Choosing and Using the Tools
A comparison written by someone selling none of them.
- Level
- Nothing assumed
- Lessons
- 76
- Reading time
- 607 min
- Price
- Free, no sign-up to read
A straight, comparative look at the AI tools people actually use — ChatGPT, Claude, Gemini, Copilot, and the open models you can download. What genuinely separates them, what each free tier really gives you, when paying is worth it, what happens to what you type, and how to re-check all of it in six months when the answers have moved.
Opens after the AI at Work exam
Sign in, finish that course, and pass its exam. You can read this syllabus meanwhile.
Go to AI at WorkModule 1
What a tool is made of
Before any comparison, you need to know what you are looking at. This block takes a product apart: the model inside it and how to read its name, the size class that decides the price, the thinking mode that decides the wait, the context window that is a ceiling rather than a promise, the unseen instructions that shape its manners, and the tools it can reach for. By the end, when a tool behaves oddly, you will know which of those to blame.
By the end you can
Take any AI product you are handed, name the exact model and size class inside it, say whether its thinking mode, context handling, system prompt or tool layer is producing the behaviour you are seeing, and predict which of those will change without notice
- 1What you are actually choosingYou are choosing three things at once: a model, the product wrapped around it, and the route you reach it by.
- 2Reading a model's nameA model name encodes family, version, size class and build date, and the size class and the build date are the two parts that quietly decide what you actually get.
- 3Small, medium, large: what the size label buysStart every task on the smallest model in the family and move up only when the failures are wrong facts or broken reasoning, because a fifth of failures are prompt problems and the price gap between sizes is around twentyfold.
- 4Thinking modes, and what you pay for themThinking modes buy accuracy on problems with a checkable internal structure and buy nothing on recall, style or extraction, while charging you for every hidden token and making you wait.
- 5The context window is a ceiling, not a promiseThe advertised context window promises only that your request will be accepted, so test where a fact placed two-thirds of the way through a long document actually gets found, and put your question after the document rather than before it.
- 6The instructions you never seeMost of what a chat product feels like — its refusals, its formatting, its persona, even whether it knows the date — comes from an unseen block of company instructions rather than from the model.
- 7What 'it can browse the web' actually meansTool use is one loop — the model chooses to call a tool, the result is pasted into its context as text, and it continues — so most tool-assisted errors are bad tool output rather than bad reasoning.
- 8'It understands images' — what that buysA vision model receives a downsampled, tokenised version of your image, so exact characters and counts should be extracted with OCR or a text layer first and only then handed to the model.
- 9When the engine gets swappedTools change underneath you through moving aliases, retuned routers and rewritten system prompts, so keep ten fixed prompts with known answers and a dated note of the model string, and pin versions wherever the work runs unattended.
Module 2
Testing a tool yourself
Every published comparison measures fit to a benchmark. This block builds the one that measures fit to your work: twelve real prompts in a plain text file, two of them with answers you already know, at least one in the language you actually use, scored blind against criteria written before you saw a single answer. It also covers what leaderboards can and cannot tell you, and how to record a decision so you stop making it again.
By the end you can
Build a personal benchmark of a dozen real tasks with written success criteria, run a blind scored comparison across three candidates while accounting for run-to-run variation, and record a decision that names the exact model versions and the conditions that would reverse it
- 10Where they differ, and where it is marketingYour five real tasks tell you more than any leaderboard, because a benchmark measures the test and not your work.
- 11The file you will keep for yearsTwelve real prompts in a plain text file, each with a written note of what a good answer contains, decided before you see any answers, is a better guide to tool choice than any published comparison.
- 12The prompts where you already know the answerInclude prompts whose answers you already know, and score not just correctness but whether the tool hedged when uncertain and whether it abandoned a correct answer when you sounded doubtful.
- 13Testing the language you actually work inLanguage quality varies between models far more than English benchmarks suggest, and non-Latin scripts also cost several times more tokens for the same text, so test drafting, register, code-switching and numbers in the language you actually use.
