Anthropic has reversed an earlier decision that quietly restricted certain AI researchers from fully accessing its latest Claude Fable 5 model.

The company launched Claude Fable 5 earlier this week as a publicly available edition of the Mythos system equipped with added protections against potential misuse.

Upon release Anthropic explained that it would redirect queries involving cybersecurity biology and chemistry to weaker models preventing users from leveraging the stronger system for planning attacks or creating bioweapons.

The firm also noted that requests tied to AI development would see reduced model performance without any notice to the user which some developers viewed as an underhanded tactic to hinder competing AI projects according to prior Business Insider coverage.

The new announcement on Wednesday alters this approach so that Fable 5 will now inform users when a prompt faces refusal or redirection.

"We are updating Fable 5 safeguards for frontier LLM work to ensure transparency" an Anthropic spokesperson told Business Insider on Wednesday. "Beginning this week flagged prompts will clearly switch to Opus 4.8. Through the API any flagged prompts will include an explanation for the refusal."

The company noted "We chose the wrong balance and we apologize for missing the right approach."

Anthropic stated that these protections aim to handle national security concerns and stop foreign adversaries from advancing in frontier chips and large language models while noting that the vast majority of coding and machine learning tasks remain unaffected.

Mythos unveiled in April ranks among the strongest AI systems created to date with officials from governments and security bodies cautioning that its skills exceed those of existing public models in areas like advanced reasoning cybersecurity and scientific study.

Both outside researchers and Anthropic have highlighted risks that Mythos could speed up cyberattacks support biological or chemical weapons efforts and provide foreign entities with a potent new capability. Access to the model stays limited to select government and approved users rather than the general public.