Choose Tools Your Team Can Use, and Afford | Heidelberg AI Curriculum
T16-L02
Choose Tools Your Team Can Use, and Afford
Choose an AI tool for a team of three to eight and know what it will cost in a busy month, not just a quiet one. Read an allowance, estimate one task from the vendor's prices, pick seats or pay-per-use, set a cap and an alert, and test that colleagues can actually pick up each other's work.
A research group of five signs up for a team plan at $25 per person a month. That's $125, and the PI, the head of the group, approves it without a second look. In March a PhD student uses the same plan to screen 2,000 full-text papers for a systematic review, and runs the screening three times as the criteria change. The included usage runs out on day two. Someone had switched on extra usage with automatic top-ups, so the work carries on at pay-per-use rates, charged per token, a small piece of text. The March invoice is about $725.
In a company it's a customer service team of four at a lab-supplies distributor. A colleague found a tool "for $20 a month". That turned out to be per person, billed annually, with extra usage billed on top afterwards. The week of a product recall tripled the email volume, and the bill with it.
Nobody did anything wrong. Most AI tools now charge a monthly plan plus usage beyond an included allowance, and the headline price describes a quiet month. This book shows you how to choose a tool for your team that people can use and the budget can carry, including the busy month. You are at Level 2 Power user: your team sees the result. A wrong choice means colleagues who can't open each other's work, a figure nobody can check, or an invoice your group leader has to explain.
After this you can
Read a plan's allowance and say what happens when a busy month uses it up.
Estimate the monthly cost of one team task from the vendor's published prices, before anyone signs up.
Choose between paying per seat and paying per use for your team's workload.
Set a spending cap and an alert, and name the admin who watches them.
Run a short team trial that tests handover, export and access, not only the output.
One shared task your team does every month, with a rough count: papers, emails, reports.
The vendors' pricing pages for two candidates, and the plan your team uses today, if any.
Two or three colleagues willing to try it for a week.
About 75 minutes: 15 to read, about 60 to do.
Last verified 23 September 2026. ChatGPT Business and Claude Team both cost $25 per person a month, or $20 billed yearly. On Claude Team, usage beyond the allowance is billed on top once an admin switches it on. Details for experts at the end of this book.
The idea in one page
Three ways AI tools bill. Know which one you're buying before you compare numbers.
Model
You pay
Good for
The catch
Per seat
A fixed price per person a month
Daily, interactive use by most of the team
Seats nobody uses; a limit per seat you only meet when busy
Seat plus usage
Seats, plus extra usage at a rate once an allowance is used up
Teams with an uneven month
The busy month; automatic top-ups
Pay per use
Per token, page, character or run, often through an API, the route programs use to call a model
Batch jobs: 200 papers, 500 replies
Someone has to build it, and a loop can run all night
Read the allowance. Five questions, answered from the vendor's own pages: What's the unit (messages, tokens, credits, runs)? Per whom (each seat, or pooled for the team)? When does it reset (monthly, weekly, every few hours)? What happens at the end (blocked, slowed, or billed extra)? Who can switch on extra usage, and is it on by default? If a page won't tell you the size of the allowance in numbers, plan for it to run out in your busiest week.
Price one task, then the month. For a pay-per-use job, the arithmetic is short:
items × tokens per item × price per token = cost of one run
cost of one run × runs per month = monthly usage
seats × seat price + monthly usage = monthly cost (+ tax)
Don't trust your guess at tokens per item. Run five real items and read the numbers off the usage page. Then work out two months: a normal month, and a busy month at about three times the volume, with one rerun.
Cap it, then watch it. A cap stops the spending, and it also stops the work. So set the cap at the busy-month figure, and an alert (or a weekly look at the usage page, if there's no alert) at the normal one. Switch off automatic top-ups until you've seen two real months. One named admin, with a deputy, is responsible for the cap. They can raise it in five minutes when a deadline needs it.
Count the costs that aren't on the invoice. Admin time: adding and removing people, checking the usage page monthly. Training: an hour per colleague is realistic. Lock-in: can you get the team's work out in a format that opens without the tool? And a cheap API job needs someone to build and maintain it. That person's hours often cost more than the tokens.
