Anthropic describes itself as a public benefit corporation, with a board elected by stockholders and a Long-Term Benefit Trust that also has trustees. Its stated mission is "to ensure the world safely makes the transition through transformative AI" (Source 1).

Its research pages group safety work into three areas: alignment (studying risks of AI models and keeping future systems beneficial, truthful and harmless), interpretability (studying how large language models work internally) and societal impacts (how AI affects people individually and collectively) (Source 1).

Claude is the name of the model family; the individual models are named by tier and version, such as claude-sonnet-5-5. Anthropic's release notes list these recent launches (Source 3):

Two points about the family:

• Version numbers move per tier. Newer versions of one tier can arrive without the others (for example Opus 5.5 appears before Haiku 5.5). "The latest Claude" therefore means "the latest in each tier".
• Older models are retired on a schedule. The release notes record retirements, and the models overview lists a retirement commitment for each current model of "not sooner than" September–October 2027. Pin a full model ID if you need a stable target (Source 2).

Anthropic's models overview describes the four current models and gives each a price and a relative latency (Source 2). All four list a 1M-token context window, 128K max output and a June 2026 knowledge cutoff, so the differences are in capability, speed and price.

Suppose one agent turn sends 200,000 input tokens (the conversation and file contents) and generates 10,000 output tokens. At the list prices above, on Sonnet 5.5 that is 0.2 × $2 + 0.01 × $10 = $0.50; on Opus 5.5 it is 0.2 × $4 + 0.01 × $20 = $1.00. Real bills differ, because the overview also describes prompt-cache reads at a discount, a 50 percent discount for batch requests, and extra charges for long prompts on Haiku 5.5. Agent sessions make many such turns, which is why model choice and context size dominate cost.

Anthropic's guidance offers two starting strategies (Source 4):

The same page says most workloads should begin with Opus 5.5, and that tuning effort — a setting that trades intelligence for latency and cost inside one model — is often a better lever than switching models.

Claude Code's model documentation describes aliases and levers (Source 7): the opus, sonnet, haiku and fable aliases pick the latest of each tier; opusplan uses Opus while planning and Sonnet while executing; /model changes the model in a session and claude --model at start-up; /effort and --effort set the effort level. The documented default is Opus 5.5 on Pro, Max, Team, Enterprise and the Anthropic API, and Fable models are never the default — you select them explicitly. An administrator's organisation default overrides these.

The documented reasons are about positioning and tooling, not scores:

• Coding is a stated focus. Anthropic describes Opus 5.5 as "for long-running agentic coding" and lists agentic coding among the uses for Fable 5.1 and Sonnet 5.5 (Source 2, Source 4).
• Large working memory. Every current tier lists a 1M-token context window, which suits reading many files (with the context-rot caveat from Chapter 1).
• Vertically integrated tooling. The company that trains the models also ships the harness, Claude Code, plus the documentation for pairing them.

The same models are offered through different products. They are easy to mix up because they share a name.

A useful mental model: the API is the engine room, claude.ai is a chat product built on it, and Claude Code is an agent harness for software work. Claude Code sessions and API calls both consume model tokens; its documentation states that Claude Code charges by API token consumption, with subscription plans priced separately (Source 6).