In brief
Anthropic is an American AI laboratory founded in 2021 by former OpenAI employees who believed that the safety of AI systems must take precedence over speed to market. In five years, the company has raised over $67 billion, reached a valuation of $380 billion, and placed its Claude family of models among the most capable on the market. The stated paradox: Anthropic believes transformative AI is inevitable, and prefers to be at the frontier rather than cede that ground to actors less concerned with its risks.
In short: imagine a pharmaceutical lab founded by researchers convinced that a powerful drug will reach the market — and who prefer to develop it themselves to better contain its side effects, rather than wait for others to do so less cautiously. That is Anthropic’s stance applied to artificial intelligence: participating in the race while claiming a stricter safety specification than the competition.
Identity card
| Data | Value |
|---|---|
| Founded | 2021 |
| Headquarters | San Francisco, California |
| Headcount | Undisclosed (several hundred engineers and researchers) |
| Status | Public Benefit Corporation |
| Founders | Dario Amodei (CEO), Daniela Amodei (President), Chris Olah, and five other former OpenAI researchers |
| Valuation | $380B (February 2026, post-Series G) |
History
It all began with a disagreement. In December 2020, Dario Amodei — then VP of Research at OpenAI — and a group of close colleagues left the company. The stated reason: a deep divergence over priorities, with OpenAI accelerating toward ever-larger models without, in their view, taking the resulting risks seriously enough.
Anthropic was officially incorporated in 2021, with a first raise of $124 million. The chosen status — Public Benefit Corporation — was deliberate: it legally enshrines the safety mission in the company’s articles of incorporation, alongside the obligation to deliver financial results.
The scale-up was rapid. In April 2022, a Series B of $580 million (including $500 million from FTX, which would become problematic after Sam Bankman-Fried’s bankruptcy in November 2022) consolidated the company’s resources. In March 2023, Claude 1 launched publicly. Google invested as early as May 2023 via the Series C, followed by Amazon in September 2023 with $4 billion. These two cloud giants became strategic partners: Claude is available on Google Cloud Vertex AI and Amazon Bedrock.
Funding rounds followed at a pace reflecting the intensity of the frontier model race: Series E at a $61.5 billion valuation in March 2025, Series F at $183 billion in September 2025, and Series G at $380 billion in February 2026 — $30 billion raised in a single round, the second largest in venture capital history. Annualized revenues reached $19 billion in March 2026.
In short: Anthropic went from $124 million in 2021 to $380 billion in valuation five years later. That is a 3,000× multiplier of the price paid by a new investor — a pace only a handful of historic tech companies have matched. This inflation reflects both the perceived promise of frontier LLMs and Anthropic’s dependence on outside capital to fund training and infrastructure.
Models and products
The Claude family has evolved through several generations, each bringing significant changes.
Claude 1 (March 2023): first public model, 9,000-token context window, positioned around conversational safety.
Claude 2 (July 2023): major leap with the context window extended to 100,000 tokens — a landmark first — and improved reasoning capabilities. Version 2.1 (November 2023) doubled this threshold again to 200,000 tokens.
Claude 3 (March 2024): introduction of the three-tier structure that still defines the lineup today — Opus (most powerful), Sonnet (balanced), Haiku (fastest and most economical). Vision capabilities added (image analysis). At launch, Opus 3 outperformed GPT-4 on several benchmarks.
Claude 3.5 Sonnet (June 2024): performance exceeding Claude 3 Opus at a lower cost, with notable gains in coding. The v2 release (October 2024) introduced Computer Use, an unprecedented capability: the model can control a graphical interface (mouse, keyboard, screen) to complete autonomous tasks.
Claude 4 (May 2025): launch of Opus 4 for complex tasks and multi-agent orchestration, simultaneously with the general availability of Claude Code.
Opus 4.6 and Sonnet 4.6 (February 2026): one-million-token context window, 14.5-hour task completion horizon, advanced multi-agent orchestration capabilities. Sonnet 4.6 keeps the previous generation’s pricing tier for performance approaching Opus on analytical tasks.
Generation 5 (summer 2026): the line-up is renamed and recut. Sonnet 5 (system card dated 30 June 2026) becomes the default model on the Free and Pro plans and drops to $2 / $10 per million tokens, against $3 / $15 for Sonnet 4.6. Opus 5 stays at $5 / $25. A fourth tier appears above Opus: Fable 5.1, at $10 / $50, with adaptive reasoning always on.
