Intelligence · AI search
Your own OpenAI API key instead of a subscription: what a second brain chat costs
With your own API key and a small model, chatting with your notes can cost cents a month. In our worked example, embedding a million words with text-embedding-3-small costs about 3 cents once, and a GPT-5 mini question about a tenth of a cent, at OpenAI's prices as of November 2026.
If you want a second brain app where you bring your own OpenAI API key instead of paying a subscription, the key is usually the cheaper way to chat with your notes. In our worked example, embedding a library of a million words with text-embedding-3-small costs about 3 cents, once. A question that sends six passages to GPT-5 mini costs about a tenth of a cent. At OpenAI's prices as of November 2026, a $20 monthly plan only pays for itself after thousands of questions a month, although a plan also buys a ready-made chat app and a bill that stays the same.
Every price below comes from OpenAI's and Anthropic's own pricing pages as we read them in November 2026, and each page is linked. Word counts, token counts and question counts are assumptions we chose for the arithmetic, and they are labeled that way. We ran no tests. We make Barside, a free desktop app for Windows, Linux and Mac that runs on the AI account you connect, and the last part of this guide shows how it handles the same costs, with its limits. For app-by-app subscription prices, read AI note-taking app pricing in 2026, and for how keys and model switching work, read AI notes apps where you bring your own key.
Should a second brain app run on your own OpenAI API key instead of a subscription?
A second brain app that runs on your own OpenAI API key instead of a subscription makes sense when most of what you want is to ask questions of notes you already have. Chatting with your notes, the setup usually called RAG (retrieval-augmented generation), runs on two meters. The first is a one-time cost to turn your notes into embeddings, lists of numbers that let software compare meanings. The second is a small cost for each question.
With the small models priced below, both meters stay tiny next to a monthly plan. A subscription earns its price in other ways: a chat app with no setup, a bill that never changes, and features that reach beyond your notes. Three things decide it for you:
- How many questions you ask. The API bills each one, so a quiet month costs close to nothing.
- Which model answers. In the example below, Claude Opus 5 costs fifty times what Claude Haiku 5.5 does for the same question.
- What else you use the plan for. If you already pay for a chat plan for other work, chatting with your notes can ride on it.
What OpenAI and Anthropic charge per million tokens
API prices count tokens, the small pieces of text a model reads and writes, and both companies quote them per million. The figures here are OpenAI's Standard prices and Anthropic's base prices, as of November 2026. OpenAI's API pricing page shows them with a dollar sign, and Anthropic's pricing documentation says all its prices are in USD.
OpenAI's official API pricing for embeddings, GPT-5 mini and GPT-4.1 mini in 2026
OpenAI's official API pricing for 2026 puts embeddings, GPT-5 mini and GPT-4.1 mini on one page, under the line "Prices per 1M tokens." As of November 2026, it lists these Standard prices:
| OpenAI model | Input | Cached input | Output |
|---|---|---|---|
| text-embedding-3-small | $0.02 | - | - |
| text-embedding-3-large | $0.13 | - | - |
| GPT-5 mini | $0.25 | $0.025 | $2.00 |
| GPT-4.1 mini | $0.40 | $0.10 | $1.60 |
| GPT-4o mini | $0.15 | $0.075 | $0.60 |
The same page lists Batch, Flex and Fast prices for many models. GPT-5 mini, for example, costs $0.125 for input and $1.00 for output per 1M tokens on Batch, which suits work that can wait. For a question you sit and wait on, plan with the Standard row.
OpenAI embedding API pricing per token
OpenAI's embedding API pricing per token in 2026 is $0.02 per 1M input tokens for text-embedding-3-small, as of November 2026, which is $0.00000002 a token. The output column for embedding models shows a dash. OpenAI's embeddings guide says "Usage is priced per input token" and, assuming about 800 tokens per page, puts text-embedding-3-small at about 62,500 pages per dollar and text-embedding-3-large at about 9,615. It also says a text-embedding-3-small vector has 1,536 numbers by default.
GPT-4o mini API pricing for input and output tokens
GPT-4o mini API pricing for input and output tokens is $0.15 per 1M input tokens and $0.60 per 1M output tokens, with cached input at $0.075, on OpenAI's pricing page as of November 2026. Batch pricing for it is $0.075 for input and $0.30 for output. It has the lowest input and output prices of the three chat models in the table. Price says nothing about answer quality, so try a few of your own questions on each model before you settle.
