What is Claude AI?

Anthropic, the business and research laboratory behind Claude, was founded in 2021 by former OpenAI employees, including siblings Dario Amodei (its CEO) and Daniela Amodei (its president), prior to their former employer’s release of ChatGPT. Central to Anthropic’s founding and ongoing approach to artificial intelligence is an emphasis on AI safety and interpretability, which the company’s co-founders felt they could pursue more meaningfully outside of OpenAI.¹ Anthropic was incorporated as a public benefit corporation (PBC), reflecting its ostensible commitment to certain altruistic goals.

While Claude models are all multilingual and multimodal, their functionality is primarily text-based. Claude lacks some multimodal features offered by competitors such as Google Gemini or OpenAI’s ChatGPT, such as image generation or native audio and video processing. Their robust text-based functionality spans a wide spectrum of both natural language and coding capabilities.

Claude models are fully closed and proprietary to Anthropic, meaning one can only use Claude directly through Anthropic’s AI chatbot offerings—available at Claude.ai or in app form for iOS and Android—or in third-party applications that access the models through Anthropic’s API, such as Microsoft Copilot. Pricing varies based on the specific plan and models being used. Anthropic also offers a free version of Claude, though it’s subject to daily usage limits significantly lower than those of its paid plans.

Beginning with the release of Claude 3 in March of 2024, Claude models are offered in three sizes: Haiku (the smallest), Sonnet and Opus (the largest). Anthropic has not published any information about how these designations translate to quantitative differences in actual model size or architecture.

Each of these model sizes offer their own advantages and disadvantages, but are broadly capable of the same tasks, including tool use and writing code. Their respective tradeoffs in performance, speed and cost when carrying out those tasks make each model size ideal for different kinds of use cases, budgets and AI systems they’re to operate within. 

All Claude models from Claude 3.7 Sonnet onward have been hybrid reasoning models. By default, Claude models directly respond to the user’s input, but they can be toggled to an “extended thinking” mode in which the model will first generate a step-by-step chain of thought (CoT) before providing a final output to the user.

Anthropic’s numbering conventions for new versions of Claude models—or rather, at least versions released to the public—have varied from generation to generation,² and not every numbered “model” is offered in all three of those sizes.

• The first generation of Anthropic’s Claude models included both “Claude” and “Claude Instant” (a faster, cheaper version). Updated versions of the former were released as Claude 1.1, Claude 1.2 and Claude 1.3; the latter was updated as Claude Instant 1.1 and 1.2.

• Claude 2 included only one model size. It was updated once, as Claude 2.1.

• Claude 3 was followed not by Claude 3.1, but Claude 3.5. Furthermore, while Anthropic released Claude 3.5 Sonnet and Claude 3.5 Haiku, it never released a Claude 3.5 Opus.

• That October, Anthropic released an updated version of Sonnet but didn’t update its version number. The “new” model was simply designated “Claude-3.5-Sonnet-20241022,” while the “old” version was retroactively designated “Claude-3.5-Sonnet-20240624.”

• The “new” Claude 3.5 Sonnet was succeeded by Claude 3.7 Sonnet, the company’s first reasoning model. Neither Haiku nor Opus sizes were ever offered in 3.7 versions. No 3.6 version was released.

• In May 2025, Claude 4 reversed the company’s naming conventions with Claude Opus 4 and Claude Sonnet 4. This generation did not include a Haiku version.

• Claude Opus 4 was followed by Claude Opus 4.1. No other model sizes were released as “4.1.”

• In the fall of 2025, Anthropic separately released Claude Sonnet 4.5, followed by Claude Haiku 4.5 and subsequently Claude Opus 4.5. 

Though Anthropic has never explained its model naming process, CEO Dario Amodei has acknowledged that AI model version numbers are often confusing. “I feel like no one’s figured out naming,” he said in a November 2024 podcast appearance. “It’s something we struggle with surprisingly much relative to how trivial it is for the grand science of training the models.” 

It’s worth noting that unclear model versioning is an industry-wide problem. OpenAI, for instance, followed “GPT-4” with “GPT-4-Turbo.” GPT-4-Turbo was succeeded by GPT-4o, which was eventually succeeded by GPT-4.5. The company’s “naming chaos” was exacerbated by its release of GPT-4.1 after GPT-4.5 (which it then retired).

Claude Haiku is Anthropic’s smallest model series, optimized for speed and cost-efficiency. “We ended up with this kind of poetry theme,” Amodei’s explained during the same podcast appearance. “And so what’s a really short poem? It’s a haiku.”