As of 6 September 2026, the current line-up documented by Anthropic is as follows:
| Model | Price (input / output, per M tokens) | Context | Max output | Retirement no sooner than |
|---|---|---|---|---|
| Claude Fable 5.1 | $10 / $50 | 1M | 128K | 1 September 2027 |
| Claude Opus 5 | $5 / $25 | 1M | 128K | 24 July 2027 |
| Claude Sonnet 5 | $2 / $10 | 1M | 128K | 30 June 2027 |
| Claude Haiku 4.5 | $1 / $5 | 200K | 64K | 15 October 2026 |
Seven earlier models remain available as legacy (Fable 5, Opus 4.5 through 4.8, Sonnet 4.5 and 4.6). Anthropic publishes a committed retirement date for each — something neither OpenAI nor Google does. It is one of the few contractual stability guarantees on the market.
In plain terms: between February and the summer of 2026, the price of the mid-tier model fell by a third while its context window stayed at one million tokens. The underlying trend of 2026 is not raw capability, it is falling price at constant performance — plus a premium tier created above Opus to capture what is left of the budgets with no ceiling.
On the product side, Anthropic addresses several audiences:
- Claude.ai: the web and mobile consumer interface, the functional equivalent of ChatGPT.
- Anthropic API: direct model access for developers.
- Claude Code: a command-line tool for delegating development tasks to Claude from a terminal.
- Claude Cowork (research preview, January 2026): a graphical interface for Claude Code aimed at non-technical users.
In short: Anthropic’s product strategy reads like a cascade of tiers — Claude.ai to open up the general public, the API for developers, Claude Code for seasoned engineers, Claude Cowork to capture the “non-tech wants agents” segment. Each tier reuses the same engine (the Claude models) and dresses the interface differently depending on the technical tolerance of the targeted audience.
Positioning
Constitutional AI: Anthropic’s technical approach
Anthropic’s most-cited differentiator is its alignment approach, formalized under the name Constitutional AI (CAI). Published in December 2022 (Bai et al., arXiv:2212.08073), the method works as follows: a set of ethical principles — a “constitution” — is defined. The model learns to self-critique according to these principles, and an evaluator model generates preference data for the reinforcement phase. This is RLAIF (Reinforcement Learning from AI Feedback), as opposed to classic RLHF which relies on human annotators.
The claimed advantage: scalability, consistency, and reduced human cost. The limit critics point to: this approach distances the feedback loop from humans, raising questions about the democratic legitimacy of the encoded values.
In January 2026, Anthropic published a new constitution for Claude, moving from a rules-based approach to a reasoning-based approach: the principles are made explicit, and a four-level hierarchy is established (safety > ethics > policy compliance > helpfulness). The document formally acknowledges the possibility of AI consciousness and moral status — a first in an official document from a major laboratory.
Strengths and limitations
Among the genuine strengths: extended context windows (Anthropic is often ahead on this), agentic capabilities for long and autonomous tasks, and a reputation for models less prone to over-refusal than some competitors.
Among the limitations: a thinner third-party ecosystem than OpenAI’s, no hardware presence (unlike Google or Meta), and dependence on partners’ cloud infrastructure. The safety-first stance is also sometimes perceived as a rationalization of active participation in the race for the most powerful models.
When to choose Anthropic over a competitor
| Context | Recommendation | Why |
|---|---|---|
| Long-document analysis (reports, codebase, legal files) | Claude (Opus or Sonnet) | Standard 200k-token window, 1M-token beta on 4.6 — few players match this capacity in production. |
| Long agentic workflow (multi-hour orchestration, interface control) | Claude Opus 4.x | Computer Use since 3.5, autonomous operation documented over 30h on Sonnet 4.5, more mature market than OpenAI o3 on this dimension. |
| Sensitive application (legal, medical, customer support) | Claude with Constitutional AI | Explicit and auditable alignment policy, calibrated refusal rate, traceability of invoked principles. |
| Native image or video generation | OpenAI or Google | Anthropic does not cover these output modalities — a structural limit it accepts. |
| Rich third-party ecosystem (plugins, partner agents) | OpenAI (ChatGPT) | Plugin marketplace and GPT Store more dense, eighteen months ahead of Anthropic on integration surface. |
Key takeaways
- Anthropic was born from a disagreement over the priority given to safety in LLM development, and enshrined that mission in its legal articles.
- The company has raised over $67 billion in five years, with two cloud giants — Google and Amazon — as strategic partners.
- Constitutional AI (RLAIF) is Anthropic’s signature technical approach: alignment through constitution-based self-critique, without human annotators in the reinforcement phase.
- The Claude family stands out for wide context windows and advanced agentic capabilities (Computer Use, Claude Code, multi-agent orchestration).
- A $380 billion valuation places Anthropic among the most highly capitalized tech companies, but the company is not yet profitable and remains dependent on investors to fund research and compute.