The team test. A tool that works for you can still fail your team: a colleague can't open the result, comments disappear in the export, or the person who has to check the output spends longer than the author saved. So the trial always includes one colleague picking up another's work, and one export opened outside the tool. The tool you use today, even a non-AI one, is a candidate and can win.
Lab: a research group that screens papers every quarter
Structure it. Petra, the group's lab manager, gets the March invoice and the job of fixing it. "Five people use AI every day for writing and reading. Once a quarter we screen about 200 new papers for our living review, a review article we keep updating. A useful result: one tool everyone can use, with a monthly cost I know in advance and a hard maximum the PI has agreed."
Understand it. She reads the team plan's allowance with the five questions. The unit is usage per seat, and it resets on a schedule. At the end, usage credits take over if they're switched on: extra usage, billed at the pay-per-use API price. They were on, with automatic reload, which buys more credits by itself, and no limit. Then she prices the quarterly job, one line at a time. She checks her guess by running five papers and reading the usage page: close enough.
Step
Figure
One ten-page paper
About 10,000 words: around 13,000 tokens, or 17,000 on a recent Claude model
Per paper, with the instructions
Call it 20,000 tokens in and 1,000 out
200 papers
4 million tokens in, 200,000 out
On Sonnet 5, Claude's mid-sized model
$8 in + $2 out = about $10 per run
Through the Batch API, the cheaper route for jobs that can wait
About $5 per run
What March did: 2,000 papers, three runs, on the largest model at $4 per million in and $20 out
$480 in + $120 out = about $600
That changes her problem. The screening was never expensive. March cost so much because big volumes went through the largest model three times, with nothing to stop it. Back to Structure it: "Seats for daily use, plus a small capped budget for the screening. The cap matters more than the plan."
Choose tools. Two candidates: Claude Team, which they have, and ChatGPT Business at the same seat price. By default, neither vendor uses the team's texts to train its models. She knocks out nothing on price, since five seats cost $125 a month either way. The team already has its projects and instructions in the current tool, and switching would cost a week of setup. She keeps it, and changes the terms:
Five Standard seats, the normal kind (Premium seats cost five times as much), billed monthly for three more months, then annually if usage holds. That saves $25 a month from then on.
Usage credits on, automatic reload off, an organisation limit of $50 a month and $25 per person.
The screening runs on Sonnet 5, not the largest model, with a test on five papers before every full run.
Admin: Petra. Deputy: the PI.
Build it. Before she writes the decision, she checks the team side with two volunteers. One screens twenty papers into a table in a shared project. The other opens it, adds a column and exports it to Excel. The export keeps the table and loses the comments, so the group agrees that decisions go into the table, not into comments. A colleague who works with screen magnification checks that she can use the project view at 200% zoom. She can.
Her decision note: normal month $125; busy month (screening, one rerun) $125 + about $20; hard maximum $175; alert when usage credits pass $25; review at the end of December.
What it taught her: the plan was fine. The missing cap and the choice of model were the problem, and neither needed a new tool.
Company: 500 customer replies a month
Structure it. Jonas leads a customer service team of four at a lab-supplies distributor. "We answer about 500 emails a month about orders, delivery dates and product specifications. A useful result: a first draft in our tone, from the order record and our product sheets, that a colleague checks and sends. Budget: under €150 a month."
Understand it. He prices both routes. Four ChatGPT Business seats cost $100 a month billed monthly. The pay-per-use route would be an automation that sends each email and its order record to Claude Haiku 4.5, Anthropic's smallest and cheapest current model, through the API. Anthropic's own example puts 10,000 conversations at about $37, so 500 replies come to about $2 a month. A recall month at three times the volume is about $6.
Then he looks at who would build it. The automation needs someone to connect the mailbox, the order system and the API, and to fix it when a field changes. Nobody on the team has done that, and the IT contractor quotes two days plus an hour a month. Back to Structure it: "The tokens are nearly free either way. What we're really choosing is whether the team drafts in a chat window, or someone builds and maintains a pipeline, an automation that runs by itself."
Choose tools. For now, seats. Four ChatGPT Business seats, billed monthly, because everyone can start tomorrow and the price is flat: there's no usage bill to surprise anyone. The customer data in the emails stays out of training by default on the Business plan. The API pipeline goes into the decision note as the next step, for when volume doubles (T05-L02 · Put AI in the Middle of a Workflow, and Check Every Answer). It will need its own spend limit in the Console, Anthropic's account page for the API, and a named person who maintains it. Admin: Jonas. Deputy: the office manager.
