Field Notes · AI spending · October 2026
Pay for seats, not tokens
Right now, a flat-rate AI subscription buys far more than the same money spent per use. If someone at your company builds tools with AI, give them a seat, keep track of what their work would cost per use, and treat the deal as a window that will close.
Building your tools on frontier AI makes sense right now. Running your business on it is a different question, which I cover in Keep your decisions in-house. This note is about the first part: what to pay for so your people can build.
The math favors seats
For an owner, the question is who gets a seat, and the math favors seats right now. A $100 to $200 monthly plan gives one person far more frontier model time than the same money buys through the pay-per-use API. At today's list prices, $200 of API use buys about 7 million words of output from Anthropic's top model, Opus 5.5. Someone building tools all day with an AI coding agent can burn through that in days, because the agent rereads files and test results on every step.
| Option | Monthly cost | Who it's for |
|---|---|---|
| Team Standard seat | $20 to $25 per person | Staff who write, summarize and look things up |
| Team Premium seat | $100 to $125 per person | The one or two people who build your tools |
| Enterprise | $20 per seat, plus usage at API rates | Larger rollouts where cost should follow use |
| API, pay per use | Opus 5.5: $4 per million tokens in, $20 out | Systems that run on their own, with no one at the keyboard |
Anthropic list prices, October 2026.
The vendors have said so
When Anthropic added weekly caps to its plans in 2025, it said one subscriber was consuming "tens of thousands in model usage on a $200 plan." It also said the caps would touch fewer than 5% of users. Most people use far less than they pay for, and the heaviest builders get far more.
Treat it as a window
This is not a permanent price. The signs that it is narrowing are already visible: weekly caps, extra usage billed at API rates once you hit them, and enterprise plans that charge for usage instead of seats. Ask your builders for a monthly estimate of what their work would cost at API rates. Then a price change is a budget line you already know, not a surprise.
Seats are for people building, not for systems running. Anything that runs on its own, on a schedule, or serves the plant floor belongs on a pay-per-use account or on your own models. Plan terms and caps are written for a person at a keyboard.
Buy a business plan for business data. Team and Enterprise plans come with commercial terms. Personal plans should never see customer drawings, pricing or employee records.
The code itself is now a utility
For writing code, the top frontier models have become dependable enough to plan around, the way you plan around power from the grid. On SWE-bench Verified, a test built from real bug fixes in open-source projects, the three best models now score between 95% and 96%. Two years ago, scores above 50% were news.
Developers' own reports point the same way. In Sonar's 2026 survey of more than 1,100 developers, AI already wrote 42% of committed code, and 72% of developers who had tried the tools used them every day. But 96% don't fully trust AI-generated code, and only 48% always check it before committing. That gap is where your risk lives.
What this means for a manufacturer:
- The model is no longer the scarce part. Any of the top models will write a working scheduler, report or integration. Pick one and stop shopping.
- Your edge is the work around the model: clear specs, tests that encode how your plant actually runs, and a reviewer who knows the process. Those are the things a vendor can't sell you.
- Small teams can now build what used to need a contractor. One person with a seat and good tests can maintain tools that once required an outside firm.
What to do this month
- Give the one or two people who build your tools a business-tier seat with the higher usage limit.
- Ask them for a monthly estimate of what their work would cost at API rates, and keep it on file.
- Move anything that runs on its own off subscriptions and onto a pay-per-use account or your own models.
Work with us
Not sure who should get a seat?
The answer depends on who builds your tools today, what they're building and what data it touches. In 30 minutes we can sort out who needs which plan and what to keep off subscriptions entirely.
Book a 30-minute callSources
- Anthropic API pricing: Opus 5.5 rates
- Claude plans and pricing: Team and Enterprise plans, usage beyond limits
- Tessl: Why Claude Code is capping power users (August 2025): Anthropic's statements on heavy use and caps
- BenchLM: SWE-bench Verified leaderboard (October 2026)
- Sonar: State of Code developer survey (January 2026)