Anthropic has released a new Claude model, Opus 5, bringing it close to the capabilities of Claude Fable 5 at half the price. The new model produces the strongest work performance the company has released, with even more gains in its coding capabilities. 

In a fight for AI supremacy, Anthropic joins Google and OpenAI, which earlier this month released new AI models — all of which are more efficient and better at certain tasks. While Fable 5 remains Anthropic’s most “ambitious” work, the latest model excels in enterprise with its coding capabilities and more. 

Anthropic says Opus 5’s strengths lie in knowledge work, autonomy and biology. It’s capable of working more independently at completing tasks without user input and with less back-and-forth than previous models. It can also now double-check its own work and recover from errors as it goes along.

There’s been a boost in the quality of scientific research with Opus 5 over Opus 4.8. Those improvements are seen most in organic chemistry tasks, like inferring molecular structures from spectroscopy data, according to Anthropic.

You can now adjust effort levels for the model as well. There’s a “dial” that can be adjusted to indicate how hard AI thinks about a specific problem or task. There are also new safety and security guardrails in place, which Anthropic calls the “most secure model yet and the hardest to trick into doing harmful things.”

While the model is not specifically designed with cybersecurity in mind, its safeguards make it more useful for tasks such as secure coding and finding vulnerabilities.

New fast mode and more

Claude now has a research preview for an Opus 5 fast mode. Anthropic says this mode provides Full Opus 5 Intelligence with 2.5x faster output token generation at 2x standard Opus 4 pricing. The mode is also available via Claude Code with extra usage credits. 

Anthropic is introducing a “fallback model” for when you prompt something that trips a strict safety filter, only to block the request. For these instances, the system will automatically switch to another model in an attempt to provide an answer to the prompt. You can choose what model to fall back on in these scenarios.

There’s also a new option for developers that will allow them to change the tools Claude can use midconversation without invalidating the prompt cache. This means each phase of an agent’s work is only exposed to the tools it needs to perform its task.

Blake has over a decade of experience writing for the web, with a focus on mobile phones, where he covered the smartphone boom of the 2010s and the broader tech scene. When he’s not in front of a keyboard, you’ll most likely find him playing video games or watching horror movies.

See full bio

Apple’s plan to buy memory from a blacklisted supplier is receiving some pushback from Micron, which claims it will destabilize the U.S. tech industry.

In June, Apple petitioned the Trump administration to allow it to buy Mac RAM chips from a supplier blacklisted in the United States. Now, a U.S. memory producer has urged President Trump not to give in to Apple’s request.

According to the Wall Street Journal on Friday, Micron lobbied the White House on the matter. Micron CEO Sanjay Mehrota and others met with Commerce Secretary Howard Lutnick and others, saying the move to allow sales from blacklisted Chinese companies to U.S. tech companies will be incredibly harmful.

To Micron, it believes that permitting such sales would harm the U.S. tech manufacturing industry in the same way that China decimated U.S. steel and manufacturing plants.

Blacklist sales

The problem revolves around the ongoing memory chip crisis and reducing the cost of components. Apple wants to secure supplies of memory at more reasonable prices, but like all other manufacturers, it has to fight for a limited manufacturing capacity, which has spiked memory prices.

June’s petition explained that Apple should be allowed to buy memory from CXMT, a company on the Chinese Military Company Blacklist, or the 1260H list. The list is made up of companies the Pentagon believes have links to the People’s Liberation Army, and so could be considered a risk to U.S. national security.

Apple isn’t explicitly banned from buying from CXMT, and it reportedly has for small-scale testing. However, the Defense Department is not able to make agreements with companies on the list, nor use products or services from third parties that use components from the list.

In effect, Apple would immediately lose sales to the Defense Department, and others who also monitor the list.

There’s the further issue of CXMT potentially being placed on the “Entity List” in the future, which would block all trade with the supplier. That would also impact Apple, as it would have to find another supplier if it had to stop using CXMT.

While it’s safe to say that Micron has a financial incentive to keep Apple buying its memory chips, it does at least offer an alternative route forward. Albeit not one that will fix things anytime soon.

Micron told the White House that the memory pressure could be alleviated by building more domestic plants faster.

This is basically the same as the general belief in the industry that the situation won’t be fixed until production capacity increases. That is a slow process, as that means getting more factories and production lines online and up to speed.

