conversation so far, so cost grows with conversation length rather than with the number of messages you type.

Does a personal AI assistant cost more than ChatGPT Plus?

At light use on a cheap model tier, a self-built assistant costs less than a $20 subscription. At the volumes people actually reach once it works, it costs several times more.

Can I cap what a personal AI assistant costs?

Only two ways. Hard spend limits on the API, which stop the assistant when they trigger, or a flat plan with no token meter, where the cap is the price itself.