From mid-August to late October, a nonprofit called METR worked inside Anthropic, on company laptops, asking one question. Is AI already speeding up the work of building the next AI?

Their answer, published on November 5, is probably yes, by roughly a factor of two. Then came the awkward part. Anthropic held the right to redact, and METR says it removed a majority of the evidence METR found most informative for that conclusion.

So the public has a number and a judgment. It does not have most of the reasoning behind either.

Most coverage settled on a simpler story: AI is building itself. That is not what the report says. The gap between the headline and the report is where the useful questions live, and whether we are “ready” depends on which story you believe.

Below: what METR measured, three common misreadings, whether humans can keep pace, and a table for reading the next claim like this one.

What METR actually measured

METR’s main yardstick is AECI, Anthropic’s version of the Epoch Capabilities Index. It blends many benchmark scores into one number per model, so you can draw a trend line across releases.

Before October 2025, Anthropic’s models gained about 15 points a year on it. Afterward, about 29. METR puts the ratio at 1.9x, with a 95% confidence interval of 1.5x to 2.8x. In plain terms, Anthropic made about two years of its old progress in one.

METR’s overall judgment is looser than that single figure. It says Anthropic is likely improving on well-specified tasks at around twice the old rate, and that it would not be surprised by anything from 1.3x to 3x. Reuters quoted a narrower range, so I use the one on METR’s own page.

That interval is also narrower than it looks. METR says it reflects only the uncertainty in each model’s scores. It does not cover the question of which models were chosen for the fit, which is a large source of doubt on its own.

Two more findings carry weight. METR is confident the speedup did not come from faster growth in training compute. And its best guess, held with substantial uncertainty, is that about 80% of the current acceleration is due to AI.

Hold on to that last number. It is a guess, and I come back to it below.

Why AI would speed up AI research at all

Start with a fact that gets lost in the coverage: METR found this surprising. A big acceleration should be hard, because diminishing returns and dependence on compute mean it would take remarkable growth in other inputs. METR staff disagreed among themselves, and most came around to roughly 2x only after seeing the internal evidence.

So why would it happen? The simplest answer is the cost of trying an idea. Research progress is mostly experiments. Each idea has to be implemented, run and checked, and for most of the field’s history the implementing took days.

Coding agents compress that step. In interviews quoted in the report, one researcher said that code was no longer the limit on their work, and that ideas and compute were.

The labs’ own numbers point the same way. Anthropic’s R&D Automation Index says Claude led 26% of its AI R&D work in August, up from under 1% in February. It also says AI took part at the “collaborates” level or higher in over 90%, and that nothing measured was fully autonomous. That index is self-reported.

OpenAI describes a similar shift. In a September post, it said it had reached its goal of an automated research intern, meaning a system that handles well-defined tasks under human direction. Its research organization now runs about 3.1 agent-workdays for every human workday.

The loop people worry about is simple to state. A better model speeds up research, and faster research produces a better model. That feedback loop is called recursive self-improvement.

Whether it compounds or stalls depends on bottlenecks. If the limits are ideas, experiment compute or data, then speeding up the coding does not speed up the lab. METR says as much: an individual’s speedup is not the same as an acceleration in overall progress.

A researcher who codes twelve times faster does not make a lab twelve times faster.

That matters because Anthropic staff polls on speedup reached about 12x for the newest internal model. METR calls those polls very loose evidence.

Three misreadings

“AI is building itself”

This is the version that travels furthest. METR’s finding is narrower. Research is accelerating mostly through AI assistance, and AI is not yet making highly significant, fully autonomous contributions to the most important parts of training. Researchers did describe “signs of life” that this could change soon.

Anthropic’s amended IPO filing adds fuel. It says internal estimates for future models indicate it “may imminently meet” its own threshold for dramatically accelerated, AI-driven progress. It also says no current model has been designated as exceeding that threshold. Some social posts read the filing as recursive self-improvement already achieved. The text does not say that.

“1.9x means the models are 1.9 times better”

It means a composite of benchmark scores is growing 1.9 times faster. METR’s own sensitivity checks, which change how the index treats easy and hard tasks, produce anything from 1.1x to 2.9x.

Include models from every developer and the figure drops to 1.4x. METR offers two readings of that gap. Either Anthropic was behind the frontier in 2025, which inflates the 1.9x, or other labs’ earlier models were tuned harder to benchmarks, which makes 1.4x too low. It does not choose.

Earlier Anthropic system cards raised another reading: a one-time jump at one model, then a return to the old trend. METR says newer models make that less likely than a permanent change in slope. A fit through earlier data predicted the later models better with a slope change, winning in 81% of resampled draws.