Micron insists that it plans to spend $250 billion in the U.S. to boost capacity.

However, Micron is framing it as a means to increase investment in the United States, as well as increasing U.S.-based manufacturing. These are two topics that President Trump has openly discussed improving during both his terms in office.

Apple, meanwhile, has gone on the offensive. The report claims Apple has argued that Micron’s gross profit margins of 80% are evidence of price gouging.

Apple also reportedly said that Micron isn’t reinvesting fast enough to meaningfully increase supply. That improved capacity will also have little effect to the supply chain, as the vast majority will be allocated to AI customers.

Micron, meanwhile, believes that Apple and others squeezed suppliers during downturns, which laid the groundwork for the current shortages to take place.

As it stands, it’s a problem that has two possible routes for President Trump to take. He could side with Apple for a more immediate short-term solution to the problem, but the U.S. manufacturing investment proposal will be a very tempting option for Trump too.

E-commerce platform Shopify wants to make reading source code a thing again. And it has an unlikely ally for its back-to-the-roots approach: AI agents. 

The company is launching a new storefront theme for its customers, aka “merchants,” that it claims is completely understandable by anyone with even a smidgen of HTML knowledge. It is designed to be simpler for both people and machines to digest and hack against. 

Shopify hasn’t named this new theme yet, but is unveiling it in developer forums as the probable successor to Horizon, its current JSON-heavy base theme, which customers use to customize their own Shopify pages. 

The new code is mostly HTML, interpolated by the company’s own block-based Liquid templating language, which is also fairly easy on the eyes. 

When merchants peek into the templates directory of this new theme, they will see plain-text files of easy-to-parse code, not long strings of inscrutable JSON. This theme has 93% fewer lines of code than Horizon. 

The company’s increasingly popular AI service is driving this rearchitecture redesign. 

“Every leading model already understands HTML. It is expressive, local, and token-efficient,” wrote Ben Sehl, Shopify product director for storefronts, in a missive on X. And Shopify’s Liquid template language meshes perfectly with HTML. 

From merchants to merchandisers

In his post, Sehl recalled a decade ago, when he was a merchant himself, and Shopify only offered a basic template. It was easy for Sehl, then a novice web coder, to modify with a third-party starter kit. 

Over time, Shopify enriched the templating, allowing users to become full-fledged designers. The Online Store Editor gave them more control over how a page and its various “block” components would look like. Every button and knob could be customized.

Flexibility begets complexity, however. The user template instructions were converted to JSON. “That trade-off gave merchants far more control, but it came with a developer-experience cost: you could no longer understand a page by reading one file,” Sehl wrote. 

“The moment templates became JSON, they stopped being a great developer surface. They became an auto-saved output,” Sehl wrote. 

Good for the bot, good for the human

Recently, Shopify added an AI assistant called Shopify Sidekick, which gave users even more control over how they designed their pages. No longer would they have to remember that #0000FF means ‘Blue’ in HTML speak. They just “declare” the background color to be blue and the agent will make it so. 

This year, 20% of Shopify merchants are using Sidekick to edit their themes, making 25 million edits this year alone. 

Consequently, Shopify faced the task of simplifying things for the AI agent itself. 

It turns out that agents want the same things as humans: code that is easy to read, explicit contracts, and constructive feedback. So the company has reorganized its theme architecture using these qualities as first principles (while maintaining existing Liquid themes in a “forever API”).

While these changes better support our new AI overlords, the underlying code also became more readable to humans as well. 

“The deeper problem with serialized configuration isn’t that JSON is inherently bad. It’s that configuration has a constrained vocabulary,” Sehl explained.

Reveal source code

For this new base theme, Liquid gained a new syntax and parser, built with some additional discipline to execute more complex instructions on the back-end. What the user sees, however, are composable, typed blocks that fit easily alongside ordinary HTML.

In this new approach, each user-designed theme is given a special directory and files with instructions and pointers of how additional artifacts should be generated. Furthermore, an expanded doc tag contains examples and instructions on handling contracts and snippets.

Key to the new theme is a new composable block tag, which contains a set of nested parameters describing page components that the developer can feed with values, configurations, overrides, and other instructions. 

React devs will recognize parameters as doing a similar job to React’s props, but without the need to create a virtual DOM.  Also familiar to React users will be the new partial primitive, which allows a specific region of a page to be updated without needing to re-render the entire page. 

