Real-life example: a very cheap setup with DeepSeek

What this example builds

One provider profile in SideNote Pro, pointed at DeepSeek, running deepseek-v4-flash-vision-exp, paid for with a five dollar top-up. That profile becomes your everyday assistant: the one that answers when you highlight a sentence and press Ctrl+Shift+E, the one that rewrites an email, and the one that reads a screenshot of an error dialog and tells you what it means.

It is a bring-your-own-key setup like any other, so nothing here is special-cased in the application. If you would rather read the general version first, see what BYOK means. If you want the mechanics of any compatible endpoint, see configuring an OpenAI-compatible provider.

Why we recommend the vision model here

There is usually a plain text variant of a fast model and a vision variant, and the plain one looks like the obvious budget choice. We recommend the vision model anyway, and the reason is that image input is a model capability rather than an application one.

SideNote Pro can capture a region of your screen, take a bitmap off the clipboard with Ctrl+V, or accept an image dragged onto the window, and it will format all of them correctly for any OpenAI-compatible endpoint. Whether anything comes back depends entirely on the model you pointed it at. A text-only model ignores the image or refuses it, and the failure is quiet enough that people spend an afternoon blaming the app. See asking AI about a screenshot.

So deepseek-v4-flash-vision-exp is not a luxury upgrade in this example. It is the difference between an assistant you can show things to and one you have to describe things to, and describing a broken layout, a spreadsheet region, a chart or a stack trace in a dialog box is slow and lossy work.

An error dialog you have never seen before

Capture it, ask what it is actually telling you and what the likely causes are.

A screenshot of a settings pane in some other application

Ask where the option you are hunting for lives.

A chart in a report

Ask what the trend is and what it does not show.

A photo or scan of a printed page

There is no OCR in SideNote Pro, so whatever text gets read is read by the model's own vision. A text-only model reads nothing at all.

A region of a spreadsheet

Faster to capture than to retype, and the numbers come along with their layout.

What five dollars actually buys

This is the part people do not believe until they try it. For ordinary desktop assistant use, a five dollar top-up on a cheap fast model lasts months. Not a week, not a fortnight. Months.

The arithmetic behind that is simple. Budget models in this class are priced in cents per million tokens, and the requests you actually send all day are small: a highlighted paragraph and an instruction, a short email to rewrite, a two line question about a screenshot. A request like that is hundreds of tokens, not hundreds of thousands. You have to work quite hard to spend a dollar a month this way.

So the honest version of the claim is this: it really does go a very long way, and how long depends on how you use it. Someone asking twenty short questions a day will not come close to finishing five dollars in a quarter. Someone attaching a hundred page PDF every morning and never starting a new thread will get through it much faster. The habits in reducing your AI API costs are what decide which of those two you are.

Fast enough to use without thinking about it

The second reason to pick a flash-class model for the everyday slot is latency. A docked assistant only earns its place if answering a question is quicker than opening a browser tab, and that means the answer should start appearing more or less as you finish reading your own question.

Small fast models are very good at exactly the work that dominates daily use: explaining a sentence, rewriting a paragraph, summarizing a page, translating ordinary prose, generating a few subject line options, telling you what a screenshot shows. None of that has hidden structure to reason about, so a heavyweight model would give you the same answer, later, for more money. See when to use high reasoning effort for where the opposite is true.

Leave reasoning effort on Provider default for this profile. reasoning_effort is an optional field that not every model implements, and on quick fluency tasks it buys nothing anyway.

Step one: the DeepSeek account and the API key

This happens on the provider's side, in a browser, and takes a few minutes. Go to platform.deepseek.com. Note that this is the developer platform, which is a separate thing from any consumer chat product: a chat subscription does not include API credit.

1. Create an account and sign in to the platformUse an email you will still have access to in a year, because it is the account that owns the billing.
2. Top up a small amountFive dollars is the whole point of this example. Start there. You can always add more, and starting small means an accident costs five dollars rather than fifty.
3. Set a spending limit if the dashboard offers oneDo it now rather than later. A cap on the provider side is the only limit that is authoritative, because it is the only one that sees every application using that key.
4. Create an API key and copy it immediatelyProviders typically show a secret key once and never again. Copy it straight into SideNote Pro, or into your password manager on the way.
5. Find the exact model ID in the model listIDs are case-sensitive and occasionally carry a version suffix. Copy the string the provider documents rather than typing what you remember.

