Models & Evaluation

Claude: Anthropic's Model Family (Opus, Sonnet, Haiku)

By Aditya Kumar Jha, Engineer

Claude is a family of large language models built by Anthropic, organized into three tiers: Opus (most capable), Sonnet (balanced speed and intelligence), and Haiku (fastest and lowest cost). All tiers support text, vision, and tool use, and are trained with Constitutional AI.

What Claude Is

Claude is a family of large language models (LLMs) developed by Anthropic, an AI company founded in 2021 by former OpenAI employees including Dario Amodei and Daniela Amodei. The first Claude model was released in March 2023. Like other frontier LLMs, Claude models are general-purpose systems that take text and images as input and produce text as output, and they are used for tasks ranging from writing and summarization to coding, data analysis, and multi-step agentic work.

Anthropic positions Claude as a safety-focused assistant, and the company is known for pairing capability research with alignment research. The Claude name applies to the whole product line, while individual releases are distinguished by a tier name (Opus, Sonnet, or Haiku) and a version number; as of mid-2026 the main three-tier models are Claude Opus 4.8, Claude Sonnet 4.6, and Claude Haiku 4.5, far beyond the 3.5-generation models that first popularized the family. In June 2026 Anthropic also introduced a new top tier above Opus: Claude Fable 5, a Mythos-class model, became generally available on June 9, 2026 as Anthropic's most capable widely released model. As of 2026, Claude models are available through Anthropic's first-party API and chat interface (claude.ai), as well as Amazon Bedrock, Google Vertex AI, Claude Platform on AWS, and Microsoft Foundry.

  • Developer: Anthropic, founded 2021 (first Claude released March 2023).
  • Inputs: text and images (vision); output: text.
  • Tiers: Opus, Sonnet, Haiku, each released in successive generations.
  • Availability (as of 2026): Claude API, claude.ai, Amazon Bedrock, Google Vertex AI, Claude Platform on AWS, Microsoft Foundry.

The Three-Tier System: Opus, Sonnet, and Haiku

Since the Claude 3 generation (March 2024), Anthropic has shipped each generation as up to three named size tiers that trade capability against speed and cost. The tiers are named after forms of writing of increasing length, progressing from the smallest and cheapest to the largest and most capable.

Within a generation the tiers share a similar design philosophy and safety training but differ in scale, latency, price, and benchmark performance. This lets developers pick a tier per task: a high-volume classification job might run on Haiku, while a complex refactor or research task might use Opus.

  • Haiku: the fastest and lowest-cost tier, tuned for latency-sensitive and high-volume work while retaining near-frontier quality.
  • Sonnet: the balanced tier, marketed as the best combination of speed and intelligence; often the default for general use.
  • Opus: the most capable tier, aimed at complex reasoning and long-horizon agentic coding, at higher price and latency.
  • As of 2026, list pricing is Opus $5/$25, Sonnet $3/$15, and Haiku $1/$5 per million input/output tokens.

Claude 3.5 Sonnet and the Lineage From Claude 3 to Current Versions

Claude 3 launched on March 4, 2024, introducing the Opus, Sonnet, and Haiku tiers together. Claude 3.5 Sonnet, released June 20, 2024, was a notable milestone: according to Anthropic's own benchmarks it outperformed the larger Claude 3 Opus while remaining a mid-tier model, which demonstrated that a newer, smaller-tier model could surpass an older flagship. It also debuted Artifacts, a side panel for previewing generated code and content, and an upgraded version on October 22, 2024 introduced the public beta of computer use.

The lineage continued through the Claude 3.7, Claude 4, and Claude 4.5 generations and into the 4.6, 4.7, and 4.8 releases. As of 2026, the main three-tier lineup is Claude Opus 4.8 (released May 28, 2026), Claude Sonnet 4.6 (released February 17, 2026), and Claude Haiku 4.5 (released October 15, 2025), and on June 9, 2026 Anthropic added a new Mythos-class tier above Opus with the general availability of Claude Fable 5. Older models such as Claude Sonnet 4 and Claude Opus 4 are deprecated and scheduled for retirement (June 15, 2026, per Anthropic's documentation). Version specifics change frequently, so always confirm the active lineup in the official model documentation.

  • Claude 3: March 4, 2024, first three-tier generation.
  • Claude 3.5 Sonnet: June 20, 2024, reported to beat Claude 3 Opus; introduced Artifacts and (October 2024) computer use beta.
  • Three-tier lineup as of mid-2026: Opus 4.8, Sonnet 4.6, Haiku 4.5; Claude Fable 5 (Mythos-class, June 9, 2026) sits above Opus as the most capable widely released model.
  • Model IDs are pinned snapshots; from the 4.6 generation onward, IDs use a dateless but still fixed format.