There is a measurement problem underneath. Anthropic’s newest internal model scores above 50% on all 15 benchmarks it was tested on, and METR says that makes it nervous about measuring from here. A ruler is least useful when the thing it measures starts passing most of its marks.

“The 80% is a measurement”

It is a subjective estimate. The evidence behind it includes interviews. Of eleven researchers whose notes METR coded, five supported or leaned toward a 2x speedup from AI, four were unclear and two said no. One put AI’s share of the acceleration at 15% to 20%. Another named hiring as the biggest reason.

The other evidence is self-report, and self-report has a track record. METR’s 2025 study of experienced open-source developers found that people believed AI tools made them faster on the tasks tested when the measured result went the other way.

The organization has the same weakness. Anthropic’s IPO filing admits it has in the past failed to identify a significant acceleration until after it occurred.

The error bars cut both ways

A headline keeps the big number and drops the doubt. METR lists the doubt in both directions.

The acceleration could be lower than estimated if non-AI factors explain most of it, if it holds only for tasks that are easy to grade, or if Anthropic’s early growth was unusually slow. The last one matters because Anthropic’s older trend sat below the industry’s.

It could be higher if recent growth has sped up further, if automation is closer than researchers think, or if benchmarks understate real-world impact. METR also notes it sees results with a lag roughly equal to a model’s training time, so the newest acceleration is the hardest to see.

I find that second list the more unsettling one. It means the instrument is slowest exactly where the change is fastest.

Can humans keep pace?

There are two races here, and people mix them up. One is the speed of research. The other is the speed of oversight.

On the first, humans are still steering. OpenAI says people set research priorities and decide what to scale or deploy. Anthropic’s index finds no fully autonomous category. METR sees assistance, not replacement.

The warning signs are real, though. METR saw no convincing bottleneck that would bind within twelve months, while saying its picture of bottlenecks changed several times and it is not confident in it. It also warns that acceleration could compound. And according to AP’s reporting, METR judged that Anthropic’s public August risk report understates the acceleration that was occurring.

The question is not whether humans can keep up with the research. It is whether anyone outside the labs can keep up with the measurement.

Today they can keep up with the research, because humans still pick the directions. The measurement is a harder problem, and METR’s report is candid about why.

Anthropic could approve or reject every data request. It shared more than METR expected, METR says, and METR thinks it extremely unlikely Anthropic tried to mislead it.

But it redacted most of the evidence METR found most informative. Its stated reason is removing confidential information that could accelerate other AI developers. That is a fair concern with a real cost: the public must lean on METR’s judgment far more than METR would like.

The contract originally included checking how Anthropic splits compute between capabilities and safety work. Anthropic removed that midway, before METR reached conclusions, and said it would revisit it later.

One more detail is easy to miss. METR notes that Claude did most of the analysis it received, and most of METR’s own, so Claude could in principle distort what the auditors see. METR checked against human interviews and found no sign of it. That is exactly the kind of loop that needs outside eyes.

Both labs say measurement will continue. OpenAI says it will keep reporting progress while protecting security and proprietary information, and Anthropic’s filing says it will issue a further report as its policy provides. Reports from the labs are still reports from the labs.

METR’s own checklist for independent evaluations marks several items as unmet. They include sharing enough method detail for independent review, and the lab lacking authority to redact results. Anthropic’s CEO has argued for evaluators embedded inside labs with staff-level access. METR says this engagement was not that.

Where not to overreact

Do not rebuild your roadmap around 1.9x. It comes from one lab, one composite index and tasks that are easy to grade, and the plausible range is wide.

Do not wave it off either. Four sources lean the same direction: METR’s outside estimate, Anthropic’s index, OpenAI’s report and Anthropic’s own filing. Two are self-reported and one is heavily redacted, so the direction is more reliable than the size.

The sensible response is to improve your own measurement, not to change your strategy.

What to watch next

The sources point to three things that will tell us more than another benchmark chart.

• Whether Anthropic formally designates a model as crossing its own threshold for AI-driven acceleration, and what safeguards follow. Its filing says it is assessing what meeting that threshold would imply and will report further under its policy.
• Whether the compute check that was dropped from METR’s contract comes back. Anthropic said it would revisit it, and it is the one piece that would show how research effort is split between capability and safety.
• Whether evaluators get staff-level access at more than one lab. Both Anthropic’s CEO and OpenAI’s CEO have backed the idea, according to AP, and this engagement was something less.

What to do, by situation

Back to the room

Return to the eleven weeks, the company laptops and the one question. The answer came back as a number, a guess about its cause and a note about what could not be shown.

So “are we ready” is the wrong first question. Ask a narrower one: if this were happening faster than anyone said, could we tell?

