Anthropic Releases Claude — A Safety-First Rival LLM

On 14 March 2023, the San Francisco AI startup Anthropic began offering its conversational large language model Claude. Two versions shipped—in the announcement's words, "Claude is a state-of-the-art high-performance model, while Claude Instant is a lighter, less expensive, and much faster option." This was not an open consumer launch: access to the chat interface and the developer API was by request, following what Anthropic called a "closed alpha" with Notion, Quora, and DuckDuckGo. Robin AI, AssemblyAI, and Juni Learning were among the named launch customers.

On the same day, OpenAI announced GPT-4.

The coincidence of two leading models shipping on one date was a fair preview of the structure that would follow. The capital and the talent that believed generative AI needs a counter-pole would gather around these two companies.

Departure from OpenAI

Anthropic's core had, until early 2021, been the core of OpenAI.

• Dario Amodei: OpenAI's VP of Research and a lead author of the GPT-3 paper.
• Daniela Amodei: OpenAI's VP of Safety and Policy. Dario's sister.
• Tom Brown: the GPT-3 team lead.
• Sam McCandlish, Jack Clark, Jared Kaplan: researchers on GPT-3 and on the scaling laws.

In late 2020 and early 2021 they left OpenAI as a group. No single official statement of reasons was ever issued by the departing group. The account that recurs in reporting is that after Microsoft's US$1 billion investment in 2019, OpenAI leaned toward commercialisation and speed in a way that no longer matched the safety research agenda—but that is journalists' reconstruction, not the founders' words.

Anthropic was founded in January 2021 by seven former OpenAI employees; it raised US$124 million in May 2021.

Claude did not begin in March 2023. The first version finished training in the summer of 2022, and Anthropic held it back, citing the need for further internal safety testing and a reluctance to be the party that starts a race. The 14 March launch came three and a half months after ChatGPT.

Constitutional AI

What set Anthropic apart was less a particular technical innovation than the explicit articulation of a safety method.

The paper Constitutional AI: Harmlessness from AI Feedback (Yuntao Bai et al., arXiv:2212.08073) was posted on 15 December 2022. The name comes from a sentence in the abstract: "The only human oversight is provided through a list of rules or principles, and so we refer to the method as 'Constitutional AI'."

The procedure has two phases:

1. Supervised phase. In the paper's words, "we sample from an initial model, then generate self-critiques and revisions, and then finetune the original model on revised responses."
2. Reinforcement learning phase. Sample two responses from the finetuned model, have a model judge which is better, train a preference model on those AI preferences, and use it as the reward signal—what the paper calls "RL from AI Feedback (RLAIF)".

The difference from RLHF is not that human judgement disappears but that it enters through the written principles rather than through per-example labels: the method works, in the paper's phrasing, "with far fewer human labels". The result, the authors report, is an assistant that is "harmless but non-evasive"—one that "engages with harmful queries by explaining its objections to them."

Anthropic presents Claude as trained with Constitutional AI. The 14 March post itself does not name the method, describing Claude only as based on "Anthropic's research into training helpful, honest, and harmless AI systems."

A Chain of Investments

What was unusual about Anthropic was that, while positioning itself as a safety-led research lab, it drew capital on the same order as OpenAI itself.

• April 2022: US$580 million raised, US$500 million of it from Sam Bankman-Fried's FTX.
• February 2023: Google invests about US$300 million, reportedly for roughly a 10% stake.
• May 2023: a US$450 million Series C led by Spark Capital, with Google among the participants.
• September 2023: Amazon announces up to US$4 billion, US$1.25 billion of it immediately.
• October 2023: Google commits up to a further US$2 billion.
• March 2024: Amazon completes the remaining US$2.75 billion; in November 2024 it adds another US$4 billion, doubling its total to US$8 billion.
• March 2025: Series E, US$3.5 billion at a US$61.5 billion post-money valuation.
• September 2025: Series F, US$13 billion at US$183 billion.
• November 2025: NVIDIA and Microsoft commit to invest up to US$10 billion and up to US$5 billion respectively; Anthropic commits to buy US$30 billion of Azure compute capacity.
• February 2026: Series G, US$30 billion at US$380 billion.
• 24 April 2026: Google puts in US$10 billion cash at a US$350 billion valuation and pledges up to US$30 billion more against performance targets—up to US$40 billion in all.
• May 2026: US$65 billion raised, at a reported US$965 billion valuation.

A common reading of Amazon's and Google's logic was that, against the Microsoft–OpenAI alliance, Anthropic was the credible alternative to back. By November 2025, Microsoft and NVIDIA had joined the cap table too. The company itself is covered on the Anthropic page.

Claude Model Generations, 2023–2026

The March 2023 pair was the first of many. Dates below are those of Anthropic's own announcements.

Two things changed shape along the way. First, the most capable tier split. Anthropic describes Fable 5.1 and Mythos 5.1 as the same model with different levels of safeguards: Fable is generally available, Mythos only through trusted-access programmes for work such as cyber defence. Second, release became entangled with government. Three days after launch, on 12 June 2026, a US export-control directive forced Anthropic to switch Fable 5 and Mythos 5 off for every customer until 1 July (the export-control episode).

Claude Code, the command-line coding tool introduced alongside Claude 3.7 Sonnet, has its own page: Claude Code.

Each announcement describes the new model as leading on coding and agentic benchmarks. Those are the vendor's figures, measured under its own conditions.

Two Frontiers to Point At

The shared 14 March 2023 launch date was a coincidence, but a telling one. The OpenAI–Anthropic duality would shape the terrain of the AI industry:

• OpenAI: speed and scale, deep partnership with Microsoft, commercial maximisation.
• Anthropic: safety and explainability, distributed partnerships with Google and Amazon, expansion into government and defence contracts—on 7 November 2024 it announced a partnership with Palantir and AWS to make the Claude 3 and 3.5 families available to US intelligence and defence agencies inside Palantir's IL6-accredited environment. That relationship broke in February 2026: after months of negotiation stalled over two exceptions Anthropic wanted—mass domestic surveillance of Americans and fully autonomous weapons—Defense Secretary Pete Hegseth said on 27 February that he was directing the Department of War to designate Anthropic a supply chain risk (AI goes to war).

The two were competitors. But for the industry as a whole, they also divided the role of preventing single-vendor dominance. From the autumn of 2023, when regulatory debate intensified, the fact that governments and researchers had two contrasting frontiers to point to owes a great deal to Anthropic's existence.

And the founders of both companies—who once shared the same OpenAI offices, building GPT-3—now build competing frontier models from separate buildings.