Build it. Two colleagues try it for a week on twenty real emails each, against the Outlook templates they use today. Drafts are faster for product questions and no better for delivery dates, where the date still comes from the order system. One test matters more than the rest: a colleague picks up a half-finished draft from another's shared project and sends it. It works. The team's tone and reply rules live in a plain text file on the shared drive, pasted into the project, so they aren't locked in.
His decision note: normal month $100; busy month $100, because seats are flat; cap: four seats, removed within a week when someone leaves; checked monthly on the admin page; review in three months or at 1,000 emails a month.
In both cases the headline price wasn't the decision. The busy month, the cap and the person who watches it were.
What goes wrong
Reading the headline price.Symptom: "$20 a month" becomes $180 on the first real invoice. Fix: read the allowance with the five questions, and price the busy month before you sign up.
Top-ups with no ceiling.Symptom: extra usage with automatic reload quietly carries a runaway job through the night. Fix: turn off automatic reload, set an organisation limit and a per-person limit, and test a new job on five items first.
A cap that stops the deadline.Symptom: the limit hits in the week of a grant deadline, and nobody knows who can raise it. Fix: alert at the normal-month figure, cap at the busy-month one, and name an admin and a deputy who can change it in minutes.
Seats nobody uses.Symptom: eight seats, three people using them. Fix: bill monthly until you know real usage, look at the admin page once a month, and remove the seats of people who have left.
Annual billing before the trial ends.Symptom: you save 20% and are stuck for a year with a tool the team dropped in week three. Fix: monthly first, annual only after two or three real months.
Counting only the invoice.Symptom: the cheap API route needs a contractor every time something changes. Fix: write the setup and upkeep hours into the decision note next to the dollars.
A tool one colleague can't use.Symptom: most of the team likes it, but the one person who checks every result can't use it with their normal keyboard or zoom settings. Fix: test with each person's normal setup. One unmet must-have outweighs several happy votes.
Do it yourself (60 minutes)
Structure it (10 min). Write the shared task, the monthly volume and who does what. Add a budget and a busy month: what happens at three times the volume?
Understand it (15 min). Open the pricing pages of your current tool and one challenger. Answer the five allowance questions for each, from the vendor's own pages. Run five real items and read the usage.
Choose tools (10 min). Work out a normal and a busy month for each candidate: seats, usage, tax. Add the hours for admin, training and building. Knock out anything whose data terms don't fit your material. Pick one, with billing period, cap, alert, admin and deputy.
Build it (25 min). Set the cap and the alert, and switch off automatic top-ups. Run the team test: one colleague picks up another's work, and one export opens outside the tool. Write the one-page decision.
Exit check
What you hand in: a one-page tool decision for your team with:
the workload: the task, the monthly volume and the busy-month volume;
the monthly cost estimate for a normal and a busy month, with the vendor's prices, the date you read them, and your tokens or units per item from a real test;
the cap and alert, as actually set, and whether automatic top-ups are off;
the admin and deputy, by name;
the team test: who picked up whose work, which export you opened, and what didn't survive;
the non-money costs in hours, and the date you'll review the decision.
It passes when a colleague can recompute your estimate from the page and the vendor's price page, and the cap on the page matches the one in the tool.
Rule to remember
Price the busy month, set the cap, and name who watches it.
Details for experts
Checked on 23 September 2026. You don't need this to follow the book; it's for setting things up or answering a colleague.
What
Details
What it means for you
ChatGPT Business
$25 per user a month billed monthly, or $20 billed annually. At least two seats. Visitors in Germany see €26 or €21. All prices before tax.
Add VAT to every figure in your decision note.
Claude Team seats
Standard seat $25 a month, or $20 billed annually. Premium seat $125 or $100, with five times the usage. For 2 to 150 people.
Worth it only for someone who regularly runs out on a Standard seat.
Usage credits on Claude Team
When a member's included usage runs out, credits let them carry on at standard API rates. The top admin (Anthropic calls the role Owner) switches them on under Organization settings → Usage, sets a monthly limit for the organisation and one per person, and can turn on automatic reload. Anthropic warns that usage can slightly exceed a limit.
