
Anthropic
AI safety company; developer of Claude; subject of the largest publicly named publisher content settlement.
On 12 June 2026 the US ordered Anthropic to withdraw Claude Fable 5 and Mythos 5 from every foreign national on earth; Washington partially lifted the curb on 27 June, five weeks before the EU AI Act's GPAI enforcement begins on 2 August.
Last refreshed: 27 July 2026 · Appears in 6 active topics
Can the EU's GPAI framework survive if US models can be suspended from Washington overnight?
Timeline for Anthropic
Mentioned in: AI Office gains enforcement powers in August
European Tech SovereigntyMentioned in: Marking code draws 190, none of them TV
Media's AI PivotMentioned in: Time starts selling ads to the machines
Media's AI PivotMentioned in: Samsung in talks to anchor Mistral at €20bn valuation
European Tech SovereigntySpent USD1.97 million on federal lobbying, more than Nvidia
AI: Jobs, Power & Money: Anthropic now outspends Nvidia in WashingtonBackground
Anthropic is an AI safety company founded in 2021 in San Francisco by former OpenAI researchers Dario and Daniela Amodei, built around the mission of researching and deploying AI systems safely. Its flagship model family, Claude, competes directly with OpenAI's GPT and Google's Gemini lines.
Amazon is its largest external backer, having committed up to $4 billion, and Anthropic runs Project Glasswing, a restricted government- and critical-infrastructure-access programme that has expanded to seventeen partners including banks, cybersecurity firms and national agencies, offering early access to frontier capability in exchange for tighter oversight .
That dual identity, safety-focused research lab and commercial frontier-model vendor, puts Anthropic at the centre of arguments over AI's economic and security impact. Its own usage research is now widely cited evidence of real labour substitution, and its models have twice drawn direct government intervention: once through a US suspension order, separately through publisher-copyright litigation that set the market's reference settlement figure.
The US suspended Anthropic's top models
On 12 June 2026 Commerce Secretary Howard Lutnick ordered Anthropic to bar all foreign nationals from Claude Fable 5 and Mythos 5; unable to check nationality mid-conversation, Anthropic switched both models off for every customer on earth . The trigger was a rival's claimed jailbreak of Mythos 5, yet the same vulnerability sat in GPT-5.5, which stayed on sale; the administration invoked national-security authority under the deemed-export doctrine built for chips.
The EU's cybersecurity agency ENISA had joined Anthropic's restricted Project Glasswing programme in April and lost access within weeks, prompting an MEP to call the episode a real kill-switch. The Commerce Department partially lifted the restriction on 27 June, but the precedent survives the reversal .
It faces its largest content settlement
News Corp chief executive Robert Thomson told investors on 8 May 2026 that News Corp anticipates a $1.5bn settlement with Anthropic, an unclosed figure roughly six times its prior five-year OpenAI deal . The same day Anthropic signed a seven-year, $1.8bn compute deal with Akamai, its largest contract ever, placing Claude inference on the EU network edge ahead of the AI Act's August deadline .
Separately, the $1.5bn Bartz v Anthropic class settlement, covering 92.77% of claimed works, remained under judicial submission through May and June with no final approval entered, leaving Anthropic's exposure as the industry's reference point for what publisher licensing now costs.
Its own product crashed a market
Anthropic's Claude Cowork launch on 30 January 2026 wiped roughly $24 billion from India's Nifty IT index within a week, as TCS, Infosys, Wipro, HCLTech and Tech Mahindra each fell 5 to 8% on fears the tool would gut labour-arbitrage coding work .
Its own researchers later measured the mechanism directly: Massenkoff and McCrory found real Claude usage covers 75% of programmer tasks, against 35.8% for computer and maths roles more broadly, with the most exposed workers skewing older, more educated and higher-paid . That observed-exposure methodology is now the most cited internal evidence of real AI-driven labour substitution.