Everyday AI Tools Guide — Part 1: Choosing the Right Tool

Updated September 25, 2026. Every limit, cap and quoted line below was read that day from the vendor's own documentation or from the standards body that published it, and each page is listed at the end. Documented limits change without notice; check the page before you design a process around a number.

Everyday AI Tools Guide — Series Navigation

1. Part 1: Choosing the Right Tool (You are here)
2. Part 2: Building a Reliable Workflow
3. Part 3: Practical Prompts and Safe Automation

The only spec that decides what you can hand over at once

Almost every "best AI tool" list compares personalities. The specification that actually changes what you can do on a Tuesday afternoon is the context window: how much text the tool can hold in front of it in a single pass. Everything else - tone, speed, which brand your colleague prefers - is downstream of whether your document fits.

As their developer documentation stands today, all three of the large vendors list roughly a million tokens of input for their current top models. Anthropic's model overview lists 1M tokens of context with 128K tokens of output; OpenAI's model reference lists about 1.05M tokens of context with 128K output; Google's Gemini 3 documentation states that its models "support a 1 million token input context window and up to 64k tokens of output," and the per-model page gives the exact integers, 1,048,576 in and 65,536 out.

Now the part that the comparison articles leave out. Those are developer-platform figures. On the consumer side, Anthropic's own support page states 200K tokens across its models and paid consumer plans, with 500K reserved for one model on Enterprise. Google documents its 1M window as a benefit of the paid AI Pro and Ultra tiers rather than the free app. OpenAI publishes no consumer context window at all - three separate help pages covering plan limits were checked and none states a number.

So the honest version of "these tools have a million-token window" is: the API does, your subscription may not, and one vendor does not say. That single correction changes which of the tasks below are possible in one step and which have to be chopped up.

Three vendors, three rules for turning tokens into words

Tokens are useless as a unit until you convert them. Each vendor publishes its own rule of thumb, and they do not agree.

Vendor's documented rule	What 200,000 tokens becomes	What 1,000,000 becomes
Anthropic: a token is about 3.5 English characters	about 700,000 characters	about 3.5 million characters
OpenAI: 100 tokens is about 75 words	about 150,000 words	about 750,000 words
Google: 100 tokens is about 60 to 80 English words	120,000 to 160,000 words	600,000 to 800,000 words

The spread between the highest and lowest estimate of the same window is about 40,000 words - roughly a short novel. Nobody is wrong; the rules are measured on different languages and different text. The practical consequence is that you should size a job using the least generous rule you can find, because running out of window mid-document is the failure that quietly produces a confident summary of only part of your file.

A 1,000-page upload against a 200,000-token window

Here is where two published limits from the same vendor stop lining up. Anthropic's upload documentation accepts PDFs of up to 1,000 pages, noting that visual analysis applies only to files of 100 pages or fewer while pages 101 to 1,000 are processed as text. Its consumer support page puts the window at 200K tokens.

Take a dense business page at 500 words. A 1,000-page PDF is then about 500,000 words, which by OpenAI's published rule is roughly 667,000 tokens - about 3.3 times the documented consumer window. Assume a sparser 250 words per page and it is still about 333,000 tokens, or 1.7 times over. The upload will be accepted. The whole document will not sit in front of the model at once.

OpenAI documents a comparable mismatch from the other direction: 2 million tokens per text file, against a consumer window it does not publish. A file allowance an order of magnitude larger than any stated window is a specification for a pipeline, not for a chat.

What to do about it is dull and effective: split long documents at chapter or section boundaries yourself, name each part, and ask for a summary per part before asking anything that spans the whole. That is three extra steps and it is the difference between an answer about your document and an answer about the first third of it.

