Anthropic

Anthropic is an AI-safety research company founded in the US by seven people who left OpenAI in January 2021, among them the Amodei siblings, Dario and Daniela. Headquartered in San Francisco, it builds Claude on two foundations: Constitutional AI, in which a written list of principles is the only channel through which human oversight enters, and mechanistic-interpretability research. Its run-rate revenue crossed US$47 billion in May 2026; the Series H that closed on 28 May 2026 raised US$65 billion at a US$965 billion post-money valuation, and on 1 June 2026 it submitted a confidential draft S-1 to the SEC.

Out of OpenAI, into a US$965 bn valuation

Dario Amodei led research at OpenAI as VP of Research, having driven GPT-2 and GPT-3, before leaving in January 2021 with six colleagues including his sister Daniela, who became COO. The stated mission was to advance safety research faster than frontier capabilities. A US$124 m Series A followed in May 2021, and the company gathered ex-OpenAI researchers Jared Kaplan, Tom Brown, and Sam McCandlish — principal authors of the GPT-3 paper.

The Constitutional AI paper, "Constitutional AI: Harmlessness from AI Feedback" (arXiv:2212.08073), appeared in December 2022. On 14 March 2023 Anthropic opened request-based access to Claude and the lighter Claude Instant — not an open release — with Claude 2 in July and Claude 2.1 (200,000-token context) in November. Google invested about US$300 m in February 2023; Amazon announced up to US$4 bn in September 2023 and later doubled that to US$8 bn. Claude 3 (Opus / Sonnet / Haiku) shipped in March 2024, Claude 3.5 Sonnet in June, and Computer Use — a model that reads the screen and drives mouse and keyboard — in October. Claude Opus 4 and Sonnet 4 arrived in May 2025, launched together with the activation of the AI Safety Level 3 (ASL-3) deployment and security standards under the Responsible Scaling Policy. Claude Opus 4.5 followed in November 2025, Claude Sonnet 5 on 30 June 2026, and Claude Opus 5 on 24 July 2026. Claude Code, the developer CLI, went generally available in May 2025.

In November 2024 Anthropic, Palantir, and AWS announced a partnership to run Claude 3 and 3.5 in Palantir's IL6-accredited classified environment for the US intelligence and defense community — the central case in the 2024 AI military-use debate. On the capital side, February 2026 brought a US$30 bn Series G at US$380 bn post-money, led by GIC and Coatue; in April 2026 Amazon committed up to a further US$25 bn on top of the US$8 bn already invested, and Google up to US$40 bn in a mix of cash and compute; and on 28 May 2026 the US$65 bn Series H closed at US$965 bn post-money, led by Altimeter, Dragoneer, Greenoaks and Sequoia. That round put Anthropic above OpenAI's US$852 bn post-money valuation of March 2026.

Claude on three distribution paths

Key events

• January 2021: Seven people including Dario and Daniela Amodei leave OpenAI; Anthropic is founded.
• May 2021: US$124 m Series A.
• April 2022: US$580 m Series B, roughly US$500 m of it linked to Sam Bankman-Fried's FTX; those shares passed to the FTX bankruptcy estate after the collapse.
• December 2022: The Constitutional AI paper is published.
• 14 March 2023: Request-based access opens to Claude and Claude Instant. OpenAI announced GPT-4 the same day.
• July 2023: Claude 2 becomes generally available; in September Amazon announces up to US$4 bn.
• March 2024: Claude 3 family (Opus / Sonnet / Haiku).
• June 2024: Claude 3.5 Sonnet.
• October 2024: Computer Use.
• November 2024: Palantir-AWS partnership to supply Claude to US intelligence and defense.
• May 2025: Claude Opus 4 and Sonnet 4 launch with ASL-3 protections activated; Claude Code goes generally available.
• November 2025: Claude Opus 4.5.
• February 2026: Series G — US$30 bn at US$380 bn post-money.
• April 2026: Amazon commits up to a further US$25 bn; Google up to US$40 bn.
• 28 May 2026: Series H — US$65 bn at US$965 bn post-money.
• 1 June 2026: Confidential draft S-1 submitted to the SEC.
• 30 June and 24 July 2026: Claude Sonnet 5 and Claude Opus 5.

Key metrics

• Run-rate revenue: US$47 bn as of May 2026 (Anthropic's own figure)
• Valuation: US$380 bn post-money at the Series G (February 2026); US$965 bn post-money at the Series H (May 2026)
• Principal investors: Amazon (US$8 bn plus up to US$25 bn more, April 2026), Google (up to US$40 bn in cash and compute, April 2026), plus Spark Capital, General Catalyst, Lightspeed, GIC, Coatue, Altimeter and others
• Listing: confidential draft S-1 filed 1 June 2026; share count and price not yet set

Where the safety claim meets the defence contract

Anthropic is unusual among large AI labs in putting safety at the centre of its public identity. Its 2023 Responsible Scaling Policy introduced graded AI Safety Levels that tie capability progress to specific mitigations. When ASL-3 was activated alongside Claude Opus 4 in May 2025, Anthropic was explicit that this was precautionary and provisional: it had not determined that Opus 4 crossed the capability threshold requiring ASL-3, only that it could no longer clearly rule out CBRN-related risk as it had for earlier models. Its mechanistic-interpretability work — sparse-autoencoder feature extraction in frontier models among it — keeps safety research at academic frontier quality. Yet the founding logic, that safety-focused labs need to be at the frontier, has been criticised as itself a contributor to the acceleration of capabilities: the "acceleration paradox".

The November 2024 Palantir-AWS partnership provoked particular debate over whether an avowed AI-safety company should supply LLMs to military and intelligence agencies. Anthropic's usage policy prohibits use of Claude for nuclear or biological weapons design and for mass domestic surveillance while permitting defence analytics and cybersecurity; in February 2026 its insistence on keeping precisely those two carve-outs put it in direct conflict with the US administration.

Litigation has expanded in step with revenue. Universal Music and other rights-holders sued over song lyrics in training data (from 2023), publishers have filed copyright cases, and Reddit sued over scraping of its corpus — the same litigation surface OpenAI carries. The depth of the Amazon and Google relationships also exemplifies the "compute-for-capital" model that has emerged at the frontier: equity in exchange for committed cloud spend. Balancing that dependency against research independence remains an open problem.