What it is
Claude is a family of AI language models made by Anthropic, an American AI company. You can talk to Claude in a chat window, much as you would with ChatGPT or Gemini, and developers can build it into their own apps. Like the others, it's a language model: a system trained on huge amounts of text that writes its replies one small piece at a time.
Anthropic was founded in 2021 by a group of seven former OpenAI employees, including the siblings Dario Amodei, who had been OpenAI's vice president of research and became Anthropic's chief executive, and Daniela Amodei, who became its president. The company is set up as a public benefit corporation, a type of business that is legally allowed to weigh a stated public purpose alongside profit. Anthropic's stated purpose is "the responsible development and maintenance of advanced AI for the long-term benefit of humanity". Its founders left OpenAI over differences about direction, and the company has made safety research a central part of how it presents itself. Whether it lives up to that is a fair question, and this page looks at both sides.
"Claude" is not one model. It's a family that has changed many times since 2023, and each generation comes in several sizes. The names can be confusing, so here is the shape of it.
The tiers. When Anthropic released the Claude 3 family on 4 March 2024, it introduced three names that it still uses: Haiku, described as the fastest and most compact; Sonnet, the middle option balancing speed and intelligence; and Opus, the largest and most capable. Bigger usually means slower and more expensive. In 2026 Anthropic added two more names above Opus, Mythos and Fable, which are explained in the first example below.
The generations, at a high level, using the dates in Anthropic's own announcements:
- March 2023: the first Claude, and a faster, cheaper Claude Instant, offered to businesses after testing with partners such as Notion, Quora and DuckDuckGo.
- July 2023: Claude 2, and the first public chat website, claude.ai, open in the US and the UK.
- March 2024: the Claude 3 family, with the Haiku, Sonnet and Opus tiers.
- June 2024: Claude 3.5 Sonnet, free on claude.ai, launched with Artifacts, a side panel where Claude shows code, documents and designs you can edit.
- February 2025: Claude 3.7 Sonnet, which added a visible thinking mode (see Extended Thinking).
- May 2025: Claude Opus 4 and Claude Sonnet 4, followed through 2025 and early 2026 by versions 4.1, 4.5, 4.6 and later.
- June to September 2026: Claude Fable 5 and Claude Mythos 5 (9 June), Claude Sonnet 5 (30 June), Claude Opus 5 (July), Claude Fable 5.1 and Mythos 5.1 (1 September), and Claude Opus 5.5 (22 September).
At the time of writing, Anthropic's model overview lists four current models: Claude Fable 5.1, which it describes as "for demanding reasoning and long-horizon agentic work"; Claude Opus 5.5; Claude Sonnet 5, "the best combination of speed and intelligence"; and Claude Haiku 4.5, "the fastest model". This line-up will change, probably within months, so check the current list yourself.
Three ways to use it. People often mix these up, and they're paid for in different ways:
- The Claude apps (claude.ai on the web, plus desktop and phone apps) are the chat product. There's a free plan and paid subscriptions.
- The Claude API is for developers. An API lets one program use another, so a business can put Claude inside its own website or tool. It's paid per use, by the token, through a separate developer account, and Claude is also sold through Amazon, Google and Microsoft's cloud services.
- Claude Code is a coding agent that works inside a software project, reading files, making changes and running tests. At the time of writing it's included in the paid subscriptions or can be paid for through the API. It's covered on the Coding Agents page.
Why it matters
It's one of the handful of families most people will actually use. Alongside GPT from OpenAI and Gemini from Google, Claude is one of the main general-purpose models on the market, and it sits inside many other products you may not realise use it. Knowing what the names mean, and what you're paying for, saves money and frustration.
The choice of model changes your results. A quick model like Haiku costs a fraction of a large model and replies much faster, which is ideal for simple, repetitive jobs. A large model may handle a tangled, multi-step task that a small one gets wrong. At the time of writing, Anthropic's API prices range from $1 per million input tokens for Haiku 4.5 to $10 for Fable 5.1, with output costing five times the input price. Picking the smallest model that does the job well is one of the easiest savings available to anyone building with AI.
Anthropic markets itself on safety, which invites scrutiny. The company publishes a lot about how it tests and constrains its models: its system prompts, a written constitution that guides Claude's behaviour (see Alignment), research on interpretability, and even an engineering postmortem when bugs made Claude's answers worse in 2025 (see Observability). That openness is useful to readers. But a company that makes safety its selling point also deserves to be held to it, and it has faced serious criticism:
- It settled a copyright class action brought by authors over pirated books used in training for $1.5 billion (see Datasets).
- In February 2026 it rewrote its main safety policy and dropped a central pledge, which critics saw as weakening its commitments (see the second example).