Step two: the profile in SideNote Pro

Now the easy half. Open Settings, then AI providers, and add a profile. DeepSeek speaks the OpenAI-compatible Chat Completions API, which is the contract SideNote Pro is built against, so there is nothing to adapt and no plugin to install. You are filling in four fields.

Settings › AI providers

Name the profile for what it is for rather than for the model inside it. Models change and prices fall; a profile called Everyday survives swapping the model underneath it, and a profile called DeepSeek V4 Flash Vision becomes a small lie the first time you edit it.

Settings › Setup guide

Step three: test before you trust it

Press Test connection. It calls GET https://api.deepseek.com/v1/models with your key as a bearer token. A pass tells you three things at once: the base URL is right, the key is valid, and the discovered model IDs can now populate the searchable model picker, so you can select the vision model from a list instead of trusting your typing.

Then send something real. Highlight a sentence anywhere in Windows, press Ctrl+Shift+E to explain it, and watch the answer stream in. Capture a region of your screen and ask what it shows. Those two checks confirm the parts of the setup that matter: text in, images in, streamed text back.

Step four: record the prices, so the ledger is worth reading

SideNote Pro tracks token counts locally, but it cannot know what your provider charges unless you tell it. Enter DeepSeek's current input and output prices in USD per million tokens for this model, and the usage view turns from a token counter into a spending estimate you can act on.

Settings › AI providers › Model pricing

Settings › Usage and cost

Optionally, switch on the local daily estimate limit as a seatbelt. It is a local guard working from local estimates and it cannot see other applications using the same key, so it complements the cap in the provider dashboard rather than replacing it. See usage and cost tracking.

What the profile is good at, and what it is not

A cheap flash model is not a compromise for most of what a docked assistant does. It is the correct tool. But it is worth being clear about the edges, because the failure mode of a budget setup is trusting it on the one task it is bad at.

The practical shape of that is a second profile. Keep this DeepSeek profile as your default, and add a Deep profile on a stronger model with reasoning effort raised, for the handful of problems a week that genuinely need it. Duplicate copies the key and base URL, so building the second one is a two minute job, and switching between them is one click in the composer picker.

Settings › AI providers

Escalating beats starting deep. Ask the cheap fast profile first, because most problems that look hard are not, and you find that out in about three seconds. Switch profiles only when the answer is visibly shallow. Done that way you pay top-tier rates only on the problems that turned out to need them, which is the entire argument of setting up a model for each kind of task.

Five habits that stretch the five dollars

1. Start a new thread for a new topic. Ctrl+N. Every turn re-sends the accumulated history, so turn thirty of one endless thread is paying for twenty-nine turns of irrelevance.
2. Detach a document when you are finished with it. A large PDF is a large input on every later turn of that thread, not just the first one.
3. Ask for a length. In three bullets or under 100 words. Output tokens are the expensive ones, and a model told to be brief usually answers better because it has to decide what matters.
4. Edit and resend rather than arguing. Three corrective turns cost three round trips and drag all three into every request that follows. Fixing the original question costs one.
5. Crop the screenshot. Region capture rather than a full desktop grab. Fewer pixels is a smaller image payload, and a tight crop also gets you a better answer because there is less for the model to be distracted by.

None of these is a sacrifice. They all make the assistant better as well as cheaper, which is the usual sign that a cost habit is a good one.

Things worth knowing before you commit

A hosted provider is a hosted provider

Whatever you send goes to DeepSeek, under their terms and their jurisdiction, exactly as it would with any other hosted API. If the material is confidential, that is a decision to make deliberately.

For confidential work, add a local profile

Ollama or LM Studio costs nothing per token and never leaves the machine. Having it already configured is what makes the right choice the easy one at the moment you need it.

Compatibility is about the request format

OpenAI-compatible means streaming chat/completions works. It does not promise every optional field behaves identically, which is why reasoning effort stays on Provider default here.

Model IDs move, and this one says so

The -exp suffix marks an experimental model, and experimental models get renamed, superseded or retired on shorter notice than settled ones. If requests start failing on a setup that worked yesterday, re-run model discovery and check the current ID before assuming anything is broken.

Credit can simply run out

A profile that suddenly returns rate limit errors is often an empty balance rather than a technical fault. Check the dashboard first.

That list is not a warning against the setup. It is a five dollar, ten minute configuration that gives a Windows desktop a fast assistant which can also read what is on screen, and for most people it will cover the whole of daily use for a season. It is simply worth knowing which parts of it are yours to watch.