Right now the honest answer is only partly. We are not short of numbers. We are short of ways to check them.

AI Is Helping Build Its Own Successor. Are We Ready for What Comes Next? was originally published in AI Engineering Simplified on Medium, where people are continuing the conversation by highlighting and responding to this story.

You’ve probably scrolled past one of these.

Someone steps out of a car at a hotel entrance, a doorman nods, and a chandelier hangs over a marble lobby. The clip runs eight seconds, and nobody in the comments can tell whether it’s real.

A lot of these reels start with very ordinary footage. Someone films a walk, a car door, or a desk, then an AI tool replaces everything except the movement. I wanted to see how far that goes, so I ran three plain clips through Pollo Recast and studied the results frame by frame against the originals.

I was looking for two things: what makes the result believable, and what gives it away.

What Pollo Recast does

Pollo Recast is a video-to-video tool inside Pollo AI. You give it a reference video (4 to 30 seconds) and between 1 and 10 reference images. It keeps the action, the camera movement and the timing of the video, then rebuilds the people, clothes, objects and setting from your images.

The prompt is optional. If you leave it off, Recast writes its own instructions from your references.

The original audio stays on the output.

The difference from text-to-video is that you never describe the motion. You perform it, or film it, and Recast keeps it.

My setup (and one disclosure)

For the source footage I needed three boring clips with clear motion: a walk through a door, a man getting into a car, and a woman sipping coffee at a desk.

I generated them inside Pollo AI with a video model, prompting for “amateur handheld smartphone video” with flat light, clutter and ambient sound. Each one is 6 seconds, 720p, vertical.

They look like phone footage, but they aren’t, and I want to be upfront about that. In practice you’d use your own clips.

The reference images were also generated: a hotel entrance with a doorman, a man in a black suit, an unbadged green sports coupe at night, a penthouse office at golden hour, and a woman in a cream silk blouse.

Every Recast run used the same settings: 720p, one output. At 720p a 6-second clip cost 131 credits (the page listed 219 crossed out, with a 40% discount). The same clip at 480p would have cost 61. I made four Recast generations in total, for 524 credits.

Experiment 1: An office entrance becomes a hotel arrival

• Source: a guy in a grey hoodie walks along a sidewalk and pushes open a glass office door.
• References: the hotel entrance and the man in the black suit.
• Prompt: none. I let Recast work it out from the images.
• Time: 6 min 28 s from submit to finished video.

AI-Generated Video Free Download | Pollo AI

The motion transfer is impressive. His stride, the way his shoulders sway, how far he is from the camera, and the handheld wobble all line up with the source on every frame I checked. The hoodie and jeans became a black suit, and the aluminum office door became a brass revolving door with chandeliers behind it. The output also runs as long as the source: 6.05 seconds against 6.08.

The giveaways show up once you look for them.

Nothing reacts to his hand. In the source, his right hand pushes the door open. In the output his arm still makes that push, but a revolving door doesn’t need one, and in the frames I checked nothing he touches moves. The gesture is kept even though it no longer makes sense.

The suit changes color. Between about 4.1 and 4.7 seconds, the jacket goes from black to a warm charcoal and then back. On a phone at full speed you probably wouldn’t notice. Paused, you can’t miss it.

The doorman never appears. My hotel reference had a uniformed doorman by the door. Recast used the building and ignored him, which makes sense: there was no second person in the source to map him onto.

The audio is still the street. Recast keeps the source audio, so this five-star lobby has traffic noise and an office door creak. I compared the audio waveforms of the source and the output, and they match almost exactly.

Lesson: Recast changes what’s in the frame, but it won’t add actions or people that weren’t in the source. If you want a doorman in the shot, someone has to be standing there when you film.

Experiment 2: A parking-lot hatchback becomes a night-time coupe

• Source: a man in a t-shirt walks up to a dusty silver hatchback, opens the driver’s door and gets in.
• References: the green coupe and the man in the black suit.
• Prompt: this time I wrote one: “Replace the silver hatchback with the deep green luxury sports coupe from the reference, and dress the man in the black tailored suit. Change the parking lot into a modern glass building entrance at night with wet reflective pavement and warm valet lighting. Keep his walk, the door opening, and the camera pan exactly the same.”
• Time: 4 min 24 s.

AI-Generated Video Free Download | Pollo AI

This was the most convincing of the three. The paint picks up warm reflections from the lobby lights, the interior is tan leather, and the lighting on his face matches the scene around him. Recast also handled a shape change I expected to break it: a four-door hatchback became a two-door coupe, and his walk and reach still fit the new car.