According to Claude’s API docs, Haiku models are optimal for applications that require low-latency, real-time inference and high-volume, straightforward tasks (such as summarization, data extraction or simple machine translation). Haiku models often serve as a component of larger agentic AI workflows. Per-token pricing for Haiku models is roughly one-third the price of Sonnet and one-fifth the price of Opus.³ 

Claude Haiku 4.5 offers a context window of 200K tokens, with a maximum output length of 64K tokens. More information is available in the model’s system card.

“Sonnet is a medium-sized poem,” says Amodei. “And so Sonnet was the middle model.”

Claude Sonnet is Anthropic’s workhorse model, aimed at what the company deems to be an optimal tradeoff of performance and efficiency for most use cases. As Claude’s model overview page advises: “If you’re unsure which model to use, we recommend starting with Claude Sonnet.”

Despite being smaller than the Opus series, there have been extended stretches throughout Claude’s history during which the latest Sonnet model served as Claude’s flagship. For instance, Claude 3.5 Sonnet significantly outperformed Claude 3 Opus—as well as nearly any competitor’s models—on a wide range of benchmarks when it released in June 2024. Sonnet models, including both the “new” 3.5 Sonnet update and (subsequently) Claude 3.7 Sonnet, remained Anthropic’s top performing model until the introduction of Claude 4 Opus almost a year later.

Though Claude Sonnet models share the same maximum context window size as Haiku and Opus by default, Anthropic offers a “beta” context window extension of up to 1M tokens. This beta feature is exclusive to Sonnet 4 and 4.5, available only to certain tiers of organizational accounts and subject to specific long context pricing.

More information is available in the model’s system card.

Claude Opus is the largest and most capable Claude model, named in reference to the concept of a magnum opus—a Latin term often used to refer to one’s greatest, most fully realized work.

Claude Opus is designed for frontier performance on challenging use cases such as complex coding tasks, agentic computer use, automated “deep research,” explaining complicated concepts and working with situations that entail significant ambiguity or brainstorming novel solutions.

That performance comes with the tradeoff of being Anthropic’s most expensive model. However, Claude Opus 4.5 is significantly more cost-effective than Claude Opus 4.1 or Claude Opus 4, both of which are three times more expensive than Opus 4.5 on a per-token basis.³

The most straightforward way to use Claude AI is directly through the Claude app. There are multiple tiers for individual accounts, as well as additional tiers for “Team” and “Enterprise” plans. All plans, including the free tier, provide users with access to Haiku, Sonnet and Opus models, but are subject to differing usage limits.

These usage limits also vary by demand, as well as by use cases and model selection: long, context heavy prompts to Claude Opus will reach a usage limit faster than short prompts to Claude Haiku. Usage is typically more restricted during peak hours in which overall usage approaches Claude’s maximum capacity. As The OC Register reports, changes to and uncertainty regarding usage limits have at times been a source of concern for users.

Individual plan options for the Claude app include:

• Free plan: Users can try Claude for free by providing an email and phone number. Usage limits are reset every 5 hours.⁴

• Pro plan: Priced at USD 20 per month as of this article’s publication date, the Claude Pro plan offers “at least five times the usage per session compared to our free service.”⁵

• Max 5x: Priced at USD 100 per month, the Max 5x plan offers five times the usage of the Pro Plan. It also includes priority access to new features and models. Max plans also provide access to Claude Code, Anthropic’s agentic coding tool.⁶

• Max 20x: Priced at USD 200 per month, the Max 20x plan offers twenty times to usage of the Pro Plan, in addition to the same early access benefits of the Max 5x plan.⁶

All individual plans share the same 200K token context window. When conversations exceed the length of that context window, the Claude app automatically summarizes earlier context to enable the exchange to continue uninterrupted.⁷

Team plans, available in Standard and Premium versions, offer the benefits of the Pro plan plus additional organization features such as managing access through single sign-on (SSO), role-based permissions and third-party application integrations.⁸

The Enterprise plan, also available in Standard and Premium versions, offers the features of the Pro plan with higher usage limits, additional enterprise tool integrations and an enhanced context window.⁹

Claude’s development is guided by Claude’s Constitution, a “foundational document that both expresses and shapes who Claude is.” Though the Constitution is freely available for anyone to read—its full text is available here—Anthropic notes that “the constitution is written primarily for Claude.” 

The function of Claude’s Constitution is to quite literally inform how the model responds to specific scenarios and implements Anthropic’s core values, during both training and inference. For instance, the Constitution directly influences the creation of synthetic data used to train Claude models, as well as the training of supplementary models used for AI alignment. Anthropic refers to its overall approach as Constitutional AI.