Claude Fable 5 Is Back, But Users Say It Feels Different

July 02, 2026

See what enterprise-ready AI support looks like

How are leading teams getting AI support to work? We're breaking down the playbook, live July 9.

Anthropic’s most powerful public Claude model is back online.

After U.S. export control restrictions were lifted, Anthropic reinstated Claude Fable 5, its high-end model built for long-horizon coding, research, and complex knowledge work. The model had been temporarily unavailable after government concerns around advanced AI capability and potential misuse. Anthropic’s official Fable page now says access has been restored, while Axios reports the reinstatement followed the lifting of U.S. controls. Anthropic, Axios

But the comeback has not been smooth.

Some users are reporting that the restored version feels more restricted, especially in coding and agentic workflows. On BridgeBench, the July 1 version shows weaker results in certain UI and reasoning tests, including a 47.6 overall score on UI tasks and 40.6 on reasoning. However, another BridgeBench coding-cost page still shows Fable 5 achieving 100% strict success across 30 coding tasks, so the picture is not a simple “model got worse” story. BridgeBench UI, BridgeBench Reasoning, BridgeBench Coding Cost

The likely reason is guardrails.

Anthropic says Fable 5 includes stronger safeguards for areas like cybersecurity and biology, with some flagged queries routed away from Fable to safer fallback models. That means users may still be selecting Fable 5, but not always getting the full Fable behavior on sensitive or borderline tasks. For normal users, this may look like refusals, weaker completions, or less aggressive problem-solving.

This is the bigger story: frontier AI is no longer just about raw intelligence. It is also about who is allowed to access that intelligence, under what conditions, and with how many safety filters in between.

Fable 5’s return shows how AI companies are now balancing three forces at once: user demand for stronger models, government pressure around national security, and safety concerns around misuse.

The model is back.

But the version users experience may now depend as much on policy and routing as on the model itself.