Claude pricing in August 2026

Every Claude price verified on 21 August 2026 - Pro at $20, Max from $100, Team from $20 a seat, Enterprise at $20 a seat plus usage, and the full per-token API rate card.

Claude costs $0 on the Free plan, $20 a month for Pro (or $200 a year), $100 or $200 a month for Max, from $20 per seat per month for Team, and $20 per seat per month plus metered usage for Enterprise. The API is billed per token, from $1 per million input tokens on Haiku 4.5 up to $50 per million output tokens on Fable 5.

Every figure on this page was checked against Anthropic's own pricing and support pages on 21 August 2026. This is an independent reference site: it cannot sell you a plan, and every purchase link points to Anthropic.

#How much does each Claude plan cost?

There are two entirely separate price lists: subscriptions for the apps, and per-token rates for the Claude API. They do not overlap. An API key gives you no Pro access, and a Pro subscription gives you no API credit.

Three details catch people out. Max has no annual option - Anthropic sells it monthly only, so there is no way to prepay for a discount. Team lets you mix seat types inside one workspace, so a 12-person team can buy ten standard seats and two premium ones. And Enterprise is annual only, with a hard seat floor: 20 seats if you buy it self-serve online, 50 if you go through a salesperson.

#What does the Claude API cost per token?

API pricing is per million tokens, split between input and output. Output is always five times the input rate on the current lineup. Full model specifications live on the Claude models page.

Sonnet 5's $2/$10 rate became permanent on 10 August 2026; a planned rise to $3/$15 on 1 September was cancelled. One caveat that materially changes the maths: Sonnet 5's tokenizer emits roughly 30% more tokens for the same English text than Sonnet 4.6 did, so the real-world saving is smaller than the sticker price implies. That comparison is worked through on the Sonnet 4.6 page.

#The levers that actually cut an API bill

Minimum cacheable prompt length differs by model: 512 tokens on Opus 5 and Fable 5, 1,024 on Sonnet 5, and 4,096 on Opus 4.5 and Opus 4.6. Below that threshold the cache simply does not engage. Server-side tools are billed separately: web search costs $10 per 1,000 searches, web fetch is free, and code execution is free when paired with web search or fetch, otherwise 1,550 free hours a month and then $0.05 per hour. Managed Agents, in public beta, adds $0.08 per session-hour on top of standard token rates. The implementation detail is on the Claude API page.

#How do Claude's usage limits actually work?

Claude subscriptions are metered by two overlapping windows, not by a message count. Understanding this is the difference between choosing a plan sensibly and guessing.

#Window one: the rolling five-hour session

Session-based usage resets every five hours. It is a rolling window, not a clock-aligned block - heavy activity at 2pm still weighs on your 4pm capacity. This is the limit most people hit first, and it is the one Max multiplies.

#Window two: the weekly cap

On top of the session window, Pro, Max and Team all carry a weekly limit across all models that resets on a fixed day and time assigned to your account. It is not the same day for everyone, and you cannot choose it. Check yours under Settings → Usage, because a weekly reset that lands on a Wednesday morning changes how you should pace a Monday-to-Friday workload.

#Opus is counted separately

Anthropic's usage dashboard reports the weekly reset “for Opus only and all other models” - meaning Opus 5 and its predecessors run on their own weekly counter alongside the general one. You can exhaust your Opus allowance while still having room for Sonnet 5. In practice this makes Sonnet the sensible default and Opus the thing you escalate to, which is also the argument the model comparison page makes on cost grounds.

#One pool, every surface

Chat on claude.ai, Claude Code in the terminal, Cowork, Claude Design and Claude for Excel all draw from the same allocation. There is no separate Claude Code quota to track and no separate Design allowance. Anthropic notes that Cowork tasks consume more than chatting does, and that Cowork's Auto mode consumes more than Manual or Skip because of the extra per-action safety checks. Claude Code in VS Code, Cursor and JetBrains counts identically to the terminal.

#What consumes the allocation fastest

Message length, attachment size, how long the current conversation already is, tool use such as Research and web search, the model selected, the thinking effort level, and artifact creation. Two things work in your favour: project knowledge is cached, so re-referencing the same uploaded document only charges the new portions, and automatic context compaction on paid plans does not count toward limits at all.

#What are extra usage credits, and when are they worth buying?

Usage credits let Pro, Max 5× and Max 20× subscribers keep working past their included limits instead of waiting for a reset. They are billed at standard API rates - not at a discounted subscription rate - so a credit-funded hour on Opus 5 costs the same as it would through the API.

Team and seat-based Enterprise plans use a separate credits flow that an owner has to enable. Usage-based Enterprise plans do not use credits at all - consumption is simply billed. If you find yourself buying credits every month, that is the signal to move up a plan: Max 5× at $100 buys five times Pro's allocation for the price of roughly $80 of credits over the base subscription.

#Which Claude plan should you buy?

Work down this list and stop at the first row that describes you.

The distinction that matters most at the Team/Enterprise boundary is compliance, not usage. If a security questionnaire is coming, read the Enterprise page before you buy Team seats - several controls cannot be added later without a plan change. If your team is under 20 people and nobody is asking for audit logs, Claude for business teams covers the Team route in detail.

#Is Claude Pro cheaper than paying per token on the API?

For a moderate chat user, the API is often cheaper. For anyone using Claude Code, it is not close. Here is the arithmetic.

#Worked example: a moderate chat user

Assume 20 working days a month, 30 exchanges a day, and an average of 4,000 input tokens per exchange - a realistic figure once a conversation has some history and a system prompt - with 700 output tokens back. That is 600 exchanges, 2.4M input tokens and 0.42M output tokens a month.

So a moderate user on Sonnet 5 would pay roughly $9 a month at API rates against $20 for Pro. On paper the API wins. In practice you would be giving up the entire product: no claude.ai chat interface, no mobile or desktop apps, no Projects, no Memory, no Research, no Artifacts, no Claude Code, no Cowork. You would be paying a developer to rebuild a worse version of all of it. Pro is not competing with $9 of tokens - it is competing with $9 of tokens plus the engineering to make them usable.

#Where the subscription wins outright

Agentic coding inverts the maths completely. A Claude Code session re-reads files, runs tools, and carries a long context through dozens of turns, so input token counts run orders of magnitude above chat. Anthropic's own Enterprise consumption guide - which prices usage at API rates - offers rough planning estimates of about $215 a month for a typical Claude Code user and around $500 for a power user. Against those figures, Pro at $20 and even Max 20× at $200 are heavily subsidised. If you write code with Claude every day, the subscription is the cheap option and it is not a close call. The developer workflows themselves are covered on Claude for developers.

#The honest verdict

Buy the API if you are embedding Claude in software, need parameter control, or want batch and caching discounts on a predictable pipeline. Buy a subscription if a human is doing the talking. Mixing both is normal and often correct - many teams run Pro or Max seats for people and a separate API key for production traffic. For how Claude's pricing lands against rivals, see Claude vs ChatGPT.

#Are there discounted Claude plans?

Three programs are published, and one common assumption is wrong.

Nonprofit eligibility is verified through Goodstack, which Anthropic says takes two to three minutes for Team. Government agencies, political campaigns, higher education institutions and healthcare systems are explicitly not eligible. Startup credits apply only to the first-party Claude API through the Console - not to AWS Bedrock or Google Vertex AI - and standard pricing resumes automatically when they run out. There is no published consumer student discount.

#Frequently asked questions