Anthropic

Anthropic, the US AI safety company behind the Claude model family, is racing OpenAI and Google to define frontier AI while navigating export controls, model IP disputes, and a near-US$1 trillion valuation ahead of a prospective IPO.

AI· ·4 論調 ·2026年7月3日

What it is

Anthropic is a US artificial intelligence safety company incorporated as a Public Benefit Corporation and headquartered in San Francisco, California. It develops the Claude family of large language models and publishes research on interpretability, alignment, and the risks of frontier AI. As of July 2026, the company holds a US$965 billion post-money valuation following its May 2026 Series H round and has filed a confidential S-1 for a prospective IPO, which would rank among the largest US technology listings on record.

History

Dario Amodei, then OpenAI's VP of Research, and his sister Daniela Amodei, along with six other OpenAI researchers, founded Anthropic on January 26, 2021, after disagreements over how seriously OpenAI was treating AI alignment risks relative to scaling. The company raised US$124 million in seed funding that spring. In December 2022, Anthropic published "Constitutional AI: Harmlessness from AI Feedback" (arXiv:2212.08073), a training method that uses AI-generated feedback guided by written principles to reduce harmful outputs with fewer human labels, becoming a reference technique in the alignment field. Claude launched publicly in March 2023; Claude 2 followed in July 2023. Google committed US$500 million in 2023; Amazon committed US$4 billion across late 2023, deploying the remaining US$2.75 billion in March 2024. Claude 3 (March 2024, three tiers: Opus, Sonnet, and Haiku) was Anthropic's first broadly competitive frontier release. The Series G in February 2026 valued the company at US$380 billion; the Series H in May 2026 pushed that to US$965 billion.

Current state

The Claude 5 family, completed with the June 30, 2026 Claude Sonnet 5 release, spans three model tiers. The flagship Fable 5, which led key agentic benchmarks at launch, faced a 19-day US export restriction starting June 12, 2026, after Amazon researchers identified a security jailbreak; the US Commerce Department lifted the ban July 1 following a patch and a new cybersecurity classifier (米国が19日間の輸出禁止を解除、クロードFable 5とMythos 5が全世界で復元). The episode created enterprise compliance exposure: customers running contracts specifying Fable 5 were silently moved to a different model for those 19 days. In June 2026, Anthropic filed suit over what it characterized as the largest-ever model distillation attack, alleging Alibaba's Qwen lab extracted Claude capabilities through 28.8 million fraudulent exchanges via 25,000 fake accounts. Competitive pressure from open-weight Chinese labs, including MiniMax M3, is compressing inference cost margins from below.

Relationships

Amazon is Anthropic's primary infrastructure partner and largest strategic investor, contributing US$5 billion to the Series H and signing a 5 GW compute supply agreement in early 2026. Anthropic's models serve as a flagship option on Amazon Web Services. Google has committed over US$3 billion across rounds and provides TPU capacity under a separate 5 GW arrangement with Broadcom. Micron, Samsung, and SK hynix joined as strategic partners at Series H, securing chip supply for planned compute scale-up. A wave of senior researcher departures from Google DeepMind to Anthropic through mid-2026, including AlphaFold co-creator John Jumper and Gemini lead Jonas Adler, reflects the pull of pre-IPO equity. Anthropic's Long-Term Benefit Trust, comprising independent trustees, holds veto rights over decisions that could compromise the safety mission, an unusual governance structure for a company at this valuation.

What to watch

Whether Anthropic converts its confidential S-1 into a public filing before end-2026 is the most immediate question. The US export-controls episode around Fable 5 demonstrated that individual model releases can trigger national-security review, a precedent with implications for Anthropic's international business. The Alibaba distillation case may establish legal precedent for model IP protections across the industry. Annualized revenue grew from roughly US$1 billion in December 2024 to US$47 billion at the Series H announcement, but sustaining that trajectory depends on Anthropic's ability to maintain model performance leads as open-weight competitors narrow the capability gap and inference prices continue to fall.

一次記録 · 4

Anthropic (company page) — Anthropic describes itself as an AI safety and research company, a Public Benefit Corporation governed by a Long-Term Benefit Trust, dedicated to building reliable, interpretable, and steerable AI systems.

Anthropic (Series H funding announcement) — Anthropic's official announcement of its US$65 billion Series H at a US$965 billion post-money valuation, led by Altimeter Capital, Dragoneer, Greenoaks, and Sequoia; states run-rate revenue exceeded US$47 billion and confirms a confidential S-1 filing.

Anthropic (Constitutional AI research page) — Anthropic's own documentation of Constitutional AI, published December 2022, describing the two-phase supervised and reinforcement-learning training technique that uses AI-generated feedback rather than extensive human labeling to align model outputs.

arXiv (Constitutional AI paper, arXiv:2212.08073) — Yuntao Bai et al., Anthropic, December 15, 2022: foundational paper introducing Constitutional AI and Reinforcement Learning from AI Feedback, with 51 co-authors across Anthropic research teams.