Put an agent to work. Measure what it does and what it costs.
Assumes Intro to AI for Business. Start here anyway if you like.
Every class teaches one idea, asks you three questions on it and then puts it to work on a real company.
An agent is a loop. It picks a tool, reads the result, and decides whether to go again.
Tool definitions and tool results are input tokens. You pay for them on every turn of the loop.
An agent resends its whole history every turn. Cost grows with the square of the loop and not the length.
Cache what is stable, size the model to the task, and apply a discount to the line it touches rather than the total.
Price the finished task and not the run. The rescue line is usually the big one.
Ask who ran it, when, and what the absolute figure was. Then go and measure your own tasks.
Write the criteria first, measure against your current process, and run it more than once.
Loud failures cost tokens. Silent ones cost trust, and only checking the output finds them.
Gate what cannot be undone. Sample what already went out. Confidence is not a control.
Tools are the permission model. Anything a tool returns is data and never a command.
Find which limit binds before you size a rollout. For agents it is input tokens, and caching raises it.
Seats price people and include the product. The API prices work and you build the rest.
Prices, limits and documentation facts below were read from the vendor's own published pages on September 20 2026 and will change. Any class using one states the date. Recheck before a price is more than 90 days old.
| Source | Figure | Value | Date read |
|---|---|---|---|
| Anthropic pricing page | Claude Opus 5 | $5 per million input. $25 per million output | Sep 20 2026 |
| Anthropic pricing page | Claude Sonnet 5 | $2 per million input. $10 per million output | Sep 20 2026 |
| Anthropic pricing page | Claude Haiku 4.5 | $1 per million input. $5 per million output | Sep 20 2026 |
| Anthropic pricing page | Batch API discount | 50% on input and output | Sep 20 2026 |
| Anthropic pricing page | Cache write multipliers | 1.25x base input for five minutes. 2x for one hour | Sep 20 2026 |
| Anthropic pricing page | Cache read multiplier | 0.1x base input | Sep 20 2026 |
| Anthropic tool use documentation | Tool definitions, tool_use blocks and tool_result blocks | all billed as input tokens | Sep 20 2026 |
| Anthropic tool use documentation | Tool system prompt overhead, Claude Opus 5 | 286 tokens on tool choice auto or none. 406 on any or tool | Sep 20 2026 |
| Anthropic tool use documentation | Tool system prompt overhead, Claude Sonnet 5 | 354 tokens on tool choice auto or none. 474 on any or tool | Sep 20 2026 |
| Anthropic prompt caching documentation | Total input identity | cache reads plus cache creation plus new input | Sep 20 2026 |
| Anthropic rate limits documentation | Claude Opus 5 and Sonnet 5 | 1,000 requests per minute. 2,000,000 input tokens per minute. 400,000 output tokens per minute | Sep 20 2026 |
| Anthropic rate limits documentation | Tier structure | identical throughput across Start, Build and Scale. Tiers differ by monthly spend cap of $500, $1,000 and $200,000 | Sep 20 2026 |
| Anthropic rate limits documentation | Input limit mechanic | only uncached input tokens count toward the input limit for most models | Sep 20 2026 |
| GAIA public leaderboard | Top entries | about 93.36% and 93.02% average, submitted June and May 2026. An entry built on Claude Opus 4.8 at 88.7%, submitted July 2026 | Sep 20 2026 |
| GAIA public leaderboard | Difficulty spread across top entries | easiest tier about 92% to 99%. hardest tier about 76% to 90% | Sep 20 2026 |
| GAIA public leaderboard | Provenance | entries are submitted by the teams that built the systems and are not independently re-run | Sep 20 2026 |
| OSWorld project page | Paper abstract figures | humans complete over 72.36% of tasks. The best model in that paper reached 12.24%. Page notes a verified update dated July 2025 | Sep 20 2026 |
| Anthropic Claude Opus 5 announcement, published July 24 2026 | Agentic benchmark results | described only in relative terms. No absolute percentages published on that page | Sep 20 2026 |
| GitHub Copilot plans page | Copilot Business / Enterprise | $19 / $39 per user per month | Sep 20 2026 |
| OpenAI ChatGPT business pricing page | ChatGPT Business seat | $20 per month billed annually. $25 billed monthly | Sep 20 2026 |
Ten minutes. No account. No card.
Take my first class