- 14Taking the brand off the answersScore answers labelled A, B and C against criteria written in advance and only then reveal which tool produced which, because brand halo, order and length shift judgements more than the tools differ.
- 15Run it five timesBecause products sample from a probability distribution rather than picking the single most likely reply, a single run is one draw and any tool difference smaller than a tool's own run-to-run variation is not a difference.
- 16A rubric you can apply in ten secondsScore each answer 0, 1 or 2 against criteria written before you saw any answers, and watch for the four biases — length, formatting, confidence and agreement — that inflate scores without improving usefulness.
- 17Reading a leaderboard without being fooledStatic benchmarks suffer contamination, saturation and best-case settings, arena votes reward length and confidence rather than correctness, and when a leaderboard disagrees with your own prompts, your prompts win.
- 18Testing the edges: refusals and house mannersFor most professionals the costly failure is a false refusal on ordinary work, and because refusals live mostly in the product's instruction layer, they vary between products running the same model.
- 19Making the decision, and recording itWrite down the choice, the exact model strings, the date, and the specific conditions that would change your mind, because a decision you have not recorded gets silently re-litigated every time somebody publishes a benchmark.
Module 3
The routes in
The same model reaches you through very different doors, and the door decides more than people expect. This block covers the web, phone and desktop apps and how they differ; the API, its request format and the spend cap you set before your first call; free front-ends that turn a key into a chat app you own; routers; where a coding assistant lives; the extension and wrapper trap; and the two routes most comparisons ignore entirely — a phone with 3 GB of RAM and a metered bundle, and a screen reader.
By the end you can
Choose the right route to a model for a given job — web, phone, desktop, API, free front-end, router, editor or terminal — set one up with a hard spend cap and a revocable key, and judge an extension or wrapper by its permissions, publisher and disclosed data use before installing it
- 20The web app, the phone app and the desktop appThe website, the phone app and the desktop app are different products sharing one account, with the web usually ahead on features and app-store subscriptions usually costing more than the same plan bought on the website.
- 21Your first API callAn API request is a model name, a list of role-tagged messages and a few settings, and because most providers accept the same format, moving between vendors or to a local model is usually a change of two lines.
- 22Keys, spend caps, and the bill nobody wantsSet a hard monthly spend cap before your first API request, keep keys out of any file you might share, use one key per use so a leak revokes one thing, and never put a key anywhere a user's device can read it.
- 23Building your own chat app for nothingA free open-source front-end turns an API key into a chat application you control — your system prompt, your stored history, several providers and a local model in one window — at the cost of polish, built-in extras and support.
- 24Routers: one key, many modelsA router gives one key and one bill across hundreds of models, at the cost of a margin, an extra party in the data path, and — for open models — a back-end host chosen per request that may differ in quantisation and context length.
- 25Where a coding assistant livesCoding assistants differ mainly by what they can see and do — a file, a project, your terminal, or a cloud clone — so keep every change small, keep version control between you and any agent, and review each line before it lands.
- 26Extensions, wrappers and the permission dialogAn AI browser extension usually requests permission to read and change every page you visit, so check publisher, age, recent reviews and the data disclosure before installing, and never hand a third-party app a key you cannot cheaply revoke.
- 27When the phone is the computerOn a phone, install web apps to the home screen rather than native apps, treat text as effectively free but photographs and voice as expensive, prefer fast models on poor connections, and use a small local model when there is no connection at all.
- 28Voice, screen readers and hands-free useRoute decides usability before model quality does: dictation and read-aloud are free on every device, streaming replies interact badly with screen readers, and a plain-text API client is accessible by construction when a product's interface is not.
Module 4
What it actually costs
Money decides more tool choices than quality does, and almost nobody does the arithmetic. This block does it: what a token is and why output costs several times input, how to cost a job from a twenty-item pilot before running fifty thousand, the caching and batching discounts most people never switch on, how to route easy work to cheap models, the three billing shapes and their different risks, the costs that never reach the invoice, and what to do when the card does not work or the price is simply out of reach.