Claude prices, and where Claude gets embeddings
As of November 2026, Anthropic's pricing documentation lists Claude Haiku 5.5 at $0.10 / MTok for input and $0.50 / MTok for output, for prompts up to 100,000 tokens, and at $0.50 and $2.50 for longer prompts. Claude Sonnet 5.5 is listed at $2 and $10, and Claude Opus 5 at $5 and $25. The page defines the unit as "MTok: Million tokens."
Anthropic's embeddings page says "Anthropic does not offer its own embedding model" and names Voyage AI as one provider. A notes chat that searches by meaning and answers with Claude therefore gets its embeddings from another company, such as OpenAI or Voyage AI.
What it costs to embed and query your notes with the GPT API
The cost to embed and query notes with the GPT API, the usual RAG setup for a personal knowledge base, splits into a one-time embedding bill and a bill for each question. The one-time part turns every note into a vector. The per-question part embeds your question, finds the closest passages, and sends them with your question to a chat model that writes the answer.
The one-time cost of embedding your library
Take a worked example with assumed sizes. Your library holds 2,000 notes of about 500 words each, which is 1 million words, and we assume about 1.3 tokens per word. That makes about 1.3 million tokens. At $0.02 per 1M tokens for text-embedding-3-small, as of November 2026, embedding all of it costs 1.3 × $0.02 = $0.026, or about 3 cents. OpenAI's page-based figure lands in the same place: 1.3 million tokens is about 1,625 pages of 800 tokens, and at about 62,500 pages per dollar that is $0.026.
Size changes little. A library five times larger, 10,000 notes of the same length, comes to about 6.5 million tokens and $0.13. On text-embedding-3-large at $0.13 per 1M tokens, the 1-million-word library costs about $0.17.
As general advice, three things push that number up:
- Overlapping chunks. Many RAG setups split notes into chunks that overlap, so some text is embedded twice.
- Edits. A changed note is embedded again. Editing 100 notes of 500 words in a month adds about 65,000 tokens, or about $0.0013.
- A new embedding model. Vectors from different embedding models can't be compared with each other, so a switch means embedding the whole library again.
The cost of one question
Your question is embedded too. A 30-word question is about 40 tokens, which costs less than a thousandth of a cent at $0.02 per 1M tokens. The answer is where the money goes.
For the worked example, assume each question sends about 2,000 input tokens (six passages of about 200 words, the instructions and your question) and gets back an answer of about 400 tokens. Both sizes are assumptions. The prices are OpenAI's Standard prices and Anthropic's base prices as of November 2026.
| Model | Input per 1M tokens | Output per 1M tokens | 2,000 tokens in | 400 tokens out | One question |
|---|---|---|---|---|---|
| GPT-5 mini | $0.25 | $2.00 | $0.0005 | $0.0008 | $0.0013 |
| GPT-4.1 mini | $0.40 | $1.60 | $0.0008 | $0.00064 | $0.00144 |
| GPT-4o mini | $0.15 | $0.60 | $0.0003 | $0.00024 | $0.00054 |
| Claude Haiku 5.5 | $0.10 | $0.50 | $0.0002 | $0.0002 | $0.0004 |
| Claude Sonnet 5.5 | $2 | $10 | $0.004 | $0.004 | $0.008 |
| Claude Opus 5 | $5 | $25 | $0.01 | $0.01 | $0.02 |
GPT-5 mini comes to $0.0013 a question, about a tenth of a cent. Token counts for the same text also differ between tokenizers. Anthropic's pricing page says Claude 4.7 and later models use a newer tokenizer that produces approximately 30% more tokens for the same text than the one before it. The table uses one token count for every model to keep the comparison simple, so check your real counts on your AI account's usage page after a few days.
If you let OpenAI host the search for you, its pricing page lists a file search tool at $2.50 per 1,000 tool calls, with storage at $0.10 per GB per day and 1 GB free, as of November 2026. At $0.0025 a call, the search alone costs about twice what the whole GPT-5 mini question costs in this example.
Three example months: light, regular and heavy
The question counts below are assumptions: 30 questions for a light month (about one a day), 150 for a regular month (about five a day) and 600 for a heavy month (about 20 a day). Each cell multiplies the cost of one question from the table above, rounded to the cent. Add about 3 cents once if your app embeds a 1-million-word library with text-embedding-3-small.
| Model | Light: 30 questions | Regular: 150 questions | Heavy: 600 questions |
|---|---|---|---|
| GPT-5 mini | $0.04 | $0.20 | $0.78 |
| GPT-4.1 mini | $0.04 | $0.22 | $0.86 |
| GPT-4o mini | $0.02 | $0.08 | $0.32 |
| Claude Haiku 5.5 | $0.01 | $0.06 | $0.24 |
| Claude Sonnet 5.5 | $0.24 | $1.20 | $4.80 |
| Claude Opus 5 | $0.60 | $3.00 | $12.00 |
Three made-up examples show how this plays out:
- Light: Lena, a freelance UX researcher, asks her interview notes about one question a day. On GPT-5 mini, her month costs about 4 cents.