The team also added standard actions, a collection of event triggers for the site, such as, say, updating the shopping cart. “This one primitive will let us remove thousands of lines of reactivity code from Horizon,” Sehl wrote. 

Shopify also developed 20 new rules to ensure generated themes meet the company’s policies around contracts, structure, validation, complexity, nesting, and file-size limits.  

And if the user wants to get fancy, Liquid now allows for logical capabilities such as Boolean expressions, infix operators with precedence, and literal arrays and objects. At some point in the future, the Tailwind CSS framework will even be supported.  

All these improvements should simplify things for the merchant at home, trying to assemble a customized storefront.

For instance: In a declarative, settings-driven architecture, the developer must anticipate how many buttons they plan to use before they start the layout. With HTML, when the merchant discovers they need additional buttons, they just wrap them all in an HTML div tag.

Palatable for AI

AI experts are slowly coming around to the idea that the best diet for AIs is not dense, symbolic-heavy code — at least for non-coding tasks. Like humans, LLMs seem to do better on a diet of more easily-understandable prose (LLMs were, after all, trained on human language). 

One developer, for instance, found that SQL was easier to process compared to the Domain Specific Languages (DSLs) of 17 different tools. For prompting, Anthropic recommends the semantically-rich XML, whereas a decade ago that format was largely cast aside by the industry for the leaner, less-verbose JSON. 

Perhaps Shopify’s efforts to appease AI may lead not to more slop but to a movement for more readable code. ®

ChatGPT is experiencing a major outage, and users are unable to load chats, including previous conversations.

The outage started at approximately 5 AM ET and is affecting users worldwide, including those in the US and Europe.

If you are affected, ChatGPT will get stuck at loading animations for the sidebar, and you won’t be able to send messages due to “too many concurrent requests” errors.

This outage also affects OpenAI’s coding platform, Codex. Thankfully, OpenAI is aware of these issues and has already acknowledged them on the status page.

“We are investigating the issue for the listed services,” the company said in an update posted at 5:30 AM ET.

OpenAI says it has applied a fix and is monitoring the situation, but we continue to run into issues in our tests.

The outage also affects the OpenAI API, with as many as 12 API endpoints listed as having issues on the company’s status page.

Update 1: ChatGPT has fully recovered and is back online as of 6 AM ET. The outage affected ChatGPT and API services for almost 50 minutes.

Security teams log 54% of successful attacks and alert on just 14%. The rest move through your environment unseen.

The Picus whitepaper shows how breach and attack simulation tests your SIEM and EDR rules so threats stop slipping by detection.

Advertisement

Get the whitepaper

More than once, Leonardo threw up a generic error pop-up with no explanation of how to fix the problem. You’re forced to dismiss the message before doing anything else, which feels like hitting a dead end before trying to troubleshoot yourself or looking up the error code in Leonardo.

Connectivity, luckily, is super simple and seamless. The Romeo connects to your computer over Wi-Fi instead of Bluetooth, which sounds like a minor detail until you’ve wrestled with the quirks that Bluetooth can sometimes pose (looking at you, Design Space). During my testing of the Explore 5, I’d occasionally lose my Bluetooth connection, requiring me to disconnect the machine from my computer, reconnect, and try again. I’m not sure if that’s a Design Space issue or my laptop juggling too many Bluetooth devices, but it can be a headache. Through my testing, I never once had the Romeo machine unexpectedly disconnect.

Making Space

There’s no getting around it: The Romeo is big—it’s got twice the cutting width of the Explore 5. It’s not the type of machine that sits easily on a closet shelf when you’re not using it. Most people will want to dedicate a table to it. That’s not necessarily a downside, but it’s intended to exist like a permanent fixture in your workspace. (And remember to account for clearance behind the machine, since materials feed through the rear!). But it is sturdy and built well; it never shifted on the table while cutting, there was very little vibration, the carriage moved seamlessly with no rattles, and nothing flexed or creaked.

Its open design is another trade-off. Unlike the Explore 5, which folds up, everything stays exposed on the Romeo. I haven’t noticed any buildup yet, but I do worry about pet hair and dust getting into the internals, and it’s something that I’ll keep an eye on over time.

When it cuts, the Romeo bleeps and chirps like a dial-up internet, while the Cricut hums and whirs more like an old inkjet printer. Neither is particularly loud, but worth noting if you share a workspace.

Time to Graduate?