By the end you can
Estimate the cost of a job from a measured sample before running it, cut that cost with prefix caching, batching and model routing without lowering quality, choose between a subscription, prepaid credit and pay-as-you-go for your own pattern of use, and set an enforced cap so the worst case is bounded
- 29The free tiers, and their trapsA free tier shows you a product's floor, not its ceiling, and it will downgrade you without saying so.
- 30When paying is worth itPay when a cap is blocking work you would have done anyway, not to own the best model.
- 31The unit you are billed inYou are billed per token, output usually costs three to five times input, hidden thinking counts as output, and non-Latin scripts can take several times more tokens for identical content.
- 32Costing a job before you run itRun twenty real items, record actual token usage, multiply and add a third for retries — and run the same test on a small and a large model so the price of the extra accuracy is a number rather than an assumption.
- 33The two discounts almost nobody usesPut every stable block at the front of a request so prompt caching can match the prefix, send anything that can wait through the batch route at around half price, and cut what you send and what you ask for before you consider a cheaper model.
- 34Spending money where it mattersA cascade — cheap model everywhere, expensive model only on flagged cases — commonly cuts cost by four fifths at equal or better quality, provided you have measured on a hundred items which cases the cheap model actually gets wrong.
- 35Subscriptions, credits and pay-as-you-goSubscriptions cannot surprise you and are easy to underuse, prepaid credits fail safely by running out, and pay-as-you-go is the most efficient and the only shape that can produce a bill you did not expect.
- 36The costs that are not on the invoiceThe invoice is the small number: verification time, rework, learning and automation bias decide whether a tool pays, and the returns are largest exactly where you can check an answer faster than you could produce it.
- 37Paying for it from where you liveThe sticker price is not what you pay — card fees, digital-services tax and local levies can add a quarter — and where a card or the service itself is unavailable, free tiers, prepaid API credit, hosted open models and a local model cover almost everything.
- 38A budget that actually holdsWrite down current spend, set an enforced hard cap with alerts on every pay-as-you-go account, and review every subscription quarterly against what you actually used rather than what you might use.
Module 5
Open models and your own hardware
Downloadable models are the route that costs nothing per use, keeps working when the internet does not, and cannot be deprecated. This block treats them as a purchasing decision rather than a hobby: the memory arithmetic including the cache people forget, reading a model card and its licence before spending the download, choosing a quantisation, renting an open model instead of running it, the honest cases where local is the wrong answer, and the two-model arrangement most people should end up with.
By the end you can
Work out what your own machine will run including context and cache, read a model card and licence well enough to decide whether to build on it, choose a quantisation and format deliberately, compare hosted open-model providers on speed, precision, context and data terms, and set a routing rule between local and hosted work you can apply without thinking
- 39The open modelsOpen weights means you can download and keep it; it rarely means open training data, and the licences differ.
- 40Running a model on your own machineA local model buys privacy, offline access and zero marginal cost, and charges you capability and speed for it.
- 41Working out what your machine will runBudget memory for the weights, the KV cache that grows with your context length, and the operating system — and when a local model is too slow, cut context and cache precision before you cut model size.
- 42Reading a model card before you downloadA model card tells you the licence, the format, the context length and the languages before you spend the download — and it almost never tells you what the model was trained on, which is the question you cannot answer for a customer later.
- 43Choosing a quantisationTake Q4_K_M unless you have memory to spare, prefer a smaller model at 4-bit over a larger one at 2-bit, and watch for drifting instruction-following as the sign that you have quantised too far.
- 44Renting an open model instead of running itHosted open models sit between a laptop and a frontier subscription, and the same model name from two providers can differ in speed, quantisation, context length and data terms — so check those before comparing prices.
- 45The licence is a business decisionRead the licence rather than the word used to describe it: Apache and MIT remove a whole class of question, community licences permit commercial use with conditions, and whether any of this counts as open source is a live, unresolved dispute.
- 46When running it yourself is the wrong answerLocal models are the right answer for confidential, offline, high-volume and permanent work, and the wrong one for hard reasoning, whole-product needs, battery life, other people depending on it, and buying hardware to avoid a subscription.
