Claude.ai pricing starts at $0 and tops out at $200 a month for one person. Free costs nothing and Pro is $17 a month on the annual plan. Everything above that sits in the table below.
The published price is the easy half of the question. The half that decides your real bill is usage, and Anthropic publishes no message count for any plan, because a message no longer costs a fixed amount.
Two prompts that look identical on screen can differ by a factor of fifty in what they consume. That number is what you are actually buying.
Claude.ai pricing plans, and what each tier costs
Here is the full ladder, taken from Anthropic's plans and pricing page and its Max plan documentation.
| Plan | Monthly | Annual, per month | What the money buys |
|---|---|---|---|
| Free | $0 | $0 | Chat on web, mobile and desktop, with web search and extended thinking |
| Pro | $20 | $17 | More usage than Free, plus Claude Code and Cowork |
| Max 5x | $100 | Not offered | Five times Pro's usage per session |
| Max 20x | $200 | Not offered | Twenty times Pro's usage per session |
| Team, standard seat | $25 per seat | $20 per seat | More usage than Pro, with SSO and admin controls |
| Team, premium seat | $125 per seat | $100 per seat | Five times a standard seat's usage |
| Enterprise | Quoted | Quoted | Seat price plus usage at API rates |
Two details in that table catch people out.
The Max tiers are quoted monthly only, so the annual discount that trims Pro from $20 to $17 has no equivalent at the top of the ladder. Committing to $2,400 a year apparently earns you no further sympathy.
The second is that the Max multipliers are stated per session. They are not a monthly allowance you can spend in one weekend, which matters a great deal once you read the next section.
Team deserves a second look if you are two people instead of twenty. A standard Team seat costs the same $20 a month as annual Pro and carries more usage, and a premium seat is effectively Max 5x with central billing attached. Team plans are sold to workspaces, so check the seat minimum before treating one as a cheaper Pro.
What Claude AI usage limits actually measure
A usage limit is a budget, and the unit is consumption. Messages are not the unit.
Anthropic's help centre describes it as a conversation budget, moved by the length and complexity of your conversations and by the model you are talking to, among others. The effort level you select moves it as well.
In practice the dominant variable is context. Every turn resends the entire conversation, so a chat forty exchanges deep costs far more per reply than the same question asked in a fresh window. The context window, meaning how much text the model holds at once, is the meter.
Three mechanics decide when you run out:
- A session limit that resets every five hours, on both Pro and Max.
- A weekly limit on Max plans, applied across all models, resetting at a fixed time assigned to your account.
- One shared pool across claude.ai, Claude Desktop and Claude Code, with no separate allowance per surface.
That third point is the one that surprises people. An afternoon in Claude Code draws on the same budget as an afternoon in the chat window, and it draws faster, because an agent will happily read forty files to answer something you could have settled by reading one.
Your next reset time lives in Settings, under Usage. Looking at it before you start something long is a cheap habit. The same Free, Pro and Max ladder governs other parts of the product too, including what you can recover after deleting a project.
Why a Claude AI rate exceeded error appears when you still have usage left
Two different failures wear similar wording on screen, and they call for opposite responses.
The first is your own limit. You have spent the session budget, Claude tells you when it resets, and the only ways forward are waiting or paying more, whether that means an upgrade or usage credits.
The second has nothing to do with your account. Anthropic's error troubleshooting guide records this message verbatim: "Due to unexpected capacity constraints, Claude is unable to respond to your message. Please try again soon." That is demand management on Anthropic's side, and it deliberately never reaches the status page, because nothing is broken. Trying again in a few minutes usually clears it.
Telling the two apart saves you an upgrade you did not need. A capacity message on your first prompt of the day is not a plan problem, and buying Max at nine in the morning will not make it go away.
On the API the same family of failure arrives as an HTTP 429. There the fix is genuinely yours: back off and retry with jitter, then apply for a higher rate limit tier once your traffic has outgrown the one you are on.
Subscription or API: which claude.ai bill comes out smaller
A subscription buys capacity you might not use. The API bills only what you consume. Which one wins depends on how many hours a week you genuinely work with the model.