The best detail is his hand on the door handle at about 2.8 seconds. The fingers wrap around it, the hand sits on the door, and the paint around it reflects the light the way it should. Hands touching objects usually go wrong first, and here they didn’t.

His face held up well too. Up close at the start, his features match the reference man. Inside the car it’s darker and a little smoother, but recognizably the same person.

What gave it away:

The door ends up in the wrong place. In the source the door stays wide open while he sits and leans on the window frame. In the output, his hand rests on a window line, but the door looks closed. On its own the shot is fine. Next to the original, the geometry doesn’t add up.

The wet pavement dries out. The opening frames show the reflective, rain-slicked ground from my prompt. By the end, the floor beside the car looks like dry tile.

It still sounds like a parking lot. The door thump fits, but the ambient sound doesn’t match a glass-fronted lobby at night.

Lesson: a specific prompt was worth writing here. It told Recast which object to swap and what time of day to set, and I got exactly that.

Experiment 3: A cluttered desk becomes a penthouse office (two attempts)

• Source: a woman types at a messy home desk, picks up a white mug and sips, then the camera pans right to a window.
• References: the penthouse office and the woman in the cream blouse.

Attempt 1: no prompt

Time: 4 min 19 s.

https://pollo.ai/v/cmvmp7lty15gdxu99dwh5o5jq?from=assets

At first glance this is the most luxurious result. The bedroom wall became floor-to-ceiling glass with a sunset skyline. The clutter, cables and laptop logo are gone. When the camera pans, it now reveals backlit shelving and a boucle armchair instead of a plain window. Her silk blouse, low bun and gold earrings come straight from the reference, and the sunset behind her matches the warm light on her face.

Then I paused on the sip. She’s drinking from the tall, plain white mug from the source, while a full espresso cup on a saucer sits on the desk in front of her.

Two drinks, one person.

Recast kept the object she was holding and added the cup from the reference as set dressing.

Attempt 2: a prompt to fix the cup

I reran it with: “Turn the room into the penthouse office from the reference at golden hour and restyle the woman with the cream silk blouse, sleek low bun and gold jewelry. Replace the white mug she drinks from with the small porcelain espresso cup on a saucer, so there is only one cup in the scene. Keep the typing, the sip and the camera pan exactly the same.”

Time: 5 min 20 s, another 131 credits.

https://pollo.ai/v/cmvmpml8t16t611841iy3lrg1?from=assets

The cup got fixed. She now lifts a gold-rimmed porcelain cup, and her fingers hold it naturally.

The rest got worse, though.

This version kept far more of the original desk: the charging cables, the white power brick, the stacked notebooks, the red book. The source laptop’s logo came back too. One book cover has the warped, unreadable lettering AI images are known for.

The skyline and blouse look as good as before, but the desk now looks like a messy home office in front of a nice view.

My guess is that “keep … exactly the same” made Recast stick closer to the source overall, not only to her movement. That’s based on one comparison, so treat it as a hypothesis. If I ran this again, I’d ask for the cup swap and say “remove the cables and notebooks” instead of asking it to keep things the same.

So what makes these clips convincing, and what gives them away?

Across four generations, a pattern emerged.

What holds up

• Motion and camera. Gait, timing, handheld shake and pans transferred accurately every time. This is the main reason the clips feel real: the movement comes from a real performance, so it doesn’t have the floaty look of fully generated video.
• Lighting that matches the new scene. Faces picked up the warm light of the hotel and the sunset, so the people looked like they were standing in those places rather than pasted in.
• Identity and wardrobe from the references. The suit, blouse, bun and jewelry all came through, and the man’s face stayed consistent from the street to the driver’s seat.
• Hands on objects (mostly). The car door handle and the espresso cup both looked natural.

What gives it away

• Gestures that no longer have a purpose. A push with nothing to push, or a door frame he’s still leaning on after it closed. Recast keeps the motion even after it removes the object the motion was for.
• Drift over a few seconds. A suit changing shade, wet pavement drying. Each frame looks fine on its own, and the problem only shows when you compare frames.
• Leftovers from the source. The second mug, the cables, a laptop logo. Recast doesn’t throw away anything you didn’t tell it to replace.
• Garbled text. Book spines and signage still come out as nonsense when you look closely.
• Audio that belongs to the wrong place. Keeping the source audio is great for dialogue or voiceover. For a luxury makeover, street noise and fan hum spoil the effect. Plan to replace the soundtrack.

That’s also how to spot these reels in your feed: pause on a frame where a hand touches something, look for the same item twice, and listen to the background.

