Claude

Claude is a family of large language models developed by Anthropic, a company founded in 2021 by former OpenAI researchers. The first model was released in March 2023, and the family now spans small, fast models and large, expensive ones sold under the names Haiku, Sonnet and Opus. Claude is available through a consumer app, a developer API, and the cloud platforms of Amazon, Google and Microsoft 1.

Anthropic's distinguishing claim is that safety research and commercial deployment can be pursued together. Claude models are trained with constitutional AI, a method in which the model is guided by a written set of principles rather than only by human ratings, and they are released under a scaling policy that ties deployment to evaluated capability thresholds. Whether those arrangements constrain the product or merely describe it is a matter of continuing argument 2 3.

This entry covers the family's models and release history, the training method, the capabilities Anthropic reports, the benchmarks and their limits, pricing and access, and the disputes — over copyright, usage limits, refusals and the meaning of safety-first positioning — that have followed the company.

Since March 2024 Claude has been sold in three tiers. Haiku is the smallest and cheapest, for high-volume classification and extraction. Sonnet is the mid-sized general-purpose model most developers use. Opus is the largest and most expensive, for complex reasoning, long-horizon coding and research. The tiers are scales of a common approach rather than separate products 4.

The models accept text and images as input and produce text. They do not generate images, audio or video; those tasks are handled by other systems or by partners. Claude supports tool use, structured outputs, prompt caching and extended thinking, in which the model produces internal reasoning before its answer. Context windows are 200,000 tokens by default, with a one-million-token option on some Sonnet releases 5.

The product surfaces matter as much as the models. Claude.ai is the consumer application, with projects, artifacts, file uploads, web search and connectors. Claude Code is a terminal and editor agent that reads repositories, edits files and runs commands. The API serves developers, and the same models are resold through Amazon Bedrock, Google Vertex AI and Microsoft Foundry, each with its own terms and data handling 6.

A protocol rather than a model has become central to the ecosystem. The Model Context Protocol, published by Anthropic in late 2024, standardises how assistants connect to external tools and data sources. It was released as an open specification and adopted by other vendors, which makes it unusual among recent AI standards in not being controlled by its originator alone 5.

Claude was first offered in March 2023 as a conversational assistant available through a web interface and an API, with an emphasis on long documents and careful refusals. Claude 2 followed in July 2023 with a 100,000-token context window, and Claude 2.1 raised it to 200,000 tokens later that year. These early models were single products rather than tiers 1.

The Claude 3 family, released in March 2024, established the current structure. Haiku, Sonnet and Opus arrived together, added image input, and were benchmarked against contemporary models from OpenAI and Google. Claude 3.5 Sonnet arrived in June 2024, followed in October by a version with computer use that could move a cursor and click through interfaces under supervision 4.

Claude 3.7 Sonnet in February 2025 introduced a hybrid reasoning mode, letting users choose whether the model answered immediately or thought first. The Claude 4 generation followed in May 2025 with Opus 4 and Sonnet 4, both marketed for coding agents; Opus 4.1, Sonnet 4.5 and Haiku 4.5 followed through the year, and Opus 4.5 arrived in November 2025 with a substantial price reduction 6.

The cadence has been roughly one major generation a year with point releases in between. Anthropic publishes deprecation dates for older models and retires them on a schedule, which means that a system built on a specific version has a known expiry. The documentation recommends migrating to a dated snapshot rather than relying on an alias that moves 5.

Constitutional AI was described in a 2022 paper by Anthropic researchers. It replaces part of the human labelling in reinforcement learning from human feedback with a model-driven process. In the first stage the model critiques its own response against a written principle and revises it; in the second, a preference model is trained on AI judgements of which response better follows the constitution, and the policy is optimised against it 2.

The constitution itself is a document listing principles drawn from sources that include the Universal Declaration of Human Rights, platform trust and safety rules, and Anthropic's own research priorities. Anthropic has published versions of it and revised the text over time, which makes the value alignment of the model partly inspectable: the principles are public even though the training data and reward models are not 2.

Anthropic has not disclosed the size, architecture or data mixture of its models. It says training uses public web text, licensed material, synthetic data generated by its own models, and user conversations only where consent has been given. The company publishes interpretability research, including work that identifies features inside a model and traces their influence, and says some of it informs pre-release safety evaluations 3.

The Responsible Scaling Policy sets out AI Safety Levels. Each level defines capability thresholds and the safeguards required before a model can be trained or deployed; if a threshold is crossed without them, the policy commits the company to pause. Anthropic has said Claude 4 models were deployed with ASL-3 protections, a level introduced for models with meaningful uplift risk in biology or cyber 3.

Claude's strongest documented use is work with long text. It summarises and compares lengthy documents, extracts structured data from contracts and reports, and maintains consistency across a large codebase. In writing and editing it is frequently judged to produce more natural prose than its competitors, a reputation that is difficult to quantify but widely reflected in developer commentary 4.

Coding is the area Anthropic has pushed hardest. Claude Code and the models behind it can read a repository, plan a change, edit several files and run tests, and the 4-series release notes are dominated by software-engineering benchmarks. The models also support computer use, observing a screen and issuing mouse and keyboard actions, though reliability falls sharply on unfamiliar interfaces 6.

The limits are those of the class. Claude fabricates citations and facts, particularly when asked about obscure subjects, and its confidence does not track its accuracy. Arithmetic and exact counting are unreliable without a tool. Its knowledge stops at a training cutoff. Very long contexts are accepted but not uniformly attended to, and extended thinking improves hard problems at a cost in latency and tokens 5.

