Claude (AI)
|
Claude
Anthropic's family of large language models
| |
 i
| |
| Developer | Anthropic |
|---|---|
| Type | Large language model family |
| Initial release | March 14, 2023 |
| Current models | Fable 5 · Opus 5 · Sonnet 5 · Haiku 4.5 |
| Limited access | Mythos 5 · Mythos Preview |
| Context window | Up to 1 million tokens |
| Training method | Constitutional AI and reinforcement learning |
| Available on | claude.ai · Claude API · Amazon Bedrock · Google Cloud · Microsoft Foundry |
| License | Proprietary |
| Website | claude.ai |
[ Show / Hide ]
| ||||||||||||
Claude is a family of large language models made by Anthropic, an American artificial intelligence company. The same models are sold three ways: as a chat assistant that anyone can open in a browser, as a set of desktop and office tools, and as a service that software developers build on top of. Anthropic first released Claude on March 14, 2023, and has shipped a new generation roughly every few months since.
Claude models are named by size tier. Haiku is the small, fast one, Sonnet is the middle option, and Opus is the large one built for hard reasoning and long jobs. In June 2026 Anthropic added a tier above Opus, called Mythos, whose first public member is Claude Fable 5. A restricted sibling, Claude Mythos 5, runs the same underlying model with fewer built-in limits and is offered only to vetted organizations.
Anthropic markets Claude on reliability as much as raw ability. Its training method, published in December 2022 under the name Constitutional AI, teaches a model to judge its own answers against a written list of principles. That approach, along with Anthropic's habit of shipping safeguards alongside capability, has shaped how the company describes almost every Claude release.
1 Background✎
Anthropic was founded in 2021 by a group of researchers who had left OpenAI, among them the siblings Dario and Daniela Amodei. The company registered as a public benefit corporation, a legal form that lets a business weigh a stated public mission alongside profit. It described its purpose from the start as building AI systems that are, in its own phrase, helpful, honest, and harmless.
In December 2022 Anthropic published the research paper that still underpins the product. Titled Constitutional AI: Harmlessness from AI Feedback, it described a way to train an assistant to avoid harmful answers without human workers labeling each bad output. The first public Claude models arrived three months later, on March 14, 2023, after a closed test with partners including Notion, Quora, and DuckDuckGo. Anthropic shipped two versions that day: Claude, the capable model, and Claude Instant, a cheaper and faster one.
2 Model families✎
The three size names arrived together with Claude 3 in March 2024 and have stayed in use since. Haiku is tuned for speed and volume, Sonnet balances cost against ability, and Opus handles the hardest work. Anthropic added the Mythos tier above Opus in June 2026, splitting it into a widely released model and a restricted one.
Claude Fable 5 is the widely released Mythos-class model and Anthropic's most capable public offering. Claude Mythos 5 shares its specifications and price but is available only through Project Glasswing, an invitation-only program aimed at defensive cybersecurity work. Figures below are Anthropic's published specifications as of August 2026.
| Model | Built for | Context window | Price per million tokens |
|---|---|---|---|
| Claude Fable 5 | Long-running agent work at the highest available capability | 1 million | $10 in, $50 out |
| Claude Opus 5 | Complex coding and business work; the recommended default | 1 million | $5 in, $25 out |
| Claude Sonnet 5 | Everyday production work where speed also matters | 1 million | $3 in, $15 out |
| Claude Haiku 4.5 | High-volume, low-latency tasks such as sorting and extraction | 200,000 | $1 in, $5 out |
A context window is the amount of text a model can hold in view at once, counted in tokens. A token is roughly three-quarters of an English word, so a million tokens is somewhere near 750,000 words. All four current models accept images as well as text, and all of them answer in text only.
3 Version history✎
Anthropic has released a named model every few weeks since 2023, and retires older ones on a published schedule. Every public release is listed below, with status as of August 2026.