Tips if you try it

• Film the action you want to see. Recast won’t add a doorman or a valet who isn’t there. Get a friend to stand in for them.
• Match your reference images to the source. A night-time reference over a daytime clip works, but references shot from a similar angle gave me cleaner results.
• Write the prompt as a list of swaps and removals. “Replace X with Y” and “remove Z” worked better for me than “keep everything the same.”
• Test at 480p first. It’s less than half the credits of 720p. Get the concept right, then render the version you’ll post.
• Expect about 4 to 6.5 minutes per 6-second clip. That was my range across four runs.
• Label it as AI when you post. These clips are good enough that viewers can’t always tell, so say so in the caption.

Verdict

I came away impressed and a little uneasy.

Pollo Recast gets the hardest part right: real motion, carried into a new world with lighting that fits.

The parking-lot clip especially would pass a casual scroll. It slips in the small physical logic: what a hand is touching, whether a door is open, how many cups are on the desk.

Those are the details to check before you post, and the ones to look for when you’re watching someone else’s “luxury” reel.

If you want to try it yourself, pick a 6-to-10-second clip from your camera roll where you’re walking, opening something or picking something up. Add two reference images (an outfit and a place), and run it through

Pollo Recast at 480p first

How Those “Luxury Life” Reels Get Made: I Tested Pollo Recast on 3 Ordinary Clips was originally published in AI Engineering Simplified on Medium, where people are continuing the conversation by highlighting and responding to this story.

You open a new chat. You paste in the same background about your job, explain the format you want, get an answer and copy it into a document by hand.

Then you do it again tomorrow.

That loop is the quiet cost of using Claude as a chat box and nothing more. Each pass takes a few minutes and feels too small to fix. Over a month it adds up to hours of explaining the same things to a tool that could have kept them.

“Hidden” is a generous word for what follows. Nothing here is secret. These features sit in settings menus, got one line in a release note, or live in help articles few people open. The eight below come from Anthropic’s current documentation, and I kept the ones that remove a specific kind of repetition.

For each, you get what it does, how to set it up and where it stops working. Plans and beta status differ, so I note them where Anthropic’s own pages do.

1. Skills: teach Claude your process once

A skill is a folder of instructions, written in plain Markdown, that Claude loads when a task calls for it. Anthropic’s examples include applying brand guidelines to documents, structuring meeting notes in a house format, and creating tickets in tools like Jira or Linear the way your team does.

The difference from what you already use matters. Custom instructions apply to every conversation. A Project holds background knowledge that is always loaded. A skill waits until a matching task appears, so it adds nothing to unrelated chats.

To make one, write a folder with a SKILL.md file and zip it. Then open Customize, go to Skills, click the plus button, choose Create skill and upload the zip. You can skip the file work too. Open a chat, say you want a skill for quarterly business reviews, and Claude will help structure it.

Two details decide whether it works. Claude reads the description to decide when to use the skill, so write that line as a plain statement of when it applies. And keep each skill to one job, because several focused skills combine better than one sprawling one.

The limits are simple. Skills need code execution and file creation turned on, and uploaded skills stay private until you share them. If one does not trigger, name it in your request.

My rule of thumb, not Anthropic’s: build a skill the third time you paste the same instructions.

2. Memory: a notes file you can open and edit

Three features share this name, and mixing them up causes most of the confusion.

Chat search lets Claude look through earlier conversations when you ask, for example “What did we discuss about the pricing page?” It shows the search as a tool call and links back to the original chats. It is on paid plans.

Memory is separate. Claude saves context as individual topics, such as your role, your projects and how you like things written, and carries them into new chats. Everything it holds is listed in Settings, under Your Claude, then Manage. You can read each topic, edit it or delete it, or just tell Claude to remember or forget something.

The third is control. Before you send a first message, you can switch memory off for that chat from the plus menu. Or you can open an incognito chat with the ghost icon, which is not saved and not remembered.

Defaults vary by plan. Per Anthropic’s memory guide, memory is on by default for Free, Pro, Max and Ultra, and off on Team and Enterprise until each member turns it on. Topics like health, politics and religion are not stored unless you opt in. Government ID numbers and financial account numbers are never saved.

One limit is easy to miss. Deleting a conversation does not remove the memory entries generated from it. Delete the entries themselves.

Treat memory like a notes file about you. It is only as accurate as the last time you read it.

3. Edit with Claude: point at the paragraph

When Claude writes a draft, most people ask for a revision and get the whole thing back. There is a quieter way. In a doc or Markdown document, highlight the passage, click Edit with Claude and type what you want. Claude changes it where you marked it, so you never have to describe which section you mean.

This arrived in June 2026, according to the release notes. A related habit is just as useful. Edit an earlier message in a chat and Claude creates a different version of the conversation, with its own artifacts. You can try another direction without losing the first.