Current rates, from Anthropic's API pricing documentation:
| Model | Input, per million tokens | Output, per million tokens |
|---|---|---|
| Claude Opus 5 | $5 | $25 |
| Claude Sonnet 5 | $2 | $10 |
| Claude Haiku 4.5 | $1 | $5 |
Sonnet 5 is the interesting row. Its $2 and $10 rates launched as introductory pricing due to expire on 31 August 2026, and the scheduled increase to $3 and $15 will not happen. Those are the standard prices now.
Two discounts change the arithmetic more than model choice does. Prompt caching bills a cache read at a tenth of the base input price, against a write at 1.25x for the five-minute cache, so one reuse puts you ahead. It is the rare optimisation that pays for itself before you have finished configuring it. The Batch API then takes 50 per cent off input and output together, for anything that can wait.
Rough shape of the decision: $20 of Pro is worth about ten million input tokens on Sonnet 5, or two million tokens of generated output. Send less than that in a month and stay out of Claude Code, and the API is cheaper. Work in the tool most days and the subscription stops being close.
A third option removes the choice. Usage credits let a Pro or Max subscriber keep working past the session limit at standard API rates, billed separately from the subscription. Credits generally do not expire, and there is a $2,000 daily redemption cap. You can set a monthly spending ceiling as well, and the moment to set it is before you enable credits.
Where Claude Code and AI agents spend your plan
Agents are the reason usage estimates built from a year of chat stop predicting anything.
A chat turn is one round trip. An agent turn is a loop: read a file, decide, call a tool, read the result, decide again. Every step carries the accumulated context forward, so total token spend grows faster than the number of steps in the loop.
On Claude Code that comes out of the same subscription pool as your chats. On Claude's Managed Agents the billing is explicit and worth reading before you build anything: tokens at standard model rates, plus $0.08 for every session-hour the agent spends running. Idle time is not charged.
Server-side tools carry their own line items. Web search costs $10 per 1,000 searches. Code execution gives each organisation 1,550 free hours a month and charges $0.05 per hour per container beyond that, and it is free outright when used alongside web search or web fetch.
None of this is expensive per unit. It is expensive per forgotten loop. An agent with no stopping condition will call the same tool until something else stops it, and the something else is usually your invoice.
If you are designing agents instead of paying for someone else's, Agentic AI Engineering covers reasoning loops and tool design at the level where these costs are actually decided.
Claude coding vs other AIs, and picking a plan without overbuying
Comparing Claude with another vendor on sticker price per million tokens will mislead you, and the reason is technical.
Anthropic's pricing documentation notes that Claude 4.7 and later models use a newer tokenizer that produces roughly 30 per cent more tokens for the same text. The rate per token came down while the token count for identical input went up. A cross-vendor comparison of dollars per million tokens is therefore comparing two different definitions of a token. LLM Systems Engineering goes into how tokenizers are built and why swapping one changes every number downstream.
The comparison that survives contact with reality is your own workload. Run the same fifty real tasks through each candidate, record the total spend and the number of attempts you had to throw away, then divide.
For coding, the variable that dominates is rework. A model costing twice as much per token that gets there in one attempt instead of three is cheaper, and cheaper again in the hour you did not spend reviewing two dead ends. This is also why a Claude AI review written by someone else is thin evidence. Their tasks are not yours, and neither is their tolerance for a wrong answer.
A rule of thumb that holds for most people: start on Free and move to Pro the first week you hit a limit twice. Move to Max only after a month of Pro in which the five-hour reset keeps interrupting work you were being paid to do.
Before any upgrade, spend ten minutes in Settings under Usage. The reset times are there, and the pattern in them usually says whether you have a plan problem or a habit problem. Forty-turn conversations that should have been four fresh ones are the most common habit, and unlike the plan, that one is free to fix.
If Claude is new to you and you are not yet sure which parts of the product your money is unlocking, Claude AI for Beginners works through model choice and the Claude Code and Cowork surfaces from the ground up.