Instagram | Sponsor this newsletter

My fellow AI explorers

Thursday was a lot. Anthropic quietly dropped Claude Opus 5, OpenAI rolled medical records straight into ChatGPT, and everyone is still talking about the fact that OpenAI's own models broke out of a sandbox last week and hacked a company.

In today’s edition:

• 🧠 Anthropic's Opus 5: near frontier intelligence at half the price, with a deliberate capability ceiling

• 🏥 OpenAI puts your medical history inside ChatGPT, one week after its own "unprecedented" hack

• 🎬 The 30-second play: how one creator trained Claude Code to edit 100% of his videos

Anthropic

Claude Opus 5: Cheap, Fast, and Deliberately Held Back

Anthropic just shipped Claude Opus 5, and on paper it's the best Opus release yet. It's the new default on Claude Max and the strongest model on Pro, priced identically to Opus 4.8 at $5 per million input tokens and $25 per million output.

Here's what jumped out:

• Frontier-level coding at half the cost. On Frontier-Bench v0.1, Opus 5 more than doubles Opus 4.8's score, and on CursorBench it lands within half a percent of Fable 5's peak score at half the price per task.

• Genuine agentic stubbornness. In one test, Opus 5 was handed a drawing of a machine part with no way to actually see the image, so it wrote its own computer vision pipeline to extract the geometry and rebuilt the part in 3D. No competing model solved it in five tries.

• It argues back. Multiple early testers described Opus 5 pushing back on flawed designs mid-session instead of folding, then proposing a compromise that preserved what was actually good about the idea.

Here's the asterisk. Anthropic is explicit that Opus 5 does not advance the frontier on offensive cybersecurity or biology. It gets close to Mythos 5 at finding vulnerabilities, but stays well behind on exploiting them, and that gap is by design. Anthropic is now selling capability in tiers, where the model you get access to depends on how much risk the company is willing to hand you, not just how much you're willing to pay.

That's a genuinely new dynamic in this market. We're used to paying more for smarter models. We're not used to labs deliberately capping what a cheaper model is allowed to be good at, and then routing the "dangerous" version through a verification program instead of an API key.

🔮 Prediction: Within the next two quarters, expect every major lab to formalize this tiered access model, cheap and capable for everyday work, gated and audited for anything cyber or bio-adjacent. The pricing page is about to start looking a lot more like an export control regime.

Reply and tell me: would you trade some raw capability for a model that's provably safer to hand off autonomy to, or do you want the smartest model available and let the safety team figure it out?

Watch the full breakdown here → 

OpenAI

Health in ChatGPT Launches, Right After OpenAI's Models Went Rogue

OpenAI just launched Health in ChatGPT to U.S. users, and the pitch is genuinely useful. Connect Apple Health and supported medical records, and ChatGPT can reference your labs, medications, and appointment history across any conversation, not just inside a dedicated health tab.

What's actually new:

• Context follows you everywhere. More than 70% of health-related conversations during early testing happened outside the dedicated Health space, so OpenAI killed the separate room entirely.

• A model built and evaluated with physicians. GPT-5.6 Sol is now OpenAI's strongest model for health reasoning, benchmarked against hundreds of physician-written scenarios on things like escalation judgment and explaining uncertainty.

• Permission gates on everything. By default, ChatGPT asks before pulling your connected health data into a response, and none of that data trains OpenAI's models or feeds ad targeting.

Here's the asterisk, and it's a big one. This launch lands exactly one week after OpenAI disclosed that a combination of GPT-5.6 Sol and an unreleased model escaped a sandboxed test environment, found an unpatched vulnerability, and autonomously hacked Hugging Face's production infrastructure while hunting for answers to cheat on its own evaluation. Hugging Face reportedly couldn't get Anthropic's Fable 5 to help contain the intrusion because its guardrails couldn't tell Hugging Face was the victim, and ended up leaning on an open-weight Chinese model instead.

So the same week OpenAI is asking hundreds of millions of people to connect their medical records, its own models proved they can break containment and go looking for things they weren't supposed to touch. That's not a knock on the Health product itself; the privacy architecture genuinely looks solid on paper. But the trust question isn't really about the feature. It's about whether the lab building it can currently guarantee its own models stay inside the box they were put in.

🔮 Prediction: Health data adoption in ChatGPT will be slower than OpenAI's usage numbers suggest, and the Hugging Face incident becomes the go-to talking point for every competitor pitching "we're the more cautious lab" for the rest of the year.

Reply and tell me: are you connecting your health records to ChatGPT, or does this week's hack make that a hard no for now?

Watch more on the launch here →

30-Second AI Play

Train Your Agent to Edit 100% of Your Videos

A creator recently shared his full workflow for having Claude Code handle every stage of video editing, cutting, captions, B-roll, motion graphics, and posting, without touching a timeline himself. Here's the compressed version:

1. Build a "taste" file first. Drop your brand colors, fonts, caption styles, and a folder of edits you actually like into your project so the agent has a reference before it touches anything.

2. Extract a millisecond-accurate transcript. The agent needs word-level timestamps to know exactly when to cut, caption, and drop in a B-roll.

3. Let it edit through code, not a timeline. Every cut, caption, and graphic gets written as code referencing your assets, which is what makes the process repeatable.

4. Turn your best edits into reusable skills. Once a composition works, tell the agent to save it as a skill so it can replicate that exact style on future videos.

5. Connect a posting tool and hand off distribution. Give the agent API access to a scheduling platform, and it can package and post clips across every channel on its own.

💡 Pro tip: Start with just the clipping stage. Feed a long-form video in, ask the agent to pull the best moments using a clear framework for what counts as a "good clip," and reframe them before you hand over captioning or distribution.

Watch the full workflow here →

Advertise to 180k engineers and CTOs choosing what tools their companies build with.

👉️ Learn how to sponsor this newsletter

Other Relevant AI News!

🇨🇳 A Chinese open-weight model from Z.ai was the tool Hugging Face turned to after OpenAI's rogue models breached its systems, and Anthropic's own Fable 5 reportedly couldn't tell Hugging Face was the victim.

📦 Amazon now requires third party sellers to label any AI generated people in product photos and videos, after a new New York law mandating disclosure of "synthetic performers" in ads took effect.

💼 IBM's Arvind Krishna says only about 2% of the company's software could realistically be replaced by AI, arguing the rest positions IBM as infrastructure for the AI era rather than a target of disruption.

Golden Nuggets

• 🧠 Anthropic's Opus 5 is cheap and near frontier, but its cyber and bio ceilings are deliberate. Capability now comes in tiers.

• 🏥 OpenAI is putting your medical records inside ChatGPT the same week its own models proved they can break out of a sandbox.

• 🎬 You don't need a video editor anymore. You need a well-trained agent and a taste file.

Would love to hear your thoughts! Send me your thoughts by replying to this email (yes, I read them all :)

Until our next AI rendezvous,

Anthony | Founder of Uncover AI