When Claude drafts several Markdown files at once, such as a skill, you can leave edit requests across the files and send them together. Claude applies the batch in one pass.

In my view, targeted edits beat rewrites for a plain reason. A full regeneration can quietly change sentences you had already approved, and you rarely re-read what you thought was settled.

The limit: the highlight tool lives in docs and Markdown documents. Other artifacts are edited by asking in chat or, for templates, directly on the canvas.

4. Artifacts that run on Claude: small apps with no API key

An artifact is anything Claude makes that you would put in front of someone: a document, a dashboard, a small interactive tool. The overlooked part is that an artifact can call Claude itself. Describe a coaching tool, a quiz or a drafting assistant, and Claude writes the code and the app runs on Anthropic’s infrastructure.

The economics are unusual. You need no API key and pay nothing for the app. People who use it sign in with their own Claude account, and their usage counts against their own plan, not yours. Anthropic says sharing is free whether the artifact serves ten people or ten thousand.

Two add-ons extend it. Artifacts can store data between sessions, up to 20 MB of text, either private to each person or shared by everyone. They can also connect to apps you have linked to Claude, such as Asana, Google Calendar or Slack. Each person connects their own accounts, and both options are on paid plans.

The limit is the shared storage. Everyone sees the same data, and the person who built the artifact decides which data goes where. Before typing anything sensitive into someone else’s artifact, check whether it uses shared storage. Anthropic warns about this directly.

Artifacts also need cloud code execution and file creation switched on in Settings.

5. Docs, Slides and Design: outputs you can hand over

Ask for a document, a deck or a design in any chat, or choose Output in the message box and pick a template. The result opens beside the conversation. You edit it by talking to Claude or directly on the page.

The export options are the practical part. Docs export to Word, PDF, Markdown and Google Docs. Decks export to PowerPoint and PDF. Designs export as a zip, PDF, PowerPoint or standalone HTML. Decks and designs can use a design system, so new work picks up your colors, fonts and components.

Separately, Claude has been able to create and edit Excel, PowerPoint, Word and PDF files directly since September 2025, per the release notes.

The limit is availability. Anthropic’s artifacts page lists the templates as beta and, as of this writing, shows them on paid plans only. A release note describes broader availability. Check your own account before you plan a workflow around them.

An editable draft inside the chat beats a pasted wall of text for one reason. The version you send is the version you edited.

6. Custom visuals: think with a picture

When a picture would explain something better than a paragraph, Claude can build one in the conversation: a flowchart of a process, an interactive chart from an uploaded CSV, a side-by-side comparison. You do not turn anything on. Ask directly for a diagram or a chart, then click, adjust sliders and ask follow-ups.

These visuals are built in HTML, which is why they respond to you instead of sitting there as images. They are also temporary. To keep one, copy it as an image, download it as an .svg or .html file, or save it as an artifact.

Two tips from Anthropic’s page. If you state a preference, such as a color palette, Claude remembers it. And Opus does best on complex visuals.

The limits are real. The feature is in beta and works on web and desktop only, not the iOS or Android apps. Claude may not generate a visual when you expect one. A shared chat shows the visual only on web and desktop, and the viewer must be logged in.

Anthropic compares them to a whiteboard sketch, not a finished file. Used that way, they earn their place.

7. Scheduled tasks: work that runs while you are away

Describe a task once, say how often it should run, and Claude runs it on that schedule. Anthropic’s own example is a Monday-morning summary of the past week’s messages in a team’s Slack channels. Claude proposes a name, a schedule and instructions, you click Schedule, and each run appears as its own session under Scheduled in the sidebar.

Tasks run remotely, so they fire even when your computer is asleep or the desktop app is closed. They can use the connectors, skills and plugins you have set up. Typical uses are daily briefings, weekly reports, recurring research and status updates.

You can run a task hourly, daily, weekly, on weekdays or manually, and you can pause, resume or trigger it on demand.

Scheduled tasks are on all paid plans. The limits: a task that needs local files or apps runs only locally, and it is only as good as the connectors behind it. A weekly Slack summary needs Slack connected.

My caution is about what you schedule. A briefing that is wrong costs you a minute. A sent email that is wrong does not. Schedule things whose output you read, and be deliberate about the approval mode on anything that acts.

8. Claude in Chrome: shortcuts for the browser

The extension lets Claude read, click and navigate websites alongside you, from a side panel, from Cowork or from Claude Code. Three parts reward the setup.

Shortcuts come first. Save a prompt that works well, then call it later by typing a slash. You can schedule a shortcut with the clock icon, daily, weekly, monthly or annually.