Photograph: Nicole Kinning

After spending time with both machines, my biggest takeaway is that the Cricut is designed to remove as many decisions and points of friction as possible. The Romeo hands you the controls and expects you to learn how to use them.

A power line went down outside of Washington, DC, this week. Normally, the grid would only need a few seconds to recover from such an event. But this one took more than 10 minutes because more than 3 gigawatts of data centers stopped drawing power nearly simultaneously.

The event caused voltage across the PJM grid to spike from Northern Virginia to Chicago, according to data collected by Ting Labs, a startup that runs an IoT sensor network out of people’s electrical sockets.

The event didn’t cause a blackout, but it did cause lights across the region to flicker. The incident demonstrated the effect that data centers can have on the grid — an outcome that experts believe will become more frequent.

Northern Virginia, which is in PJM’s territory, is home to the highest concentration of data centers in the world.

“It’s the canary in the coal mine,” Ricardo de Azevedo, CTO at ON.Energy, told TechCrunch. These sorts of events involving large loads like data centers are “happening more and more,” he added.

The event echoes one that happened two years ago, also on PJM’s grid, and it could foreshadow larger events if data centers aren’t built to more elegantly handle disruptions to power supplies. The PJM Interconnection manages grids from New Jersey to Illinois and serves 67 million customers, making it the largest grid operator in the United States.

When the power line went down this week, it triggered data centers to switch to backup power, and about 3.1 gigawatts of load vanished in about 30 seconds, according to PJM data. The grid appeared to recover somewhat, but a short time later additional loads dropped off. At its peak, PJM’s grid had an extra 3.49 gigawatts of electricity on it. It took another 11 minutes before it stabilized. The disconnected data centers represented around 3% of total demand on PJM at the time, according to Reuters.

A few percent may not sound like much, but the electrical grid needs to operate in a state of near-perfect balance, with supply and demand closely matched. If they don’t, voltages can sag or spike. The grid and devices connected to it can tolerate small fluctuations, but if those fluctuations grow too large, they’ll trigger failsafes within the grid or within individual facilities, causing them to disconnect. 

When data centers in Northern Virginia sensed the fluctuation caused by the failed power line, they switched to backup power, which removed their load from the grid. As more data centers made the switch, they removed greater amounts of load from the grid. What started as a relatively small drop in supply became an even larger drop in demand, sending supply surging and causing light bulbs to flicker. 

Most data centers make decisions in a split second, and those that disconnected this week appear to be no different. When the voltage dip reached them, they all decided to disconnect within a few seconds of each other, Ali Zain Banatwala, senior market models specialist at the Independent Electricity System Operator, told TechCrunch. 

“We need to figure a way for these loads that are located next to each other to sequentially either disconnect or reconnect,” he said. A more orderly process would allow grid operators to develop more robust procedures in advance.

Alternatively, data centers could be built to absorb disruptions and not turn their backs to them. One startup, ON.Energy, has been working on a product to help data centers — and the grid — ride through events like the one that occurred this week. 

The company has developed an uninterruptible power supply for an entire data center campus, covering not just servers but also chillers and other equipment. The company essentially hides the data center behind a bank of batteries connected to sophisticated power conversion equipment. All the grid “sees” is one consistent, well-behaved load rather than the peaks and valleys from each individual part of the data center. ON.Energy’s system allows data centers to ramp computing workloads up and down, including AI training, without bothering the grid. 

Perhaps more important, it also means that data centers can absorb power fluctuations from the grid. Rather than disconnecting from the grid, ON.Energy’s system can use any extra power to charge its batteries, and if the flow dips, the system can dispatch power to servers. Plus, it can follow the grid’s lead within milliseconds, preventing sags or surges like the ones that caused this week’s problem for PJM.

ON.Energy is currently installing a total of 3 gigawatts worth of its systems at four different data center campuses, de Azevedo said.

Grid managers have also woken up to the problem.

ERCOT, for example, is going to require large loads like data centers to “ride through” disruptions, de Azevedo said.

The clock is ticking, though. The mass disconnection this week was twice as large as a similar event in 2024, when 60 data centers simultaneously disconnected, pulling 1.5 gigawatts of load from the grid. Back then, data centers accounted for about 6% of PJM’s load, according to Synapse Energy Economics. By 2040, they are expected to make up 24%. If the problem isn’t addressed soon, things could get a lot worse.