Where the real ceiling usually sits

Documented limit	Claude apps	ChatGPT	Gemini apps
Per file	500 MB in chat; 30 MB in projects	512 MB; about 50 MB for spreadsheets; 20 MB images	100 MB; video up to 2 GB
Files at once	20 per chat	80 per rolling 3 hours; 3 per day on free	10 per prompt
Document depth	PDF up to 1,000 pages; visual analysis only up to 100	2M tokens per text file	Audio 10 min free, up to 3 h paid; video 5 min free, up to 1 h paid
Project or workspace	File count unlimited, but total must fit the window	5 files per project on free, 25 on the mid tiers, 40 on the top ones	Notebook sources: 50 free, 100, 300, up to 600 by tier

Google's notebook product is the outlier worth knowing about, because it is the only place in this comparison with hard published daily numbers: up to 500,000 words per source or 200 MB per uploaded file, Sheets capped at 100,000 tokens, Slides at 100 slides, and chat query allowances of 50, 200, 500, 2,500 or 5,000 a day depending on tier. If your work is "read these forty documents and answer questions about them for a month," that published structure is a better fit than any window figure.

The caps published as multipliers instead of numbers

Two of the three vendors decline to publish consumer message caps. Anthropic's usage page says allowances differ by plan and names five-hour session limits and weekly limits without a single figure. Google publishes only ratios: one paid tier at twice standard limits, another at four times, the top tier at five or twenty times the one below it. OpenAI is the exception and prints ranges for business seats - roughly 5 to 45 messages per five-hour window on one model, 250 to 2,000 on another - with the caveat that "actual usage varies by model and task."

The planning rule that follows is unglamorous: if a workflow has to finish by Thursday, do not build it on an allowance nobody has quantified. Either use a tier with published numbers, or test a full day of the real workload before you commit a deadline to it.

One month of ordinary work, converted

All of the above becomes concrete only when you put a real month through it. Take a small team with three recurring jobs, and convert each using the published figures rather than a guess.

The job	The conversion	What the limits force
Review 12 vendor contracts, 40 pages each	480 pages at 500 words is about 240,000 words, or about 320,000 tokens by OpenAI's rule	Will not fit a 200K window in one pass. One contract at a time is about 27,000 tokens and fits easily, so the answer is 12 passes, not one
Summarise 12 meetings of 50 minutes	Google documents audio at 32 tokens per second, so 600 minutes is about 1,152,000 tokens	More than even a 1M window. Per meeting it is about 96,000 tokens and fits. A 200K window holds roughly 104 minutes of audio; a 1M window roughly 8.7 hours
Upload 40 reference files on a free tier	40 divided by 3 uploads per day	About 14 calendar days before the set is complete, which is usually the real reason a pilot stalls

Notice what the conversion actually decided. It did not pick a brand. It told you the contract job is a twelve-step process rather than a one-step one, that meeting audio must be handled per meeting, and that a free tier turns a two-hour setup into a fortnight. Those three findings are worth more than any head-to-head comparison of writing style.

Transcription: count the languages, because nobody publishes accuracy

If meetings are the job, the documented specification to compare is language coverage, and the differences are large. Microsoft documents 43 spoken languages and locales for Teams transcription and captions, on a page last updated July 9, 2026, with translated captions requiring the premium add-on. Google Meet documents 8 languages for transcripts, and notes that for other Workspace AI features "English is the only supported language." Zoom documents 19 languages for cloud recording transcription and 29 for meeting summaries.

Accuracy is a different story: none of the three publishes a word error rate for these products, and no standards body publishes one either. One vendor publishes a commissioned comparison claiming the best error rates without stating an absolute figure, which is marketing rather than a specification. Treat any accuracy claim you cannot trace to a number as unverified, and budget review time accordingly - especially for names, figures and anything you intend to quote.

Hard caps do exist at the interface level and are worth knowing: OpenAI's speech-to-text documentation states that files "can be up to 25 MB" and advises splitting longer recordings without cutting mid-sentence; Google's audio documentation allows up to 9.5 hours per prompt with inline audio bounded by a 20 MB request, and notes that audio is downsampled and multi-channel input combined to one channel.

What happens to your text by default

This is the specification that most changes what you are allowed to paste, and the three defaults differ.