- Regular: Daniel, a product manager, checks decisions and customer calls about five times a day. On Claude Sonnet 5.5, his month costs about $1.20.
- Heavy: Mei, a PhD student, questions her literature notes about 20 times a day while she writes. On Claude Opus 5 her month costs about $12, and on GPT-5 mini about 78 cents.
These figures cover single questions. An assistant that works in several steps costs more per request, because each step is another exchange with the model. Barside's assistant, for example, can take up to 40 steps on one request, and a request also carries the list of Barside's tools.
How those totals compare with a subscription
As of November 2026, Claude's pricing page lists Pro at $20 if billed monthly, or $17 a month with the annual subscription discount ($200 billed up front), and Max at $100 a month. It notes that the prices shown don't include applicable tax. Notion's pricing page lists Business at $20 per member a month, with prices in USD and a choice between paying monthly and yearly, where yearly saves up to 20%, though the page text we could read did not say which of the two the $20 is. Business includes Notion Agent, AI Meeting Notes and Enterprise Search, while Free and Plus include a trial of Notion AI.
We could not confirm ChatGPT's prices. The ChatGPT pricing page we could open lists the Free, Go, Plus, Pro, Business and Enterprise plans and says paid plans are priced per user per month, but the text we could read showed no amounts, and the pricing page on openai.com and OpenAI's help center refused our page reader.
To find your break-even point, divide the plan's monthly price by what one question costs you. Against a $20 month, the worked example gives these counts:
| Model | One question in the example | Questions a month to reach $20 |
|---|---|---|
| GPT-5 mini | $0.0013 | about 15,400 |
| GPT-4.1 mini | $0.00144 | about 13,900 |
| GPT-4o mini | $0.00054 | about 37,000 |
| Claude Haiku 5.5 | $0.0004 | 50,000 |
| Claude Sonnet 5.5 | $0.008 | 2,500 |
| Claude Opus 5 | $0.02 | 1,000 |
On cost alone, the key wins in every row of the example. A plan still makes sense if you use its chat app all day for writing and research, if a team shares one Notion workspace with AI meeting notes and agents, or if you want a bill that can't move. For web research, where searches add a second meter, AI research assistant pricing works through the same break-even math.
How Barside handles the cost of chatting with your notes
Barside is a free desktop app for Windows, Linux and Mac, and it brings no AI of its own. Its AI search, in Library & briefing, uses no embeddings, so the one-time line in the example above is zero. The search runs on your computer and matches the words of your question, with no model, no download and no index to build, and it reads your current library each time. It keeps the six passages that match most closely, with at most two from any one note or bookmark, and sends only those and your question to the AI you connected. AI search that cites its sources explains every step.
The trade-off is that search finds words as written. A note that only says "bill" won't come up when you ask about an "invoice", and words of one or two letters, such as "Q3", are skipped, so ask with the names and terms your notes use. If you're weighing search by meaning on a big vault, why semantic search gets slow in a large vault explains what embeddings and an index cost on your own computer.
To chat with your notes on your own key:
- Download Barside and open it.
- Open Settings → AI and choose ChatGPT. Barside opens that company's key page in your browser. Make a new key and copy it, and Barside picks it up from the clipboard, checks it and saves it encrypted. Claude and Gemini work the same way, and OpenRouter is a one-click sign-in.
- Pick a model with Choose a model. A ChatGPT connection starts with the newest balanced GPT model on your account and a Claude connection with Claude Opus 5, so choose a smaller model if you want each question to stay cheap.
- Under AI spending limit, turn on Limit what Barside's AI spends.
- Open Library & briefing, stay on the Ask library tab, type your question in Ask your entire library (up to 2,000 characters) and press Ask library.
- Read the answer under From your library, where each source is numbered, and click a source's title to open the note.
The AI company bills your own account for what you use, and Barside charges nothing for AI. Barside Plus is optional at US$20 a month or US$200 a year plus sales tax, for encrypted cloud backup, sync and cloud agents, with no AI credits inside it.
Cap the bill, or use a plan you already pay for
The spending cap sits in Settings → AI under AI spending limit. Turn on Limit what Barside's AI spends and pick an amount from US$0.50 to US$10,000 for a day, a week (from Monday) or a month.
- Barside estimates each answer from the AI company's list prices for your model. With OpenRouter, it uses the exact cost OpenRouter reports.
