Sponsored by
Welcome, AI enthusiasts
Anthropic handed Fable 5.1 unsolved problems before putting it live for public and it came back with a new map covering a third of Venus. Fable 5.1 reached every major cloud today at up to 45 percent less to run than Fable 5. Mythos 5.1 went only to vetted cyber defenders and life scientists. Let's dive in!
In today’s insights:
Anthropic's Fable and Mythos 5.1 just Dropped
OpenAI's most capable model hits Critical Cyber Limit
Your Lab Report goes through ChatGPT at 99.1% Safe
Read time: 4 minutes
LATEST DEVELOPMENTS
MODEL LAUNCH
🚀 Anthropic's Fable and Mythos 5.1 just Dropped
Evolving AI: Anthropic doubled Fable 5.1's science score and cut running costs by up to 45 percent.
Key Points:
Mythos 5.1 designed protein binders ten times tighter than the top entries on three competition targets.
Fable 5.1 trained a neural network that mapped a third of Venus from thirty year old NASA radar.
Millennium engineers watched it explain a rare crash that had gone unsolved for close to five years.
Details:
Fable 5.1 worked through a set of hard programming tasks on its own and finished 55.8 percent of them, where Fable 5 finished 42 percent. On a harder test built around scientific research tasks it reached 52.6 percent against 24.7 percent for Fable 5. Anthropic says Fable 5.1 came out ahead of Opus 5 on every one of the eight benchmarks it published. For 38 hours the payments company Ramp left it running unattended, and it caught a mistake in its own earlier work then set six experiments going overnight.
Why It Matters:
In the past few months open source models have caught the frontier, and several free releases now match what the paid labs charged a year ago. Anthropic has to keep pushing to stay ahead of them, and OpenAI has an unreleased model family it named Astra in August. Fable 5.1 and Mythos 5.1 are the state of the art models today, though how far ahead stays unclear until OpenAI ships.
Find the Perfect Voice for Your Brand's AI
Consumers can tell the difference. 79% say AI voices should come from real, attributed actors, and 61% say a voice is most memorable when it's exclusively tied to one brand.
Voices is the only platform that designs, licenses, and captures your brand's AI voice from real, consenting professional talent—never from scraped data. From strategic voice casting and brand calibration to consent-based rights management and studio-grade capture, we handle the entire process end to end.
BMW, SuperBloom, and Cresta already trust Voices to build their signature AI voice. Get a demo and see how a licensed, brand-owned voice becomes a real asset—not a legal liability.
AI CYBER THREAT
🚨 OpenAI's most capable model hits Critical Cyber Limit
Source: OpenAI
Evolving AI: Astra, Next-Gen OpenAI Model can find unknown flaws and build working exploits without a human guiding it.
Key Points:
ExploitBench gave Astra a flawless 100%, so OpenAI rebuilt the test with flaws disclosed after the model's cutoff.
The Preparedness Framework has never placed a model at Critical for cybersecurity until this one.
Thomas Larsen and other researchers warn that models showing less reasoning would make misbehavior harder to catch.
Details:
OpenAI ran the model against 20 high severity flaws in Google's V8 engine disclosed after its training cutoff. Astra cleared that set with far higher code execution rates than GPT 5.6 Sol, and found two unknown flaws now going to the maintainers. From there, expert reviewers watched it build a chain that escaped a hardened browser and ran commands on the host machine from one HTML file. The older model tried to attack nearby systems in 56% of honeypot tests, and Astra made no such attempt at any point. OpenAI is sending advanced cyber work to a small group of alpha testers before opening wider access.
Why It Matters:
ChatGPT and Codex users meet the safeguards first, since long tasks can pause for approval and API jobs stop when a monitor fires. Those monitors work by reading what the model is thinking, which is how investigators piece things together when an agent goes wrong. Sam Altman says OpenAI is slowing its later models, which buys time if people can no longer follow what a model works through.
Source: OpenAI
Evolving AI:Doctors can now ask ChatGPT what changed in your chart since your last visit.
Key Points:
Epic, the record system used by most large US hospitals, can now push your history and labs into ChatGPT.
Physicians gave 4,363 ratings across 27 clinical tasks including pre-visit review and marked 99.1% safe.
A separate plugin opens nine official sources to clinicians, including PubMed, ClinicalTrials.gov, DailyMed and CMS coverage.
Details:
ChatGPT for Healthcare now connects to a hospital's Epic system so authorized clinicians can pull a patient's history, recent lab results and medication changes into one place without digging through separate chart sections. OpenAI ran the evaluation with physicians across 27 clinical use cases covering pre-visit review, clinical timelines, medication review and handoff summaries, and those physicians returned 4,363 ratings with 99.1% marked safe. A workspace administrator has to switch the integration on, and individual ChatGPT accounts cannot get it at all.
Why It Matters:
UCSF Health is testing this on live patient records, so the summary your doctor reads before your appointment may soon be written by a model. Epic keeps records for more than 325 million people, and every large hospital group in the country runs the same software. Physicians did not mark every answer safe, and those answers now go to real patients in real appointments.
👀 Click on the image you think is real
QUICK HITS
☁️ Anthropic signed a $35B cloud deal with Nvidia-backed Lambda for a 350MW Texas data center built by Hut 8.
🦄 AfterQuery hit a $3.2B valuation, 10x in five months, making it Y Combinator's fastest unicorn ever.
🔥 AIR emerged from stealth with $50M from Sequoia and Greenoaks to build a firewall for AI agents, after finding 27% of agent add-ons fail security checks.
🛡️ CrowdStrike launched Falcon Guardian to find and shut down unapproved AI agents running on company laptops.
📈 Trending AI Tools
🎞️ Guideless - AI tool that automatically turns your clicks and workflows into step-by-step video guides*.
📄 Murmell - Google Docs for AI agents, keeps running after you close the laptop,
🎓 Creatium Coach - Multimedia mentor that takes your creative work from mid to great.
🔒 TrustedRouter - Every model behind one interface, with provable privacy.
*partner link