This is where you set the cap from this book. Leave a little room below the budget.
Claude API prices
Per million tokens, input / output: Haiku 4.5 $1 / $5, Sonnet 5 $2 / $10, Opus 5.5 $4 / $20. The Batch API halves these.
Input is what you send, output what the model writes. Output costs five times as much, so long answers cost more than long documents.
How much text a token is
Roughly 0.75 English words. Claude models from version 4.7 on count about 30% more tokens for the same text.
Price with the model you'll actually use, and read the real count off the usage page after five items.
API spend limit
In the Claude Console under Settings → Billing → Spend limits. At the limit, requests fail until the next month or until someone raises it.
The same trade-off as the cap: the admin and deputy must be able to raise it quickly.
Claude: Taught. The team plan with usage credits and spend limits in the lab example.
ChatGPT: Taught. The Business plan with flat seats in the company example.
Claude API: Taught. Pay per use, with a spend limit in the Console.
n8n: Catalogued. One way to build the pay-per-use pipeline later.
Cursor: Catalogued. Another seat-plus-usage example, for teams that write code.
Watch
OpenAI API Tutorial: Tokenizers & Response Explained!
OpenAI API Tutorial: Tokenizers & Response Explained!Prof. Ryan Ahmed
T16-L02: Learn to distinguish prompt, completion and total tokens using the described OpenAI API demonstration, then estimate request costs. This 2025 API lesson is not seat-plan or spending-cap administration. Use measured usage and current prices for your actual provider and model.
OpenAI COOKEDVideo · Matthew Berman · Sep 30, 2026
A recap of OpenAI Dev Day 2026 by Matthew Berman covering several announcements. DOTS is OpenAI's always-on personal AI assistant competing with Grok, Muse, and Instinct and inspired by OpenClaw; it lives within ChatGPT, controls ChatGPT and Codex threads, connects to personal accounts (Gmail, Google Docs, Calendar), and its conversations don't count toward usage unless they control threads. Ultrafast is GPT-6 Astra powered by Cerebras chips, running eight times faster but at six times the cost; a new Pro 500 plan ($500/month) offers 25x usage limits. GPT-6.1 is described as affordable, fast, and near GPT-6 Astra quality at a fraction of the price ($2 per million input, $10 per million output). Codex updates include Codex in the Cloud, a refreshed Codex CLI with voice control, Codex code review, a Decisions API, a plugins relaunch, 'Sign in with ChatGPT + bring your tokens,' and ChatGPT Space Links.
I Tested Sonnet 5.5 vs Opus 5.5. What You Need to Know.
I Tested Sonnet 5.5 vs Opus 5.5. What You Need to Know.Video · Nate Herk · Sep 29, 2026
In a video by Nate Herk, Claude Sonnet 5.5 is tested against Opus 5.5 across seven real workflows (landing pages, messy data turned into spreadsheets, motion design/video editing, live dashboards, project planning) to determine when to use each model. Sonnet 5.5 is described as a clear upgrade over Sonnet 5, running 30% faster and costing 30% less for most work. Pricing: Sonnet 5.5 costs $2 per million input tokens and $10 per million output tokens, while Opus 5.5 costs $4 and $20 respectively—roughly double. Drawing on Claude's own guidance, Sonnet suits well-scoped, high-volume, clearly-specified tasks with a way to verify results (everyday coding, bug fixes, polished documents, slides, spreadsheets), whereas Opus suits complex work requiring judgment, creativity, and the most intelligence. In a landing-page demo for 'Perk Form' using identical prompts with both on high effort, Opus produced the better overall output by using real generated product images within brand guidelines, though Sonnet's hero section was stronger. Sonnet ran 28 minutes with 20M input/134K output tokens; Opus ran 52 minutes with 37M input/148K output tokens, costing about double.
Jev - The New AI model that has people talkingVideo · Matt Wolfe · Sep 28, 2026
This video introduces Jev, an AI model from Typesafe AI that skips generating text and instead outputs structured values—a choice, a score, or true/false—making it fast and cheap. Input tokens cost 4¢ per million (or $42 per billion) and output tokens are free because they are too cheap to meter. The speaker notes you can plug Jev into tools like Claude Code to write the code that runs it, giving an example app that detects whether a destination URL is malicious before visiting it.
