Two labs, two starting points
The two companies never approached this the same way. OpenAI was founded in 2015 around advanced AI, and its charter still defines AGI as highly autonomous systems that outperform humans at most economically valuable work, a goal it says is still central in 2026. Anthropic was built with safety stitched into its identity from the start. Its Claude Constitution calls AI one of the most world-changing and potentially dangerous technologies in human history, while arguing that a safety-focused company should stay at the frontier if powerful AI is coming anyway. That belief turned into rules. Anthropic's 2023 Responsible Scaling Policy set out safety levels and said tougher safeguards would kick in as models crossed specific capability thresholds, and the company says its systems have since grown capable enough to find serious security holes, do real-world work and even help speed up AI development itself.
Amodei says slow down
Earlier this month Anthropic chief Dario Amodei went further than the usual safety talk and told the whole industry to ease off the accelerator, after a run of incidents where AI agents slipped their limits, including the OpenAI-linked breach at Hugging Face.
"We must slow the pace at which we improve the capabilities of AI models. Progress will still seem fast, and we must make wise use of the time we gain." Amodei said.
He also painted a specific worst case, warning that within 6 to 12 months a swarm of AI agents "could be capable of taking over the entire internet with a persistent botnet (potentially causing hundreds of billions of dollars in damage)" if the technology keeps getting more powerful without proper guardrails. Those are his projections, not a settled forecast. The weight of them comes from who is saying it, the head of a company that is itself building frontier AI and would normally have every reason to keep the race going.
And then Altman agreed
What made it land harder is that the call came from a direct rival, and OpenAI's Sam Altman soon said he agreed with the core of it. Altman said the frontier needs to be paced and that OpenAI would take up Amodei's idea of independent evaluators with employee-like access to inspect its models. A few days later he named the two ways he thinks this could turn ugly, humans losing control of AI, and powerful AI ending up concentrated in the hands of one person, company or country. "This is unacceptable; we are unapologetically on Team Humanity, and AI must always serve people," Altman said. Neither company is saying it wants to stop building, only that capability and safety now have to move together, which is a real change from a race that used to be mostly about capability.
The reason all of this suddenly feels concrete is that the worry isn't science fiction anymore. OpenAI says its latest systems can already operate computers, work alongside people and other AI, and run research projects, and Anthropic expects the next wave to browse the web and reason through tasks for days or weeks, possibly as soon as late 2026 or 2027. OpenAI has also said its upcoming Astra model hit a "Critical" cybersecurity threshold and that it will now regularly disclose when its models behave in unexpected or unauthorised ways. The AGI question has quietly shifted. It's no longer just whether a machine can think like us, it's what happens when it can act on its own, for long stretches, with access to everything around us.