- 47A working setup with two modelsRoute work by what the material contains rather than by how hard it looks — confidential to the local model, routine and bulk to a cheap hosted one, the hard remainder to the strong one — and write the rule down once so the choice stops being a decision.
Module 6
Data, terms and the law
What happens to what you type is the question most likely to matter and least likely to be answered on a comparison page. This block reads a privacy policy in ten minutes, says what a business agreement actually buys and what no agreement can give you, sets out the shared shape of data protection law without pretending to give legal advice, separates the three tangled questions about who owns the output, and covers redaction, secrets, shared devices and the workplace policy your employer probably has not written.
By the end you can
Establish for any tool whether your content is trained on, retained or read by people and whether that covers your tier, redact material so the sensitive part is never sent, recognise which questions about ownership and legality are genuinely unresolved, and write the one-page record of tools and boundaries that a client or regulator would ask for
- 48Who reads what you typeNot training on your data and not storing it are different promises, and most providers only make the first.
- 49Reading the terms in ten minutesSearch a privacy policy for train, retention, human review, third part, delete and transfer, read the sentence around each, check whether it covers your tier, and screenshot it with the date — and treat an unfindable answer as the answer.
- 50What a business agreement actually buysA business tier buys a contract rather than a better model — a training commitment, a data processing agreement, a sub-processor list, retention controls and regional processing — and none of it makes the output correct or moves the responsibility off you.
- 51Your country, your rulesMost data protection regimes share the same shape — you are the one deciding, you need a lawful basis and a purpose, people have rights, and sending data abroad is restricted — and your profession's rules sit on top of your country's.
- 52Who owns what it producesProvider terms usually assign you their rights in the output, but whether that output is copyrightable at all and whether it infringes anyone else's rights are separate, unsettled questions answered differently in different countries.
- 53Redaction that actually worksReplacing identities with roles preserves everything the model needs and removes most of the risk, but files carry authorship metadata, hidden rows, extractable text under black rectangles and location data, and removing a name column is not anonymisation.
- 54Secrets, and what not to pasteNever paste credentials or identity numbers at any tier, substitute placeholder values before copying config files and logs, and if it happens anyway rotate the secret first — deletion limits exposure but does not undo storage.
- 55Accounts, sharing and devicesYour conversations are visible to whoever holds the device, to an administrator on a work account, through browser sync and through any share link — so keep work and personal accounts apart, audit memory and connected apps, and treat a shared link as published.
- 56What your workplace probably has not decidedCheck the confidentiality policy, your contract and your client agreements before asking about AI specifically, ask a specific bounded question rather than an open one, and keep a one-page record of which tools you use, for what, and what you never send.
Module 7
Beyond the chat box
Most of the money in this field is spent outside the chat window, on tools nobody compares carefully. This block works through the categories: image generation, transcription and translation, voice, video, research and search, documents and OCR, spreadsheets and data, meeting notetakers, the per-seat office suite assistants, automation platforms, and the specialist products sold to your profession. Each one gets the same treatment — what genuinely separates the products, where the free path is good enough, and the specific way each category fails.
By the end you can
Judge a tool in any of the main non-chat categories on the property that actually separates them — legible text, verifiable citations, licensed data, visible code, consent and retention — name the free path that does the same job, and identify the failure mode specific to that category before it costs you
- 57Choosing an image toolImage tools differ most in rendering text, keeping a subject consistent, and editing an existing picture rather than in raw quality — and the free local route removes both the per-image cost and the content restrictions, along with the indemnity.
- 58Transcription and translationTranscription is available free and locally at commercial quality, so accuracy is decided by microphone placement, overlapping speech and vocabulary hints rather than by which service you pay — and both transcription and translation fail silently.
- 59Voice and speech toolsText-to-speech is free, local and good enough for accessibility, proofreading and reaching people who do not read the language — while voice cloning carries a real fraud risk that makes explicit consent and a household callback rule the load-bearing parts.
