A $100 Claude seat holds up to $4,000 of API tokens
A Team Premium seat includes billions of tokens a month. Route your people through your own API key instead, and you pay list price for work the seat already covers.
Most teams treat a Claude subscription and an API key as two ways to pay for the same thing. Subscriptions for people, API for the platform, and if the platform has a nicer UI, point people there.
As a cost decision, that’s wrong by an order of magnitude.
A subscription seat is the cheapest Claude you can buy. Every request you route around it through your own API key is paid at list price, while the seat you already bought sits unused.
To keep it simple, I’ll use one subscription: a Claude Team Premium seat at $100 a month (billed annually).
How many tokens are in a seat?
Anthropic doesn’t publish token allowances. It only says a Team Premium seat gets 6.25× the Pro plan’s usage per 5-hour session, where Max 5x gets 5×.
For absolute numbers, the best source I found is a measurement of Claude Code logs from September 2026. On Max 5x, the 5-hour limit kicked in at about 39 million tokens, and one week used at least 1.3 billion. Scale that by 6.25/5 and a Team Premium seat holds about 49M tokens per session and 1.6B per week, or roughly 7B a month.
That sounds absurd until you see the mix: 96% are cache reads, the same conversation re-read on every request. Priced at Opus 5.5 list rates, that mix costs about $570 per billion tokens. So:
| Team Premium seat | Tokens | At API prices | × the seat |
|---|---|---|---|
| One full 5-hour session | ~49M | ~$28 | |
| A month of working days, two sessions a day | ~2B | ~$1,200 | ~12× |
| The weekly cap, every week | ~7B | ~$4,000 | ~40× |
These are estimates from one measurement, not invoices. But even the average user is well past break-even: Anthropic puts typical Claude Code use at $150–250 a month at API rates.
Paying twice for the same person
A company I know gave every employee a Claude subscription. They also built an internal AI tool with its own UI that talks to Claude through an API key. People like the internal tool, so that’s where a lot of the work happens.
Say 30 messages a day per person, each carrying 25k tokens of context and returning 800 tokens, without caching. At Opus 5.5 rates that’s about $3.50 a day, or $70 per person per month. All of it at list price, on top of a $100 seat that could have absorbed it in two or three sessions. Add agentic loops and it climbs fast.
Plugging people’s subscriptions into the internal tool isn’t an option. Anthropic’s terms say subscription logins are for Claude’s own apps, and it enforces that technically. The cheap tokens only come through Anthropic’s front door.
Bring your tools to Claude, not Claude to your tools
What makes an internal AI tool valuable usually isn’t the chat window. It’s the access to internal systems and knowledge, and you can give that to the apps people already pay for:
- Expose internal systems as MCP connectors, so they show up inside Claude, Claude Code and Cowork.
- Turn your prompts and workflows into skills that everyone can load.
- Keep the API for work without a person behind it: backend features, pipelines, customer-facing products. That’s what it’s priced for.
The caveats
The numbers rest on one coding workload. Chat users get fewer tokens from the same limit, and Anthropic can change limits whenever it likes.
I also suspect the limits aren’t fixed. My impression is that you get more when Anthropic has idle capacity and less when it doesn’t, though I can’t confirm that. Either way, a subscription gives you no guaranteed amount of tokens. With the API, you know what you’re paying for and that it’s there when you need it. “Subsidised” means compared with list price, not with Anthropic’s costs: flat-rate plans work like a gym membership, where light users pay for heavy ones.
On Enterprise none of this applies: it’s $20 a seat plus usage at API rates. And there are good reasons to run your own harness, like provider independence and audit trails. Just don’t assume it’s the cheap option.
Do the division
Take last month’s API invoice for your internal AI tool, divide it by the number of active users, and put it next to $100. If you’ve run this comparison in your own organisation, I’d like to hear what you found. You can find me on LinkedIn.