- In August 2025 it changed its consumer terms so that chats on the Free, Pro and Max plans, and in Claude Code, could be used to train future models if users allow it, with those chats kept for up to five years. Users were asked to choose, and business and API use were excluded.
- In late February and early March 2026 the US Department of Defense labelled Anthropic a "supply chain risk" after the company refused to drop limits on using Claude for mass domestic surveillance and fully autonomous weapons. Anthropic said it would challenge the designation in court. People disagree sharply about who was right, but it shows that the rules a company writes for its model can have large real-world consequences.
Real-world examples
A model held back, a public version, and a 19-day shutdown (April to July 2026)
On 7 April 2026, Anthropic announced a model called Claude Mythos Preview and said it would not release it to the public. Its reason was cybersecurity. Anthropic said the model could find and exploit software vulnerabilities (flaws that let attackers break in) better than all but the most skilled human experts, and that it had found thousands of previously unknown flaws in major operating systems and web browsers. Instead of selling it, Anthropic launched Project Glasswing, giving access to a small group of organisations, including Amazon Web Services, Apple, Google, Microsoft, Nvidia and the Linux Foundation, so they could find and fix flaws in important software first.
On 9 June 2026, Anthropic released two versions of its next model. Claude Mythos 5 went only to vetted security professionals and selected biomedical researchers. Claude Fable 5 went to everyone. Anthropic said the two were "the same underlying model", with Fable carrying extra safeguards: when classifiers (automatic checkers) flag a request about cybersecurity, dangerous biology or chemistry, or copying the model, the answer is handed to the older Claude Opus 4.8 instead. Anthropic said more than 95 percent of sessions never trigger this fallback.
Three days later, on 12 June, the US government placed export controls on both models, restricting access by foreign nationals. Anthropic said it couldn't check users' nationality in real time, so it suspended the models for all users. According to Anthropic, the trigger was that researchers at Amazon had found a way around Fable 5's safeguards to get it to identify software vulnerabilities; Anthropic also noted that several less capable models could find the same flaws. The controls were lifted on 30 June and access returned on 1 July, after Anthropic deployed a stronger classifier that it says blocks that technique in over 99 percent of cases.
The episode shows both sides at once. A company chose not to release its most capable model, which is unusual. It also shows that safeguards can be bypassed, and that anyone relying on one AI product for daily work can lose it overnight for reasons outside their control.
The Responsible Scaling Policy: a promise made, tested, and rewritten (2023 to 2026)
On 19 September 2023, Anthropic published its Responsible Scaling Policy, or RSP. It set out "AI Safety Levels", loosely modelled on the biosafety levels used for laboratories that handle dangerous germs. ASL-2 covered the models of the day, including Claude. ASL-3 described models that could meaningfully help someone cause a catastrophe, for example with biological weapons, and required much stronger security and safeguards. The policy also said Anthropic would pause scaling up its models if its safety measures couldn't keep pace.
The policy was put to use on 22 May 2025, when Anthropic released Claude Opus 4 with ASL-3 protections switched on. It was careful to say this was "a precautionary and provisional action": it had not yet determined that the model definitely needed them.
Then, in version 3.0, effective 24 February 2026, Anthropic rewrote the policy. The new text explains the change as a collective action problem: if one developer paused to add safety measures while others raced ahead without them, the developers "with the weakest protections would set the pace". It now separates what Anthropic plans to do as a company from what it recommends for the whole industry, which it says it "cannot commit to following" on its own. It added new requirements to publish Frontier Safety Roadmaps, which it describes as public goals rather than hard commitments, and regular Risk Reports reviewed by outside experts. TIME, which broke the story, described it as dropping Anthropic's flagship safety pledge, and quoted Chris Painter of the evaluation nonprofit METR warning that without firm tripwires risks could rise gradually without setting off alarms. Anthropic later added, in version 3.1, that it remains free to pause development whenever it judges that right. At the time of writing the policy is on version 3.4.
Whatever you conclude, the lesson for a reader is practical: company safety promises are voluntary and can change. Read the current version, and read the change log.
Weekly limits for the heaviest users (August 2025)
In late July 2025, Anthropic announced that from 28 August it would add weekly usage limits to its Pro and Max subscriptions on top of the existing limits that reset every five hours. It said the change would affect fewer than 5 percent of subscribers, and explained that some people were running Claude Code around the clock and some were sharing or reselling accounts. Heavy users complained, and many people learned for the first time that a flat monthly price didn't mean unlimited use. The structure has stayed: at the time of writing, Anthropic's help pages describe both a five-hour session limit and a weekly limit on paid plans.
An everyday scenario: choosing a model for a weekly newsletter
Imagine Priya, who runs a small online pottery shop and writes a weekly email newsletter. She uses the free plan and notices the model menu. For quick jobs, such as ten subject-line ideas or tidying a product description, she tries the fastest model and finds it good enough and almost instant. For the monthly long-form issue, where she wants a thoughtful story about a firing that went wrong, she switches to the more capable model and finds the draft warmer and better organised. When she upgrades to a paid plan months later, she keeps the same habit: small model for small jobs, bigger model only when the result is noticeably better. That also stretches her usage allowance further.