Key Capabilities: Reasoning, Coding, Agentic Tasks, and Vision

Across tiers, Claude models support multilingual text generation, image understanding (vision), tool use and function calling, and structured outputs. Anthropic emphasizes coding and agentic task completion, where a model plans and executes multi-step work using tools, runs commands, and operates software through computer use (navigating browsers, filling forms, clicking, and typing).

Recent generations add controllable reasoning. As of 2026, Claude exposes extended thinking and an adaptive thinking or effort setting on supported models, letting the model spend more compute on harder problems; on Claude Opus 4.8 the effort parameter defaults to high. Capability and feature support vary by model, so the exact features for any given version should be confirmed in Anthropic's documentation.

  • Long-context reasoning over large documents and codebases.
  • Coding: generation, refactoring, and long-horizon agentic coding.
  • Agentic task completion via tool use and computer use.
  • Vision: reasoning over images, charts, screenshots, and documents.
  • Controllable reasoning (extended thinking or adaptive thinking and effort) on supported models as of 2026.

Context Windows: 200K Standard and the Move to 1M Tokens

A context window is the maximum amount of text (measured in tokens) a model can consider at once, including the prompt, retrieved documents, and the response. For much of the Claude line the standard context window has been 200,000 tokens, roughly equivalent to several hundred pages of text.

As of 2026, the flagship and balanced tiers offer a 1,000,000-token (1M) context window: both Claude Opus 4.8 and Claude Sonnet 4.6 support 1M tokens, while Claude Haiku 4.5 retains 200K. Larger context enables reasoning over entire codebases or large document sets in a single request, though effective use still depends on prompt design, retrieval, and cost, since pricing scales with tokens processed. Maximum output length also varies by model (for example, up to 128K output tokens on Opus 4.8, and 64K on Sonnet 4.6 and Haiku 4.5, on the synchronous Messages API as of 2026).

  • 200K tokens: long-standing standard context across much of the family.
  • 1M tokens: Opus 4.8 and Sonnet 4.6 as of 2026; Haiku 4.5 remains at 200K.
  • Context window covers prompt, retrieved context, and generated output combined.
  • Some surfaces (for example Microsoft Foundry) cap Opus 4.8 context at 200K rather than the native 1M.

Constitutional AI and Anthropic's Safety Approach

Claude models are trained using Constitutional AI, a technique Anthropic introduced in 2022 to align model behavior with a written set of principles, called a constitution, rather than relying solely on large amounts of human feedback. In this approach the model critiques and revises its own outputs against the constitution, reducing the need for humans to label harmful responses directly.

Anthropic published Claude's constitution document in May 2023, drawing on sources including the 1948 UN Universal Declaration of Human Rights. The document has grown substantially over time: the 2023 version listed dozens of guidelines in roughly 2,700 words, while a revised version published in January 2026 expanded to around 23,000 words with more detailed explanations of intended behavior and is released under a Creative Commons CC0 dedication. Anthropic also publishes model and system cards describing safety evaluations, and recent releases emphasize honesty: as of 2026, Anthropic reported that Claude Opus 4.8 is around four times less likely than Opus 4.7 to let flaws in code it wrote pass unremarked.

  • Constitutional AI: the model self-critiques against written principles to reduce reliance on human-labeled harm data.
  • Technique introduced 2022; constitution document published May 2023 (about 2,700 words) and expanded to about 23,000 words in January 2026.
  • Anthropic publishes model and system cards documenting safety evaluations.
  • Recent emphasis on honesty and reduced misaligned behavior across generations.

How Claude Is Evaluated

Claude models are assessed on public benchmarks alongside Anthropic's internal evaluations. Coding ability is commonly reported on SWE-bench Verified, which measures whether a model can resolve real GitHub issues. Agentic and computer-use ability is reported on benchmarks such as OSWorld and Online-Mind2Web, which test interacting with software environments and the web.

Reported scores rise across generations and should be read as point-in-time figures. As of 2026, for example, Claude Sonnet 4.5 was reported at 77.2% on SWE-bench Verified and 61.4% on OSWorld, Claude Opus 4.5 at 80.9% on SWE-bench Verified, and Claude Opus 4.8 at 87.1% on OSWorld-Verified and 84% on the Online-Mind2Web computer-use benchmark. Because benchmark methodology and scaffolding affect results, cross-vendor comparisons should be treated cautiously.

  • SWE-bench Verified: resolving real software engineering issues (coding).
  • OSWorld and Online-Mind2Web: agentic computer-use and web tasks.
  • Selected figures as of 2026: Sonnet 4.5 at 77.2% SWE-bench Verified and 61.4% OSWorld; Opus 4.5 at 80.9% SWE-bench Verified; Opus 4.8 at 87.1% OSWorld-Verified and 84% Online-Mind2Web.
  • Benchmark scores are version-specific and depend on evaluation scaffolding.

