The real cost of AI for a five-person firm, and when local models start to make sense

The base plan is the cheap bit. ChatGPT Business for a five-person firm is £100 a month on Standard seats. Claude Team is similar. None of that is the line item that hits the budget in the second half of the year.

The line item is the working plan. Most teams that take AI seriously outgrow Standard seats inside a month. ChatGPT Business Premium is £100 a seat. Claude Team Premium is £80. Five Premium seats is £400-500 a month before anyone makes an API call. Anthropic Max and ChatGPT Pro both run at $100 or $200 a month per person for the heavy-use tier. Five of those is £800 a month. Some teams pay overage on top.

I run a $200 a month Anthropic Max plan as a single person. I sometimes run out. The limit is weekly, not monthly, and it bites during a heavy code week.

Add the API spend and the line shifts again. A team of five engineers with a Cursor or Claude Code setup can run through £500 to £2,000 of API tokens a month, depending on how much code review and refactor they push through. A working five-person team on the cloud is paying £1,000 to £1,500 a month in total. The line is not the subscription. The line is the working plan plus the embedded use. The maths is the maths.

A £2,500 workstation in late 2026 can run a 70-billion-parameter model at 4-bit quantisation, at around 5 to 15 tokens a second, on the firm's own hardware, on the firm's own data. The hardware is a one-time spend, the model is free, the data does not leave the room.

Three things drive the break-even: the working plan cost today, how much of the use is sensitive enough to want on-prem, and whether the firm's existing staff can stand up a local model box. For a five-person firm at £1,000 a month of combined spend, the workstation pays for itself in three months. The cut is at the working plan, not the base plan. The cut is at the volume.

The local model will not match the frontier on every task. A 70-billion-parameter open-weight model in late 2026 is roughly at the level of GPT-4 class. It is good enough for most code review, most first-draft writing, most customer triage. It is not good enough for the hardest reasoning tasks. You can run the easy work local. Send the hard work to the frontier. The shape of the bill changes.

The first move is to count the bill. Pull the last three months of subscription and API spend. Multiply by four. If the annual number is more than the cost of a workstation and a day of staff time, the math is on.