Try it yourself
This exercise gives the same writing task to two Claude models so you can judge the difference for yourself rather than trusting anyone's rankings, including Anthropic's. Allow about twenty minutes. Menus change often, so treat the labels as a guide.
- Sign in at claude.ai. A free account works. At the time of writing, Anthropic's pricing page lists Sonnet and Haiku models on the free plan; paid plans add Opus and larger models. Check what your plan offers.
- Find the model picker. It's usually a small menu near the message box showing the current model's name. Note which models you can choose.
- Start a new chat and pick the faster, smaller model (for example a Haiku model). Paste this prompt, filling in the brackets:
Write a 150-word welcome email for new subscribers to my newsletter about [your topic]. The readers are [who they are]. Use a friendly, plain tone, one short personal story, and end with one clear thing to do next. No exclamation marks. - Time it roughly from pressing send to the end of the reply. Copy the result into a document and label it with the model name.
- Start a second new chat, switch to the other model (for example a Sonnet model, or Opus if you have a paid plan), and paste exactly the same prompt. Starting fresh matters: in the same chat, the second model would see the first answer.
- Time it and save it alongside the first.
- Compare them side by side. Score each from 1 to 5 on: did it follow every instruction (length, no exclamation marks, one story, one call to action)? Does it sound like a person? Would you send it with only light edits?
- Ask a follow-up to both to test how each handles revision. Paste into each chat:
Make it 20 percent shorter and replace the story with one about a mistake I made when I started. Keep everything else. - Decide. If the faster model was good enough, use it for this kind of task. It's quicker and, on paid plans, uses less of your allowance. If the larger one was clearly better, you've learned where it earns its cost. Repeat the test with a harder task, such as summarising a long document, because results vary by job.
Common mistakes and misunderstandings
Treating "Claude" as one fixed thing. A review praising or criticising "Claude" might refer to a model two generations old. Always check which model and which date. Anthropic retires older models on a published schedule, and behaviour can differ a lot between versions.
Assuming the newest or biggest is always best. The largest model is slower, uses more of your allowance and, on the API, costs more. Anthropic's own advice for developers offers two routes: start with a small model and upgrade only if tests show it falls short, or start with a capable model and move down once the task works. Either way, test on your own work.
Treating company benchmark claims as independent proof. Launch posts say things like "state-of-the-art", but those numbers are chosen and measured by the company. That's true of every AI company, not just Anthropic. See Benchmarks.
Confusing usage limits with the context window. Usage limits control how much you can use Claude over time. The context window is how much Claude can read in one conversation. Very long chats, big attachments, larger models, web search and higher thinking effort all use up your allowance faster, and at the time of writing the apps, Claude Code and the desktop app share one allowance. Starting a fresh chat for a new topic helps with both.
Thinking a subscription includes the API. At the time of writing, a Pro or Max subscription covers the Claude apps and Claude Code, but API use is billed separately through a developer account. At the time of writing, Pro is listed at $20 a month and Max at $100 or $200 a month. Prices and allowances change, so check the current pricing page before you buy.
Following old tutorials. Advice written for 2024 or 2025 Claude models may not work on newer ones. Several developer features, such as prefilling and custom temperature settings, are rejected by the newest models.
Believing it's always right. Claude can hallucinate, inventing facts, quotes or sources with confidence, like every language model. It can also agree with you too readily. Check anything important against a reliable source.
Not checking your privacy settings. On consumer plans, whether your chats may be used for training is a setting you control. Look at it in your privacy settings, and don't paste sensitive information you wouldn't want stored. See Privacy.
Related elements
- GPT: OpenAI's model family, the most direct comparison to Claude.
- Gemini: Google's model family, another major alternative worth testing on the same tasks.
- Language Model: the kind of AI that every Claude model is.
- Alignment: covers Constitutional AI and Claude's written constitution.
- Interpretability: research into what happens inside models, an area Anthropic publishes widely on.
- Coding Agents: where Claude Code is explained in depth.
- Computer Use: Anthropic's 2024 release that let Claude operate a computer screen.
- System Prompt: Anthropic publishes the instructions behind Claude's apps.
- Extended Thinking: the visible reasoning mode that arrived with Claude 3.7 Sonnet.
- Context Window: how much a Claude model can read at once, and why bigger isn't automatically better.
- Guardrails: the classifiers and safeguards described in the Fable 5 example.
- MCP: the open standard for connecting AI to tools and data that Anthropic introduced in 2024.
- Datasets: the Bartz v. Anthropic case about the books used to train Claude.