- 60Video tools, generated and editedNearly all the practical value in AI video is in editing, where free tools like DaVinci Resolve are professional-grade, while generation is priced per second with a high discard rate and works best as short ingredients cut into real footage.
- 61Research and search toolsSearch-grounded answers are bounded by what the search returned, so click two citations to check they support the claim and read the source list before the report — and where you can supply the sources yourself, the failure mode becomes a safe one.
- 62Documents, PDFs and OCRCheck whether a PDF already has a text layer before anything else, because exact extraction beats recognition and both beat asking a vision model to read characters — and validate extracted tables by checking whether the rows still add to the printed total.
- 63Spreadsheets and data toolsData work is the category where checking is cheapest, so choose tools that show you the code, and make printing row counts before and after every operation the habit that catches silently dropped records.
- 64Meeting notetakersNotetakers are adopted faster than the rules that govern them: consent for recording varies by jurisdiction, retention and who can search the archive are the questions to settle first, and summaries turn inconclusive discussions into decisions.
- 65Office suite assistantsAn office suite assistant sells access to your own mail and files inside the organisation's boundary rather than a better model, so pilot five seats across different roles and read the distribution — and tidy file permissions before deploying, not after.
- 66Agents and automation platformsMost real value is in fixed workflows with a model in one step rather than in self-directing loops, because per-step reliability compounds badly — and any agent that reads untrusted text while holding permission to act is an unsolved security problem.
- 67Specialist and industry toolsA specialist tool is worth its premium when it has licensed data you cannot get, cites into verifiable sources, integrates with your working system, or carries accountability — and the professional stays responsible for the advice regardless.
Module 8
Living with it
The choice is the beginning. This block covers the years after it: turning repeated chatting into saved instructions you own rather than rent, keeping your work portable enough to act on your own review, what to do in the first hour after a deprecation notice, introducing a tool to colleagues without a mandate, measuring whether it actually helped, keeping up on an hour a month, when the right answer is to decline — and which parts of everything you have just read will still be true in three years.
By the end you can
Turn repeated work into saved instructions kept in your own files, migrate to a replacement tool when one is withdrawn without losing your working setup, measure a tool's effect on your own work honestly enough to act on it, and separate the parts of any AI comparison that expire within months from the structural facts that do not
- 68From chatting to a workflowThe people who get most from these tools have stopped starting from nothing: four or five saved instructions containing context, format, standards and one example they liked, kept in their own plain text file rather than only inside a product's project feature.
- 69Keeping your work portableKeep instructions, reference material and findings in your own plain text files, and test a product's export function in the first week rather than the week you decide to leave.
- 70Two tools, and a date in your calendarKeep a default and a checker, and re-run your own five prompts twice a year.
- 71When your tool is withdrawnExport first because the data deadline is often earlier than the shutdown, treat the recommended successor as a suggestion to test with your own prompts, and expect migration regressions to come from expectations the old model inferred and the new one does not.
- 72Introducing a tool to other peopleStart from one colleague's tedious task rather than from the tool, share the saved instruction rather than the enthusiasm, take specific objections as information, and measure a distribution across roles rather than an average.
- 73Measuring whether it actually helpedRecord minutes saved, minutes spent checking and one thing it could not do, on the day, per task — and never measure messages, words or lines produced, because those rise without anything being completed to a standard.
- 74Keeping up without drowningOne hour a month on changelogs, one substantial piece and one thing tried is enough, because your decision record's change-my-mind conditions filter almost everything else — and only price changes, data-term changes, deprecations and capability arriving at a smaller size deserve interrupting it.
- 75When the right answer is not to use itDeclining is sometimes the right decision — where the point is that a person did it, where the struggle is the learning, where nobody can check the answer, and where somebody is owed an accountable human judgement — but an untested objection is a preference rather than a principle.
- 76What changes, and what does notPrices, model names and capability claims expire within months, while the structural facts — fluency is not reliability, the product layer dominates experience, verification cost decides value, and your own prompts beat any comparison — have held for years.