| Version | Release date | Status |
|---|---|---|
| Claude | 2023.03.14 | Discontinued |
| Claude 2 | 2023.07.11 | Discontinued |
| Claude Instant 1.2 | 2023.08.09 | Discontinued |
| Claude 2.1 | 2023.11.21 | Discontinued |
| Claude 3 Opus | 2024.03.04 | Retired |
| Claude 3 Sonnet | 2024.03.04 | Discontinued |
| Claude 3 Haiku | 2024.03.13 | Discontinued |
| Claude 3.5 Sonnet | 2024.06.20 | Discontinued |
| Claude 3.5 Sonnet (new) | 2024.10.22 | Discontinued |
| Claude 3.5 Haiku | 2024.10.22 | Discontinued |
| Claude 3.7 Sonnet | 2025.02.24 | Discontinued |
| Claude Sonnet 4 | 2025.05.22 | Retired |
| Claude Opus 4 | 2025.05.22 | Retired |
| Claude Opus 4.1 | 2025.08.05 | Retired |
| Claude Sonnet 4.5 | 2025.09.29 | Active |
| Claude Haiku 4.5 | 2025.10.15 | Active |
| Claude Opus 4.5 | 2025.11.24 | Active |
| Claude Opus 4.6 | 2026.02.05 | Active |
| Claude Sonnet 4.6 | 2026.02.17 | Active |
| Claude Mythos Preview | 2026.04.07 | Limited availability |
| Claude Opus 4.7 | 2026.04.16 | Active |
| Claude Opus 4.8 | 2026.05.28 | Active |
| Claude Mythos 5 | 2026.06.09 | Limited availability |
| Claude Fable 5 | 2026.06.09 | Active |
| Claude Sonnet 5 | 2026.06.30 | Active |
| Claude Opus 5 | 2026.07.24 | Active |
An active model can be called on the Claude API today. A retired one still runs for existing customers but is no longer offered to new ones, while a discontinued model has been switched off entirely. Mythos-class models marked as limited are reachable only through Project Glasswing.
4 How people use Claude✎
Most people meet Claude through claude.ai or the phone app, where it works like any other chat assistant. Anthropic sells paid tiers that raise usage limits and unlock the heavier tools. Developers reach the same models through the Claude API, and through Amazon Bedrock, Google Cloud, and Microsoft Foundry.
The bigger shift since 2025 has been in tools that do work rather than answer questions. Claude Code puts the model in a terminal or an editor, where it can read a codebase, write changes, and run them. Claude Cowork applies the same idea to office work for people who do not write code. Extensions place Claude inside Chrome, Excel, PowerPoint, Word, and Outlook, and Claude Design turns a written brief into slides, mockups, and prototypes.
Tying these together is the Model Context Protocol, an open standard Anthropic published on November 25, 2024. Before it existed, connecting an assistant to a calendar, a database, or a ticketing system meant writing a custom bridge for each pairing. The protocol replaces those one-off bridges with a single format, and other AI companies have since adopted it.
5 Training and safety✎
Constitutional AI works in two stages. First the model is shown its own answers and asked to criticize and rewrite them against a written list of principles, drawn partly from sources such as the Universal Declaration of Human Rights. Then a second model scores pairs of answers against the same principles, and those scores train the released model. Anthropic argues the method scales better than paying people to label harmful output by hand, and that it makes the values being taught easier to inspect.
Not everyone accepts the framing. Legal scholars have argued that calling a short list of principles a constitution borrows authority the document has not earned, since it is written by one company and can be revised at will. Anthropic has run at least one public experiment in response, training a model on a constitution drafted with input from about 1,000 Americans.
Recent releases pair capability with limits built into the product. Claude Fable 5 ships with classifiers that watch for requests in sensitive areas such as cybersecurity and biology, and quietly hand those requests to a less capable Claude model instead. Anthropic says the filters are tuned cautiously and sometimes catch harmless questions. Claude Sonnet 5 was the first model in its tier to carry real-time cybersecurity safeguards of the same kind.
6 The June 2026 export controls✎
Three days after Fable 5 and Mythos 5 launched, on June 12, 2026, the United States Department of Commerce ordered Anthropic to bar foreign nationals from both models, anywhere in the world. Reporting traced the order to a security finding from Amazon researchers, who had shown that a jailbreak could push Fable 5 into identifying software flaws and writing code to exploit one. Because Anthropic could not check a user's nationality in real time, it switched both models off for everyone.
The shutdown lasted nineteen days. Commerce cleared a limited return of Mythos 5 to vetted American organizations on June 26, then lifted the controls entirely on June 30. Fable 5 came back worldwide on July 1 across claude.ai, Claude Code, and Claude Cowork. Anthropic agreed in exchange to work with the government on testing procedures and on release standards for future models.
7 The copyright settlement✎
In August 2024 three authors sued Anthropic over the books used to train Claude. The company had bought and scanned millions of print copies, which was not in dispute, but it had also downloaded hundreds of thousands of digital books from two piracy sites. In June 2025 a federal judge split the question: training on lawfully obtained books could count as fair use, while building a permanent library of pirated files could not.
Anthropic settled in September 2025 for $1.5 billion, the largest copyright payout in American history. The class covered 482,460 works and set payment at roughly $3,000 each, and the company agreed to destroy the pirated files it had kept. A judge granted final approval on July 20, 2026. Because the case ended in a settlement rather than an appeal, the fair use ruling binds no other court, which leaves the wider legal question open.