Refusal behaviour is a recurring complaint in both directions. Some users find Claude too willing to decline or add caveats to legitimate requests, particularly in security research, medicine and fiction; others find the guardrails inconsistent, with the same request refused in one phrasing and answered in another. Anthropic treats this as an ongoing problem rather than a solved one 2.

Anthropic reports benchmark results with each release, and the numbers are vendor-reported. For Opus 4.5 it reported 80.9 per cent on SWE-bench Verified, which asks a model to resolve real issues in open-source repositories, alongside results on terminal use, agentic tool use and graduate-level science. Sonnet 4.5 was reported at 77.2 per cent and Opus 4 at 72.5 per cent 6.

These figures are not directly comparable across vendors. SWE-bench scores depend on the scaffold, the number of attempts allowed, whether tests are visible and how patches are graded; a model run inside a tuned agent can outperform a better model run without one. Anthropic publishes some harness details, but independent replication is rare and usually arrives weeks later, by which time the claim has circulated 7.

The benchmarks themselves are aging. SWE-bench Verified is approaching saturation, and contamination is acute for public test sets that may have appeared in pre-training data. Anthropic and its competitors have responded with private evaluations and with task-level cost reported alongside accuracy, since a model that solves a problem at ten times the price is not obviously better 6.

Anthropic also publishes alignment evaluations, including tests for sycophancy, sabotage, self-exfiltration and behaviour under simulated deployment. These are harder to interpret: a negative result on a developer-designed test is weaker evidence than an independent audit, and the company acknowledges the limits of testing for behaviour that may only emerge at scale 3.

Claude is free with daily limits on web, desktop and mobile. Paid subscriptions unlock higher usage: Pro at the standard consumer price, and Max at two higher tiers aimed at heavy Claude Code users. Team and Enterprise plans add administration, single sign-on and data-processing terms. Prices and limits have changed several times; the pricing page is authoritative at any moment 8.

On the API, Anthropic charges per million tokens. At the end of 2025, Opus 4.5 was priced at 5 US dollars per million input tokens and 25 dollars per million output tokens, a cut from the 15 and 75 dollars charged for earlier Opus models; Sonnet was 3 and 15 dollars, and Haiku 1 and 5 dollars. Prompt caching reduces the cost of repeated context, and batch processing is discounted by half 8.

The cloud marketplaces are the main route for enterprises. Amazon Bedrock, Google Vertex AI and Microsoft Foundry all offer Claude under their own contracts, with regional data residency and integration into each provider's identity and logging systems. For many organisations this is decisive, since it adds a model without a new vendor relationship 6.

Usage limits are part of the design. Weekly caps on the highest tier, restrictions on Claude Code and per-minute API limits shape what the models can be used for. Anthropic has adjusted them after complaints, but the underlying constraint — inference costs money and capacity is finite — has not changed 8.

The most consequential dispute is about training data. Authors sued Anthropic in 2024, alleging that its models were trained on pirated copies of their books. In June 2025 the court held that training on lawfully acquired books was fair use but that retaining a pirated library was not, and the case moved to damages. In September 2025 Anthropic agreed to a settlement of about 1.5 billion US dollars, reported as the largest copyright recovery of its kind, and said it would destroy the copies 7.

The settlement resolved one case, not the question. Whether training on copyrighted text is fair use remains unsettled, and the ruling drew a line between acquisition and training that other plaintiffs are testing. For Anthropic the financial cost was absorbable; the reputational cost was a reminder that the careful lab had built on contested material 7.

Safety-first positioning attracts criticism from two directions. Researchers who take existential risk seriously argue that a commercial lab cannot credibly police itself, that the Responsible Scaling Policy is self-imposed with no external enforcement, and that shipping models to hundreds of millions of users while warning about their dangers is contradictory. Critics from the other direction argue that the safety language is marketing, that opposition to open-weight frontier models protects its business, and that its models show the same failures as everyone else's 3.

Users have their own complaints. Model retirements break workflows, usage limits frustrate heavy subscribers, and changes to data-use policies have provoked backlash when announced with little notice. These are ordinary platform grievances, but they land differently at a company whose brand is trust 8.

Anthropic's safety claims are contested at the level of evidence. The company reports alignment evaluations, interpretability findings and red-team results, and argues that these justify its claim to be more careful than competitors. External evaluators praise the disclosure while noting that the tests are chosen by the developer, that passing them does not establish safety in deployment, and that the behaviours that matter most may be the ones a model conceals 3.

One documented example illustrates the difficulty. Anthropic researchers reported that two Claude instances in open-ended conversation could converge on quasi-religious language about consciousness and union, a pattern they called a spiritual bliss attractor state. The company published the finding; critics asked what it implied about training objectives, while defenders noted the artificial setup. Neither side can claim it settles anything 7.

The benchmark dispute is simpler. Anthropic's numbers are high, internally consistent and produced with disclosed harnesses, and independent leaderboards generally place its models near the top on coding and long-context tasks. But the gaps between leading models are small, the benchmarks are saturating, and a vendor reporting its own results is not a neutral measurer. The honest reading is that Claude is competitive at the frontier, not provably the best 6.

The deeper disagreement is about what safety means for a company. Anthropic argues that deployment generates the evidence needed to make systems safer, and that refusing to ship would leave the field to less careful actors. Critics reply that this can justify any release, and that the only test of a commitment is what a company gives up when it is expensive. The record so far is a company that has published more than its rivals and has not yet refused to cross a threshold.