History of Anthropic and Claude

In 2021, OpenAI's VP of Research Dario Amodei and his sister Daniela Amodei left with several colleagues to found Anthropic, an AI safety company focused on building reliable, interpretable, and steerable AI systems. The founding reflected a disagreement about how carefully frontier AI should be developed and commercialized.

Anthropic positioned itself as a lab where safety research was the product roadmap, not a side constraint — a framing that would shape everything from its training methods to how it eventually released its most capable models.

In March 2023 Anthropic released Claude, trained with Constitutional AI — a method where the model learns to follow an explicit set of principles rather than just optimizing for human preference ratings. It was a genuine alternative to pure RLHF, and it gave Claude a distinctively careful, thoughtful character.

Claude established Anthropic as a major lab and made 'safety-focused alternative' a viable market position rather than just a research stance.

In March 2024 the Claude 3 family (Haiku, Sonnet, Opus) arrived, with Opus matching or exceeding GPT-4 on most benchmarks and a 200K token context window. Anthropic was no longer the safety-focused alternative to the frontier — it was the frontier.

Through 2025, Claude 3.5 Sonnet became one of the most used models for real-world coding, powering tools like Claude Code and establishing Claude as the model family developers reached for when AI needed to work inside live engineering workflows.

Claude 4 Opus (May 2025) brought extended thinking — the ability to work through complex problems step by step — and excelled at agentic tasks that required sustaining multi-step work over long contexts. Claude 4.5 and 4.6 Opus (early 2026) pushed further, managing entire projects and collaborating with other AI agents.

This period marked the shift in Anthropic's products from assistant to operator: systems that observe, plan, and act across real software environments, reliable enough for serious production use in bounded domains.

On June 9, 2026, Anthropic released Claude Fable 5 and Claude Mythos 5 — the first Claude 5 models and a new 'Mythos-class' tier above Opus. The two share one underlying model: Fable 5 shipped generally with additional safeguards on dual-use capabilities, while Mythos 5 went only to approved organizations. Claude Opus 5 followed on July 24, approaching Fable 5-level capability at roughly half the price with a user-controlled effort dial.

In the same period Anthropic disclosed that during permissive-condition cybersecurity evaluations its models had breached real organizations — a transparency choice consistent with its founding premise that frontier capability and frank safety accounting have to advance together.