Anthropic's newest AI launch has triggered debate over the abilities and limits built into its updated models.

On Tuesday, Anthropic released Claude Fable 5 and Mythos 5, its anticipated Mythos-class models.

With the launch, the firm outlined two notable protections: the models may quietly deliver weaker help when they sense users pursuing advanced AI research, and some queries are sent automatically to lower-performing models.

In a system card, Anthropic stated the steps seek to lower the chance that strong AI tools assist users in building competing frontier models or hastening risky developments.

Critics countered that the protections could sideline researchers, consolidate influence among leading AI firms, and reduce response quality without users knowing.

Here are perspectives from informed voices in AI, cybersecurity, and policy on Anthropic's recent steps:

David Kasten, head of policy at Palisade Research

David Kasten expressed that Anthropic appears sincere in its effort to address risks tied to Mythos.

"I accept that Anthropic is working seriously to address what it views as the two main risk areas in Mythos," Kasten told Business Insider.

Even so, Kasten noted that releasing the model brings exposure because "it remains an ongoing contest between those attacking and those defending."

Kasten added that Anthropic likely views itself as competing against other AI labs that could soon release comparable models.

"For them, the question is how much of their current advantage they believe they can risk losing," he said.

Davi Ottenheimer, VP of Trust and Digital Ethics at Inrupt

Davi Ottenheimer, a noted digital safety specialist, challenged whether Mythos posed the level of danger Anthropic described.

"Anthropic's messaging in April indicated Mythos was too risky for public access," Ottenheimer told Business Insider. "Yet now they are offering it, without changes, to the public."

Mythos was kept from public release in April after Anthropic cautioned that the model excelled at spotting cybersecurity weaknesses and therefore raised safety concerns.

"They are treating security as a promotional tactic," Ottenheimer added.

Roman Stanek, founder of GoodData

Roman Stanek stated that raw AI power is not the core issue in cybersecurity.

"Regarding AI and cybersecurity, the weaknesses that raise concern have been known for about two decades," Stanek said in a note to Business Insider. "They simply were never addressed."

"No one was willing to fund an engineer to resolve them," he added. "They will not fund an AI to resolve them either."

Elie Bakouch, research engineer at Prime Intellect

Elie Bakouch faulted Anthropic for deliberately reducing Mythos performance on select AI development tasks.

"Mythos will intentionally underperform on frontier LLM research tasks," Bakouch posted on X on Tuesday. "This is deeply disappointing for the research community."

Bakouch also described the hidden nature of the adjustment as "unreasonable."

Jeremy Howard, cofounder of AnswerDotAI

Jeremy Howard argued that Anthropic's protections may heighten consolidation within the AI sector.

"Anthropic has taken the reverse of the cautious route," Howard posted on X on Wednesday. "They permit themselves, as the current leader, to apply their top model to frontier AI research."

Howard said the outcome is that "the AI frontier moves forward while power gaps widen."

Patrick Moorhead, founder and CEO of Moor Insights & Strategy

Patrick Moorhead reported that his initial use of Fable 5 proved disappointing.

In a post on X on Wednesday, Moorhead said the model declined to assist with earnings analysis points and a board presentation because Fable 5 judged the requests "too risky."

"Is this the model that prompts such concern and values Anthropic at $1T?" he wrote. "Alright then."

Deedy Das, partner at Menlo Ventures

Deedy Das emphasized the model's abilities over its protections.

"Claude Fable 5 stands as the most striking model that genuinely concerns me about the future of software engineering," Das posted on X on Wednesday.

Das highlighted cases such as migrating a 50 million-line codebase, creating advanced 3D graphics, and surpassing other models on optimization work.

He also noted that Fable 5 carries a price similar to OpenAI's GPT-5.5 and costs six times less than GPT-5.5 Pro.