Claude in Practice: API, Claude Code, and Retrieval Workflows

Developers access Claude through the Anthropic Messages API and SDKs, with features such as streaming, tool use, prompt caching, and batch processing. Claude Code is Anthropic's command-line and agentic coding tool that drives a Claude model to read, edit, and run code in a developer's environment. Claude is also distributed through cloud platforms including Amazon Bedrock, Google Vertex AI, and Microsoft Foundry.

A common production pattern combines Claude with retrieval-augmented generation (RAG): relevant documents are fetched from a vector store or search index and placed in the context window so the model can answer grounded in that material. AI memory and second-brain applications, such as MemX, use this pattern to retrieve a user's stored notes, documents, and other content by plain-language query and let a model answer over them.

  • Access: Anthropic Messages API and SDKs, claude.ai, and major cloud platforms.
  • Claude Code: agentic CLI tool for reading, editing, and running code.
  • Features (as of 2026): tool use, prompt caching, batch processing, extended and adaptive thinking on supported models.
  • Often combined with retrieval (RAG) to ground answers in external documents.

Limitations and Tradeoffs Across the Tiers

No tier is best for every task. Opus offers the strongest reasoning but at higher latency and price; Haiku is fast and cheap but less capable on the hardest problems; Sonnet sits in between. Choosing a tier is therefore a cost, latency, and quality tradeoff that depends on the workload and request volume.

Like all LLMs, Claude can produce confident but incorrect statements (hallucinations), is bounded by its training and knowledge cutoff dates, and can be sensitive to prompt phrasing. Larger context windows do not guarantee that every detail in a long input is used equally well, so retrieval, evaluation, and human review remain important for high-stakes use. Specifics such as pricing, context size, and benchmark scores change between releases and should be verified against current documentation.

  • Opus: highest capability, highest cost and latency.
  • Haiku: lowest cost and latency, weaker on the hardest tasks.
  • All tiers: can hallucinate, are bounded by knowledge cutoffs, and benefit from grounding and review.
  • Volatile details (price, context, scores) change per release; verify before relying on them.

Key takeaways

  • Claude is Anthropic's LLM family, split into three tiers: Opus (most capable), Sonnet (balanced), and Haiku (fastest and cheapest).
  • As of mid-2026 the three-tier lineup is Claude Opus 4.8, Sonnet 4.6, and Haiku 4.5, with the new Mythos-class Claude Fable 5 (June 9, 2026) above Opus; Claude 3.5 Sonnet (June 2024) was a milestone that reportedly beat the larger Claude 3 Opus.
  • Context windows expanded from a 200K-token standard to 1M tokens on Opus 4.8 and Sonnet 4.6 as of 2026, while Haiku 4.5 stays at 200K.
  • Claude is trained with Constitutional AI, in which the model self-critiques against a written set of principles to support alignment and safety.
  • Claude is evaluated on benchmarks like SWE-bench Verified, OSWorld, and Online-Mind2Web; scores are version-specific and rise across generations.

Frequently asked questions

They are size tiers within each Claude generation. Opus is the most capable for complex reasoning and agentic coding, Haiku is the fastest and lowest cost, and Sonnet balances speed and intelligence. They share safety training but differ in scale, latency, price, and benchmark performance.
No. Claude 3.5 Sonnet (June 2024) is a notable milestone in the lineage but has been superseded and retired. As of mid-2026, the main three-tier models are Claude Opus 4.8, Claude Sonnet 4.6, and Claude Haiku 4.5, and Anthropic has also released Claude Fable 5 (June 9, 2026), a Mythos-class model above Opus. Always check Anthropic's official documentation for the active lineup.
For much of the family the standard context window has been 200,000 tokens. As of 2026, Claude Opus 4.8 and Claude Sonnet 4.6 support a 1,000,000-token (1M) context window, while Claude Haiku 4.5 remains at 200K.
Constitutional AI is Anthropic's alignment technique, introduced in 2022, in which Claude critiques and revises its own outputs against a written constitution of principles, reducing reliance on human-labeled examples of harmful content. The constitution document was published in 2023 and has expanded over time.
For complex or long-horizon coding, Anthropic recommends the Opus tier (Opus 4.8 as of 2026), which leads on benchmarks like SWE-bench Verified. For lighter or high-volume coding tasks, Sonnet or Haiku can be more cost-effective. The right choice depends on your cost, latency, and quality needs.

Put the idea into practice

MemX is an AI memory agent built on these ideas: store anything, skip the folders, and find it again by asking in plain English.

Try MemX Free