🔑 Fable 5 is Back ONLINE Worldwide

For nearly three weeks, one of the world's most powerful AI models simply vanished. Not because it failed, but because the US government pulled the plug over a single security flaw. Now it's back with tougher safeguards, closer government oversight, and a glimpse of what future AI launches might look like. Let's dive in!

In today’s insights:

LATEST DEVELOPMENTS

Evolving AI: Anthropic is bringing Fable 5 back worldwide today after agreeing to deeper government testing.


Key Points:

Anthropic retrained a classifier that blocks the reported bypass in over 99 percent of cases, and independent CAISI evaluators validated the updated safeguards. It is also drafting a shared scale with Amazon, Microsoft, and Google to rank how dangerous any jailbreak really is. That wider safety margin has a tradeoff, as requests that trigger Fable 5's cybersecurity safeguards are now automatically handled by Opus 4.8 instead.

Opus 4.8 now handles requests that Fable 5's safety filters reject, allowing Anthropic to restore its most capable model while adding stronger protections against misuse. Most users simply get Fable 5 back across Claude Code, Cowork, and the web app, while only requests that trigger the new cybersecurity safeguards are redirected to the older model.

Evolving AI: Keeping a character or product consistent across an AI video is the hardest part of the medium. Faces drift, products morph, and the whole thing breaks the moment you go past a single shot.

invideo Agent One is an AI video platform for serious creatives - you direct a crew of agents to make films, ads, and series, not one clip at a time. Its team spent a week, and an absurd number of credits, cracking consistency for real and now they're handing over the exact workflow.

Whether you're a filmmaker building a series or an agency shipping branded ads, this is how you keep everything locked from the first shot to the last.

invideo's Global Creative Head, Hridaye Nagpal, is hosting a live session where he'll build a production-ready character and product workflow inside Agent One, explain the thinking behind it, and answer questions live.

July 2, 7:30 PM IST / 7 AM PT → Save my spot

Evolving AI: Sonnet 5 makes near-flagship autonomy the baseline across Claude's tiers.

Anthropic frames Sonnet 5 as the point where its mid-tier model stops quitting early, and early partners report it finishing jobs past Sonnets abandoned. Testers say it audits its own results unprompted. One asked it to investigate a bug and it wrote a reproducing test, applied the fix, then stashed the change to prove the bug returned. Safety scores beat Sonnet 4.6, and its cyberattack skills stay below Opus. Introductory pricing runs $2 and $10 per million tokens through August 31, beside Opus at $5 and $25.

Sonnet 5 is the default model on Free and Pro right now at claude.ai.

Why It Matters:

Sonnet 5 is the free default in the Claude apps, so the model doing everyday coding, research and long multi-step chores is the one most people already have open. Anyone paying mainly to get tasks finished has less to gain from upgrading now. Most user will see the value proposition from day one.

You’ve had that moment. Walking into a call, scrolling through old emails, trying to remember what you promised. Lindy texts you a brief 15 minutes before: attendee context, past discussions, open items, talking points. All pulled automatically. Try Lindy free.

Try free today

Evolving AI: OpenAI is now measuring the scientific judgment that separates an expert from a novice.

OpenAI's GeneBench-Pro is built from 129 synthetic cases so the underlying truth is fixed and the reasoning itself can be graded. Experts reviewed most of the tasks and judged them realistic. When models fail they miss data artifacts and cannot frame a problem the way a seasoned researcher does. It has posted ten problems on Hugging Face and sent a 50-question set to Artificial Analysis for outside scoring. One task would cost an expert thousands of dollars in time, while the model attempts it for a few dollars.

Artificial Analysis will soon show whether that score holds up away from OpenAI's own lab. If it does, drug work that now needs whole teams of specialists could move quicker and cost less. The models still stumble on the messy parts, so people stay in the loop. The bar here is human expertise, and it is a high one.

QUICK HITS

💸 AWS commits billions to bring AI deeper into government and public sector workloads.

🌍 UN launches a global AI commission to help shape international AI governance.

🚘 Wayve Launches $85M Employee Share Sale at an $8.5B Valuation.

🧠 Nvidia Challenger Etched Hits $5B Valuation With $1B in AI Chip Sales.

 *partner link