• Anthropic documents that consumer conversations are not used for training unless the user switches on a setting called Model Improvement, and that incognito chats are excluded even then. Page updated March 16, 2026.
• OpenAI documents the opposite default in plain words: "ChatGPT, for instance, improves by further training on the conversations people have with it, unless you opt out." The same page states that business products and the API are not trained on by default. Updated March 13, 2026.
• Google documents that conversations improve its AI while the Keep Activity setting is on, which is the default, and publishes the bluntest sentence any vendor prints on this subject: "Please don't enter confidential information that you wouldn't want a reviewer to see or Google to use to improve our services, including machine-learning technologies." Human-reviewed conversations may be retained up to three years, disconnected from the account; with the setting off, chats are kept 72 hours.

Government guidance points the same way. The UK government's AI playbook security guidance states that you "must not use public or unassured generative AI tools to process OFFICIAL information" that could cause harm if compromised and carries markings such as sensitive or personal data - though that page shows no publication date, so treat it as living guidance rather than a fixed rule.

Tip: Since no vendor publishes a transcription accuracy figure, input quality is the one variable you control; a dedicated USB conference microphone does more for a usable transcript than switching tools does. (These are Amazon Associate links - we may earn a small commission on qualifying purchases.)

A four-step check before the first upload

The published framework for this is NIST's AI Risk Management Framework, AI 100-1, issued January 2023, and its generative AI profile, AI 600-1, issued July 2024. The framework is organised into four functions - govern, map, measure and manage - carrying 23, 16, 18 and 12 subcategories respectively. Two of its actions translate directly into desk habits: AI 600-1 asks organisations to "deploy and document fact-checking techniques to verify the accuracy and veracity of information generated by GAI systems, especially when the information comes from multiple (or unknown) sources," and to "review and verify sources and citations in GAI system outputs."

1. Check the training default for the exact product and tier you are using, not the vendor in general. Two of the three are on by default.
2. Size the document before uploading it. Pages times words, divided by 0.75, gives tokens; compare that with the window your plan actually documents.
3. Verify every figure and citation in the output against the original. This is not caution, it is the documented control.
4. Keep anything confidential, personal or contractually restricted out of a consumer tier entirely, whatever the window says.

Worth remembering on the claims side too: when the US Federal Trade Commission announced its first sweep of deceptive AI marketing on September 25, 2024, it stated that "there is no AI exemption from the laws on the books," and it has continued to act against inflated AI claims since.

Choosing on the published numbers rather than the pitch

Weighed only against documentation, the sensible order of questions is the reverse of the usual one. Start with what your text is allowed to do, because a default that trains on your chats rules out whole categories of work regardless of capability. Then size the largest document you actually handle and check it against the window your plan documents, not the one in the launch post. Then check the upload ceiling, which is where most small teams hit the wall first. Only after those three does the choice of brand matter, and by then the answer is usually obvious - and often it is two tools, one for long-document work and one for the recurring library of sources.

Part 2 of this series turns those constraints into a repeatable workflow, and Part 3 covers prompts and safe automation.

Documentation pages behind every limit above

• Anthropic model overview and file upload limits - context and output tokens, upload sizes, PDF page handling. Upload page updated July 23, 2026.
• OpenAI on tokens and file uploads FAQ - the 75-words rule, per-file and per-project limits, free-tier upload cap.
• Google on tokens, audio documentation and Gemini apps file limits - token conversion, 32 tokens per second of audio, per-tier file and length limits. Token and audio pages updated September 23, 2026.
• Google notebook source limits - 500,000 words per source, per-tier source and query allowances.
• Microsoft Teams transcription languages - 43 locales, updated July 9, 2026 - and Google Meet transcripts.
• OpenAI data-usage policy (March 13, 2026), Anthropic training policy (March 16, 2026) and Google Gemini apps privacy.
• NIST AI Risk Management Framework - AI 100-1 (January 2023) and AI 600-1 (July 2024).
• FTC, crackdown on deceptive AI claims - September 25, 2024 - and UK government AI security guidance, undated.