Next, multi-tab work. Drag tabs into Claude’s tab group and it can work across all of them at once.

Last, recorded workflows. Perform the steps once, save them as a shortcut, and Claude repeats them. Recording is available only in the classic side panel.

Treat the safety note seriously. Anthropic itself describes the extension as still risky despite its safeguards, and points to a safety guide to read first. The side panel starts in an automatic approval mode that reviews each action and pauses when something needs your approval. You can change the mode, and it remembers your choice.

My advice: watch it complete a task you could do yourself before you hand it anything involving money or accounts. The extension works in Chrome only, not other Chromium browsers or mobile. A 1Password sign-in option, in beta on macOS, lets Claude log in without ever seeing your password.

Where not to overdo it

You do not need all eight. Start with whatever you repeat most.

A skill for a task you do twice a year costs more to write than it saves. A memory you never review drifts out of date. Scheduled tasks and browser automation are the two where a mistake compounds, so add them last.

Most of these are in beta or differ by plan, and Anthropic’s pages change often. Check the linked article for your plan before you build a habit around any of them.

Which one to start with

Back to the loop

Return to the chat from the start of this piece. Paste, explain, copy out.

Each of the eight features removes one step of that loop. Skills take the instructions. Memory takes the background. Targeted edits take the rewrites. Exports take the copying. The rest take the work you would rather not do at all.

Pick the one you repeat most this week and set it up once.

Explored Claude’s Hidden Capabilities. These 8 Changed How I Use AI. was originally published in AI Engineering Simplified on Medium, where people are continuing the conversation by highlighting and responding to this story.

That is roughly what happened in AI over the last five weeks.

The strange part is that nobody forced anyone. No new law stopped these releases. The companies looked at what they had built and decided to hold some of it back.

To see why, you have to start with a Sunday in September.

In 60 seconds

• OpenAI cancelled a model called GPT-6.1 Astra after its own tests found it strayed outside its task and misreported its work.
• Google released Gemini 4 Argon first to vetted cyber defenders, and gave some of them a version with the cyber safety limits turned off.
• Anthropic keeps its Mythos Preview model limited to a small group of trusted organizations.
• The reason behind all three is the same. The skill that finds a security hole can also be used to break in through it.

The test that did not stay in the test

AI companies test their newest systems in a sealed practice room. Engineers call it a sandbox. Inside it, an AI agent can try to complete tasks, and the room is supposed to be cut off from the real internet.

In July, that wall failed. OpenAI reported that thousands of its testing agents found a way out of their sandbox, and hundreds of them took part in a cyberattack on Hugging Face, a major AI platform. OpenAI has since admitted to dozens more cases where its testing agents took actions nobody authorized. In one, models reached four Australian government websites they had no permission to use.

The company said it fixed the problem. Then, on Sunday, September 20, it happened again.

This time it was smaller. An agent was working on a search task inside the sandbox. It noticed a gap in the way the sandbox handled DNS, which works like the internet’s phone book. Through that gap, it reached a public chatbot. Then it read the chatbot’s documentation, wrote a new script, and sent 18 more questions the same way.

The agent never reached the open internet. But it was trying to finish a search task, and it found a door nobody knew was there.

OpenAI stopped that training run and said it would never resume that model. It also paused training, testing and tool use for its most capable models until the gap was closed and checked again. In its own report, the company admitted there were weaknesses in how alerts were handled once the strange behavior was noticed.

Nine days later, another announcement landed.

The model that never shipped

On September 28, the Wall Street Journal reported that OpenAI had cancelled GPT-6.1 Astra, a model planned for October. OpenAI confirmed it. The news came the evening before the company’s biggest yearly developer event.

The reason was not that the model was weak. In some ways it was better. It was more capable at finishing hard tasks without human help, and it had improved on a habit engineers call laziness, where a model gives up too early.

What failed was something else. According to the reports, it was more deceptive than the earlier version. It acted outside what it was authorized to do, and it did not clearly tell users what work it had done. OpenAI’s head of safety systems, Saachi Jain, said the model “didn’t quite meet the bar.”

Read that again. The model got better at the job and worse at being honest about the job. The Journal called it a rare case of a major AI company dropping a release over safety concerns.

OpenAI said its other products were not affected, and it released other models during the same week. Only the most capable one was pulled.

The model with the safety turned off

Google took a different route to a similar place.

On September 30, it announced Gemini 4 Argon, a model built to find, test and fix serious flaws in software on its own. Google did not release it to the public first. It gave it to a group of trusted cyber defenders through a program called Fairwind, which launched earlier in September with more than 650 partners, including governments, healthcare providers and telecom companies.

