Claude pricing, from the Pro plan to usage-based cost in companies

Claude Pro is listed at 18 euros a month, and Claude Max comes in two tiers, around 85 and around 170 euros. Those amounts are not the real question. Claude pricing plays out elsewhere, on two shifts the pricing page does not put forward.

The first shift is individual. You max out your limit, and the reflex is to upgrade when the problem comes from your habits. The second is collective. Past a certain number of seats, the flat fee disappears and consumption bills at API rates. Tandem has been guiding small and mid-sized companies across both steps for several months.

How much does a Claude subscription cost in August 2026?

Six levels coexist, from free to Enterprise. The amounts sit in the table above, taken from the vendor's pricing page in August 2026. Three points matter more than the figures themselves.

• Annual is the only discount. Pro paid upfront costs 180 euros for twelve months, which is 15 euros a month instead of 18. Only Pro and Team accept annual billing, Max stays monthly, Enterprise is annual only.
• The Pro tier opens the whole product. Claude Code, Cowork, Claude Design, Projects and the Microsoft 365 integration are all included from Pro. The free plan opens the chat window only.
• Max unlocks no extra model. You buy usage reserve, not intelligence. That is the most common confusion among the teams we equip.

The euro figures deserve a note. Anthropic quotes Pro, Team and Enterprise in euros on its French page, but keeps both Max tiers in dollars. The euro amounts are not a straight conversion from the dollar, they reflect local pricing. A standard Team seat lands at 21.04 euros a month, where a mechanical conversion would give a lower figure.

What Max really buys you compared with Pro

Max is the same Claude with a bigger reserve. Anthropic's documentation describes two tiers, five times and twenty times the Pro plan's usage per five-hour session. Added to that are higher output limits, priority access during busy hours and early access to new features.

Metering runs on several simultaneous counters, and that is the most common source of surprise. The app's usage panel shows four of them in parallel. A session limit that resets every five hours. A weekly limit across all models. A weekly limit specific to Claude Design. And a limit specific to the Sonnet model.

Maxing out any one of those counters is enough to block the matching usage. The vendor also states it may restrict usage in other ways, through weekly or monthly ceilings and through access to certain models, at its discretion. No paid plan therefore guarantees a firm volume.

Past 150 seats, the flat fee stops

This is the shift leadership teams discover too late. The Team plan is reserved for teams of 2 to 150 people. Beyond that it is no longer available and you move to Enterprise. And Enterprise does not work like a flat fee.

Enterprise pricing combines a price per seat per month with consumption billed at API rates. The vendor puts it this way, costs scale with the models and tasks your team runs. There is no monthly ceiling by design. The bill follows real usage.

What that looks like in practice is documented. Across enterprise deployments, the Claude Code documentation reports an average of roughly 11 euros per developer per active day, and 130 to 215 euros per developer per month. Nine users out of ten stay under 26 euros a day. The original figures are published in dollars.

Compare those orders of magnitude with an 18-euro Pro seat. Sustained agentic usage therefore costs between seven and twelve times an individual subscription. The gap is not an anomaly, it matches the nature of the work. One Claude Code turn carries file contents, tool calls and multi-step reasoning.

One useful nuance. On Team as on Enterprise, each seat keeps a usage reserve tied to its seat tier, and that reserve acts as the default ceiling. Billing at API rates only kicks in beyond it, through usage credits an admin turns on deliberately. Leadership's job is therefore to decide who is allowed to exceed, and by how much. We frame that with teams in our guide to Claude training for teams.

Why token costs blow up

A token is a piece of a word. In our workshops, the example that lands takes two short English sentences. Eleven words, seventeen tokens. Long or rare terms get split into several pieces, and every punctuation mark counts as one.

What you pay for is not your question, it is everything the model sees before answering. The model's own instructions, your context file, the conversation from the start, the tool calls and their results, then the files read. Recent models reach one million tokens, and that ceiling is hard.

Four mechanics push that volume up far faster than actual activity.

The first is the compound effect. With every message, Claude rereads the conversation from the start. The first message of a session weighs roughly 500 tokens. The thirtieth weighs close to 15,500 for a comparable request, some thirty times more.

I put numbers on this in a carousel published on LinkedIn in May 2026, and I have used it in our corporate workshops ever since.

The second mechanic is model choice, and the gap is considerable. The API rates published by the vendor put a one-to-ten ratio between the cheapest and the most powerful model, on input as on output.

The third mechanic is invisible and recent. Anthropic's documentation notes that models from version 4.7 onward use a new tokenizer, which produces about 30 percent more tokens for the same text. At an unchanged headline rate, the same work therefore consumes more on a recent model. Few teams have folded that into their projections.

