The U.S. government's move to stop Anthropic's Fable 5 and Mythos models turns a long-theoretical debate into a real-world test of who should have the power over world-changing AI.

Why it matters: Handing the reins to companies means trusting firms with huge commercial incentives. Leaving decisions up to governments could lead to secretive high-stakes decisions with little public approval.

Catch up quick: Anthropic was given 90 minutes last week to take down Fable and Mythos after Amazon raised concerns with U.S. officials about a jailbreak that could bypass Fable's guardrails and expose cybersecurity capabilities, sources familiar told Axios.

• The government imposed stringent export controls that ultimately led Anthropic to take the models offline entirely.

State of play: AI policy researchers and industry critics say the U.S. is now regulating frontier models through emergency intervention rather than a clear process.

• "Things feel very ad hoc," Adam Gleave, CEO at FAR.AI, an AI safety research firm, told Axios, adding "you're only as safe as the least safe model on the market."
• "In every sport, you have to separate the referee from the team players ... regulation should never come from industry," Connor Leahy, U.S. executive director of ControlAI, told Axios, arguing that government should have the authority on AI despite its current "capabilities."

Case in point: When Amazon flagged its concerns to the administration, it went to Treasury Secretary Scott Bessent, who was already close to the Mythos rollout because of its planned use by major U.S. banks.

• Model oversight has implications for national security, competition and free expression. And right now these decisions are being made through improvised channels, rather than a clear AI regulatory framework.

Between the lines: The most workable answer may be neither company self-policing nor unilateral government control, but an entirely new AI-fluent agency tasked with combining company testing, outside audits and government authority.

Keep reading.

Microsoft is moving Copilot Cowork to usage-based pricing as it expands access to the enterprise AI tool — and is considering a Microsoft-hosted version of DeepSeek as a cheaper model option.

Why it matters: Microsoft's move to add a model from a Chinese AI company could draw criticism.

The big picture: Agentic tools like Copilot Cowork, Anthropic's Claude Code and OpenAI's Codex can keep calling AI models as they work through tasks — boosting productivity but also creating bonkers AI bills.

Driving the news: Microsoft says companies using Copilot Cowork will pay based on how much compute they use.

• The company tells Axios it is exploring a fine-tuned version of DeepSeek V4, or another open-source model, as a lower-cost alternative to the Anthropic and OpenAI models now powering Copilot Cowork.
• Microsoft says it expects to make a lower-cost model available in the coming weeks and will confirm its choice then.

Zoom out: The testing reflects Microsoft's broader push toward a multi-model approach, rather than relying only on models from OpenAI and Anthropic.

Between the lines: If Microsoft goes forward with DeepSeek, the company says, the model would be optional for customers and fully hosted on Azure, keeping customer data within Microsoft's cloud and covered by Azure's enterprise security, compliance and data-residency controls.

• Microsoft says it has also fine-tuned the model and added safeguards, including changes aimed at reducing bias.

What they're saying: Charles Lamanna, Microsoft's executive vice president for Copilot, agents and platform, told Axios that testing showed Copilot Cowork could not be offered on an unlimited-use basis.

• "We have users who do hundreds of tasks a week, which is great — they're way productive — but the consequence is the costs can go very high," Lamanna said.

AI systems are beginning to replicate the same anti-LGBTQ bias and misinformation problems that have long plagued social platforms, according to a new GLAAD report previewed at Axios' AI+NY Summit.

Why it matters: The problems GLAAD flags — biased training data, privacy risks, automated discrimination, misinformation and the suppression of legitimate speech — extend beyond LGBTQ users to other minorities and groups in political disfavor.

Driving the news: GLAAD's "Build for Everyone: A Framework for LGBTQ Representation and Safety in AI" report argues that AI is already shaping how people search for information, access health care, apply for jobs and participate online.

• The report cites harms including Meta's updated Llama 4 model recommending conversion "therapy" in response to queries and moderation systems that wrongly suppress legitimate LGBTQ content.
• It offers blunt and concrete recommendations to the industry: Fix the biased foundation; don't automate discrimination; maintain human oversight; respect data privacy and engage civil society.

What they're saying: "Neutrality is no longer an option," GLAAD CEO Sarah Kate Ellis says in the report, warning that AI systems can threaten LGBTQ people's "health, safety, and civil rights" when they treat LGBTQ lives as fringe or fail to catch disinformation.

Between the lines: GLAAD is trying to push AI companies before bad practices harden — especially because a small group of foundation models from OpenAI, Google, Meta and Anthropic now feeds a much larger app ecosystem.

The big picture: GLAAD's report uses a similar approach to the one it has taken for several years in evaluating the safety of various social media platforms.

• In its first report, back in 2021, the organization laid out key issues and challenges, while subsequent reports graded the companies.

Nauk Nauk has raised $20 million to expand its AI video app that turns photos of toys into short animated videos, co-founder Daniel Liu tells Axios.

Why it matters: The company says it's targeting the growing market of "kidults" who want to use technology to indulge their playful side.

The big picture: Nauk Nauk is launching as free AI video-generation options become more limited, creating an opening for specialized consumer apps.

Driving the news: Nauk Nauk (pronounced knock-knock) is coming out of beta with an app that can create 15- to 20-second videos with audio or music soundtracks from a single image and a user-written prompt.

Keep reading.

• Apple has plans to launch AirPods with cameras by 2027, as well as foldable phones. (Bloomberg)

Shot: Mady here after having four back-to-back interviews with sources who told me AI would lead to a net loss of jobs.

Chaser: Brad Smith, vice chair and president of Microsoft, said "let's not panic" about AI's alleged hit to the labor market, signaling that the doom and gloom over AI and work may be related to investment goals rather than truth about the technology.

More of his conversation with Axios' Mike Allen here.