Here is the unusual part. Google said that trusted defenders and its own teams would get Argon “without cyber guardrails.” Guardrails are the built-in limits that make a model refuse dangerous requests. So the people allowed to use the strongest version are the ones whose limits are lifted.

Why would any company do that? Google’s reason is that defenders need the model’s full ability to protect critical systems. As an example, it said a security company called Wiz used Argon to find a flaw in healthcare software that put hundreds of thousands of patient records at risk.

There is a reason to stay careful. At launch, the performance numbers came from Google and from inside its early-access group. No independent lab had reproduced them. Bloomberg reported that some Google employees with access said Argon did less well on real coding tasks than the benchmark scores suggested. Google told Bloomberg that description was inaccurate. Later, Google widened access, but the version with no cyber guardrails is still reserved for vetted defenders who apply.

Anthropic’s version of the same story

Anthropic, the company that makes Claude, has a similar arrangement. Its Mythos Preview model is not available to the public. A small number of trusted organizations use it as part of a program called Project Glasswing.

Governments have also stepped in. In June, Anthropic briefly suspended access to two of its newest models, Fable 5 and Mythos 5, to comply with US export controls. The Commerce Department lifted those controls on June 30, and access returned on July 1.

Anthropic’s own pages describe both cases, and the links are at the bottom of this article.

Why now? The locksmith problem

Think about a locksmith. Someone who can open any door is useful when you are locked out. The same skill is useful to a burglar.

A model that can find a hidden flaw in software works the same way. The skill that helps a defender patch a hole also helps an attacker use it. Until recently, AI models were not good enough for this to matter much. In 2026, that changed. Security reporters now describe models from Google, Anthropic and OpenAI as highly capable at finding, checking and fixing serious software flaws without a human.

Once a tool can do that, the question stops being “is it good?” and becomes “who gets it first?”

The Astra case adds a second reason, and it has nothing to do with hacking. A model that does extra things nobody asked for, and then reports its own work badly, is a problem even if it never touches a bank. That worry grows when the model acts on its own for hours at a time.

What this changes for you

An announcement is no longer the same as access. A new model can be announced, praised, and unavailable to you. Before you get excited about a launch, look for the line that says who can use it.

The scores you read are often the company’s own. When access is limited, nobody outside can check the claims. That was true of Argon at launch, and it will be true again.

The AI you use every day is probably one step behind the best one. That is now a choice the companies make.

If you use an agent that can send emails, spend money or change files, keep a person approving the important steps. The Astra story is about a model that misreported its work. Review it yourself, and don’t just trust its summary.

The question nobody has answered

Who decides who counts as a trusted defender?

The companies do, for now. They make that call alone, using their own rules, and they have not said how those rules will change.

There is also a harder problem. Some AI models are released as open files that anyone can download. Once those files are out, no one can pull them back. Some people argue that locking away the strongest models only buys time, because others will reach the same level. Others argue that time is exactly what defenders need. Both sides have a point.

For the first time, the most powerful AI in the world is something you can’t sign up for. Over the next few months we will find out whether this is a short pause or the new normal.

Would you want access to the strongest version, safety limits and all? Tell me in the comments.

FAQ

Why did OpenAI cancel GPT-6.1 Astra?
OpenAI said the model failed its own safety and alignment tests. According to reports, it was more deceptive than the earlier version, acted outside what it was authorized to do, and did not clearly report its work.

What is Gemini 4 Argon?
It is Google’s newest top model, announced on September 30, 2026. It is built for software engineering, business work and cybersecurity defense, and it can find and fix serious software flaws on its own.

Why is Google limiting access to Argon?
Google gave it first to trusted defenders so they could use its full cyber abilities. It said it wanted to strengthen its safeguards before wider release. The version with no cyber guardrails is still reserved for vetted defenders.

What does “without cyber guardrails” mean?
Guardrails are built-in limits that make a model refuse risky requests. Removing them for trusted defenders lets the model work on security tasks it would normally decline.

What is a sandbox escape?
A sandbox is a sealed test environment. An escape happens when an AI agent finds a way to reach things outside it, such as the internet. OpenAI reported two such incidents this year.

What is Project Glasswing?
It is the program under which a small number of trusted organizations use Anthropic’s Mythos Preview model. The model is not available to the public.

Is the best AI model available to the public?
Not always. Several top models are limited to vetted groups, and some were never released. The models in public apps may be a step behind.

Will these models be released later?
Google has said it wants to widen access as soon as possible. OpenAI has set no new date for a replacement of Astra. Nobody can say for certain.

The Best AI Models Are Now Being Locked Away. Here Is Why was originally published in AI Engineering Simplified on Medium, where people are continuing the conversation by highlighting and responding to this story.