- At 80%, you get a notification and a banner in the assistant and Library.
- At the limit, the assistant, agents, Ask library, Quick AI and the Space helper stop sending requests until you raise the limit, turn it off or the count starts again.
- Each computer counts its own use, and cloud agents aren't counted, so set a limit on your AI company account as well.
If you already pay for a Claude or ChatGPT plan, Barside 1.59 and later offer My plan on the Claude and ChatGPT cards in Settings → AI. Barside then runs Claude Code or Codex on your computer, signed in with your plan, so the plan's limits apply instead of a bill per use. Web search inside Barside and cloud agents still need an API key. Anthropic and OpenAI have not said that apps may use a plan this way, so an API key is the safer choice.
What leaves your computer when you ask
Your library stays in Barside's data folder on your computer. When you ask a question, your question and up to six matching passages go to the AI company you connected, and that company's data policy applies. Barside does not train any model on your notes. Your key is saved encrypted on your computer, and for work on your computer, requests go straight to the AI company with no Barside server in between. Cloud agents in Barside Plus are the exception, since they send your key to Barside's server, sealed, to run your scheduled agents.
To keep a note out of answers, open the Sources tab in Library & briefing and untick it, or a whole workspace, under Choose included workspaces & notes. Do it before you ask, because unticking later can't recall an excerpt that was already sent. Read how AI search uses your data and the AI you connect for the details, and AI second brain apps that answer only from your notes for how other apps handle the same question.
Works well with
These guides go further into the search, the keys and the prices behind this page.
FAQ
Questions and answers
Is my own OpenAI API key cheaper than an AI notes subscription?
For chatting with your notes, usually yes. In our worked example, with assumed sizes, a question that sends six passages to GPT-5 mini costs about $0.0013 at OpenAI's prices as of November 2026, so a $20 monthly plan only matches it at about 15,400 questions a month. A plan still wins on a fixed bill, a ready-made chat app and use beyond your notes.
How much does it cost to embed my notes with text-embedding-3-small?
OpenAI's pricing page lists text-embedding-3-small at $0.02 per 1M tokens as of November 2026. In our worked example, 2,000 notes of about 500 words, with an assumed 1.3 tokens per word, come to about 1.3 million tokens, so embedding them once costs about $0.026. Edits, and a change of embedding model, mean embedding some or all of the library again.
What is the GPT-4o mini API price for input and output tokens?
OpenAI's API pricing page lists GPT-4o mini at $0.15 per 1M input tokens, $0.075 per 1M cached input tokens and $0.60 per 1M output tokens on Standard pricing, as of November 2026. Its Batch prices are $0.075 for input and $0.30 for output.
How much does GPT-5 mini cost on the OpenAI API?
As of November 2026, OpenAI's pricing page lists GPT-5 mini at $0.25 per 1M input tokens, $0.025 per 1M cached input tokens and $2.00 per 1M output tokens on Standard pricing, and at $0.125 for input and $1.00 for output on Batch. GPT-4.1 mini is listed at $0.40 for input and $1.60 for output on Standard.
Does Barside use embeddings to search my notes?
No. Barside's AI search matches the words of your question on your computer, with no model, no download and no index to build, so there is no embedding bill. It sends your question and up to six matching passages to the AI you connect. The trade-off is that it finds words as written, so ask with the words your notes use. See AI search that cites its sources.
Can I use a ChatGPT or Claude plan instead of an API key in Barside?
Yes, from Barside 1.59. Choose My plan on the ChatGPT or Claude card in Settings → AI, and Barside runs Codex or Claude Code on your computer, signed in with your plan, so its limits apply instead of a bill per use. Web search inside Barside and cloud agents still need an API key. Anthropic and OpenAI have not said that apps may use a plan this way, so an API key is the safer choice.
How do I keep API costs from running away?
Set two limits. In Barside, open Settings → AI and, under AI spending limit, turn on Limit what Barside's AI spends, with an amount from US$0.50 to US$10,000 for a day, a week or a month. Barside warns you at 80% and stops its AI at the limit. Set a limit on your AI company account too, because each computer counts only its own use and cloud agents are not counted.
Is Barside free?
Yes. Barside is free on Windows, Linux and Mac, with no note limit and no account for your notes. It has no AI of its own, so the AI company you connect bills your account for what you use. Barside Plus is optional at US$20 a month or US$200 a year plus sales tax, for encrypted cloud backup, sync and cloud agents.
Sources
Where the facts come from
Checked on November 13, 2026. Products change: if something here is out of date, tell us at support@barside.space.