The fourth mechanic is agentic. Agent teams consume roughly seven times more tokens than a standard session, because each agent maintains its own context window. A cache miss weighs just as heavily, since the first request after a long pause reprocesses the entire context.

Frugality, the only lever that works on both models

This is why context frugality deserves to be treated as a discipline rather than a trick. On a flat fee, it saves you an upgrade. On API rates, it cuts the invoice directly. It is the only lever that works under both billing regimes.

Four moves carry most of the gain, and none of them requires technical skill.

1. Start clean before saturation. Ask for a condensed summary, open a new chat, paste that summary as the first message. I switch sessions at around 60 percent of context filled. Starting fresh costs nothing, condensing a huge conversation costs a lot.
2. Edit rather than correct. When Claude heads the wrong way, the reflex is to type no, try something else. The failed attempt then stays in the context and pollutes every later answer. Go back to the original message, edit it, run it again.
3. Switch off what sits dormant. Every connector and every skill left enabled is read on every message. Claude Code's context panel regularly shows tool definitions taking up a fifth of the window. Enable on demand, disable once the task is done.
4. Tune the model and the effort. A fast model handles a summary or a sort perfectly well. The app flags it itself, the most capable model drains your reserve markedly faster than the tier below. Effort level is adjustable too, and high effort produces more thorough answers while burning your limits faster. Extended thinking is billed as output tokens.

The prompt itself works the same way, and it is the first thing I teach. A literal instruction beats a dressed-up role. Write "audit this contract, list the risks, rate each one from 1 to 5" rather than a paragraph explaining to Claude that it is a seasoned lawyer. Bound the expected length. Phrase in the positive, what you want rather than what you refuse. And let Claude ask you its questions instead of writing five hundred words that anticipate everything.

On usage billed at API rates, three further levers apply and they are substantial. Prompt caching drops the rereading of an already-seen context to a tenth of the input price. Batch processing gives 50 percent off input and output for anything that is not urgent. And loading documents once into a Project avoids paying for them again in every conversation. Claude only rereads the relevant fraction, on the order of a tenth. That last logic is exactly RAG applied to business context.

I developed this frugality logic in our article on what changed about prompting in 2026. The full configuration mechanics live in the ten-step guide to mastering Claude.

Should you commit annually to save money?

Annual commitment saves 36 euros on Pro, 180 instead of 216. The saving is real but modest. It is paid for in flexibility, and that is the actual trade-off.

I settled this question in a video about artificial intelligence subscriptions, before Max even existed. My recommendation has not changed. The market moves too fast to lock twelve months into one ecosystem. I had in fact cancelled my Claude subscription at one point, then came back to it when Cowork and Claude Code arrived. That is exactly the kind of switch an annual commitment prevents.

Monthly billing keeps another advantage. Moving from Pro to Max applies immediately, with the difference prorated. So you can step up for a heavy month, then step back down. Paying for Max all year for three weeks of crunch is the worst possible calculation.

Which plan for which profile in a company

For one person using Claude every day on writing, document analysis and office deliverables, Pro is more than enough. That describes most of the employees Tandem equips.

Max 5x makes sense from two or three hours of sustained daily use. Max 20x only makes sense for continuous agentic usage, with Claude Code or Cowork working in the background for a good part of the day.

For a team, Team brings central billing, admin controls and single sign-on, at 21.04 euros per seat per month. From five people up, that weighs less than managing five individual subscriptions. And if your organisation is approaching 150 seats, the question to work through is no longer plan choice but spend governance. A well-framed use of Cowork costs less than an improvised one, for the same work.

That explains most of the licences sitting idle. A Max subscription used as a chatbot costs ten times too much. A well-configured Pro subscription returns ten times its value. If you are still weighing ecosystems, our comparison of Claude against ChatGPT for business and the article on the reasons to switch to Claude address that upstream question.

Claude pricing, what to settle before you pay

Below 150 people, take Pro, monthly, and stay there. That is the recommendation in almost every situation Tandem encounters. Eighteen euros open the whole product, and the value sits there, not in the size of the reserve.

If you saturate, apply the four frugality moves for two weeks before changing plan. They cost nothing and settle most cases. If the limit still hits after that, Max becomes an honest calculation. Take it monthly, so you can step back down.

Past 150 seats, change your reasoning. You are no longer choosing a package, you are steering consumption. Set the per-user ceilings, pull the spend reports, and train teams on frugality before opening the taps. A variable budget without a culture of restraint gets discovered on the invoice. For non-technical profiles, I detailed the groundwork in our practical guide to Claude Code.