Fable 5 Is Back After the US Lifts Anthropic's AI Ban — Jul 3, 2026

0:00The most powerful AI the U.S. banned last month is back, and biomedical labs want it. Welcome to MedAI Times Podcast, your daily update on medical AI. Don't forget to like and subscribe. This is the MedAI Times Weekly Wrap for Friday, July 3rd.

0:15Here's the week in one breath. The United States lifted its export ban on Anthropic's two most powerful models. Anthropic shipped a cheaper agentic model called Sonnet 5. It also launched Claude Science, aimed straight at pharma.

0:29The FDA cleared an on-premise brain tumor AI and an auto-contouring tool for radiation therapy. And federal regulators slapped a corrective action plan on a company using AI to review Medicare claims.

0:42On Monday, we told you Washington had locked Anthropic's smartest models behind a government allow list. This week it opened the gate for everyone. Start with the reversal. On June 30th, the Commerce Department lifted the export controls it imposed on June 12th.

1:00And by July 1st, Anthropic's flagship Fable 5 and its cyber-focused sibling Mythos 5 were back worldwide, this time behind a new safety classifier. The order was triggered when Amazon researchers found a jailbreak that slipped past one of Apple's cybersecurity guardrails.

1:20And Amazon's chief executive flagged it to federal authorities. Here's the medical part. And it's the part that matters for this show. Mythos 5 was the model biomedical labs reached for roughly 80% of the time when they wanted molecular hypotheses and drug design ideas.

1:37For three weeks, those academic and startup teams were locked out. Now they're back in. So treat this as a restored tool, not a proven one. Signal for the drug discovery bench, noise at the bedside. Mythos designs molecules.

1:51It does not read your chart. Which lines up with the second Anthropic move this week? On the same day, it unbanned Fable. The company launched Claude Science, a research workbench that actually runs the analysis instead of just planning it.

2:05It submits compute jobs, spawns specialist sub-agents, and includes a reviewer agent to catch its own errors. Early users are named and serious. UCSF's Brain Tumor Center said it cut a genetic analysis to about one-tenth of the time.

2:21Novo Nordisk and Manifold Bio are on the list. Company reported, not peer-reviewed. But the direction is clear. Anthropic also shipped Sonnet 5. And this one's about cost. It lands close to Opus Class quality at $2 per million input tokens and $10 per million window.

2:44For hospitals and startups building agents on a budget, cheaper frontier-adjacent inference is the quiet story under the louder headlines. To the clinic. The FDA cleared Sercari Medical's Brain Tumor Segmentation module, and the interesting part is where it runs.

3:00Entirely on-premise, no cloud connection, built for HIPAA-strict systems and VA hospitals. It maps a tumor from a standard brain MRI without manual contouring, with a DICE overlap score of 0.82 across 100 patients.

3:15In a year of cloud everything and ambient scribe consent lawsuits, a model designed to never leave the building is a deliberate countermove. Staying in imaging, GE Healthcare picked up FDA clearance for the latest version of its auto-contouring software for radiation therapy planning.

3:34Contouring is the tedious step where someone outlines every organ to protect before the beam goes on. Automating it saves hours per plan. It's a workflow clearance, not a diagnostic one. But workflow is where most of these tools actually earn their keep.

3:50Now the reprimand. Federal regulators found Vertix Health out of compliance in the Wizzr pilot, the program using AI to screen Medicare prior authorization requests in six states. The company missed the 72-hour turnaround rule, with decisions stretching to about five days before improving, and it's been ordered onto a corrective action plan.

4:13The design flaw underneath is the one to remember. The contractor is paid a share of what it saves the government, which means it earns more the more it denies. AI didn't invent that incentive, but it scales it.

4:27One general note worth your time. This week also brought fresh money into the space, with Qualified Health raising $125 million to push enterprise AI deeper into health systems. The capital keeps arriving faster than the outcomes data does.

4:43Spotlight. The quiet frontier this quarter is surgical robotics foundation models. NVIDIA's GR0TH is a 3-billion-parameter vision-language-action model that turns a text command about a clinical task into robot motion, trained on OpenH, a 700-plus-hour surgical dataset from 35 organizations.

5:04It's a research and simulation stack today, not an autonomous surgeon, and the gap between a benchmark and an operating room is enormous. But it's the same recipe that reshaped language models, pointed at the OR.

5:18Every source and link is in the description. If your research or drug discovery team lost a frontier model for three weeks, what did you switch to, and did anything actually break? Thanks for listening. Find us on YouTube and your favorite podcast app.

5:32See you tomorrow.