Last week I helped a friend debug a problem: Claude Code on his machine asked him to log in the moment it started, and after he logged in, it returned a 401.

It took me five minutes to spot the issue. The project gateway address his company gave him was correct, but he had put the key in the wrong variable. A one-word difference, and he spent a whole night going in circles between 401s and “please log in.”

The short version: connecting to a gateway comes down to two environment variables. But behind those two variables is a pile of pitfalls you only learn about by stepping in them, especially if you use a relay service, a company-wide gateway, or a proxy.

I’ve put together the pitfalls from that night, plus the error-reference table from the official docs, into this article.

Read more

Zuckerberg once mocked Apple() for not inventing anything significant since the iPhone. Now he has posted a photo of AI at work on a Mac, thanking Apple and its new CEO John Ternus for their support.

What brought the two sides together is an 816-billion-parameter large model and a shared goal of putting it on users’ desks.

Today, Meta released the open-weight model Muse Spark Open and, together with Apple, demonstrated how to run it on two or four Mac Studios. Developers can use the model locally to write code, analyze video, and run agent tasks, with no inference passing through the model vendor’s servers.

As AI starts handling personal files and corporate data, where data is processed and who controls it matters more and more. Open models and local AI complement each other well: users can deploy models themselves and keep inference on their own devices.

Apple’s years of investment in its own silicon, unified memory, and developer tools now have room to pay off. With an 816B-parameter model running on a desktop, the Mac is one step closer to becoming the PC of choice for the AI era.

Read more

The software industry seems to be collapsing at remarkable speed these past few days.

One developer, using Opus 5.5, has built open-source alternatives to seven Adobe products, including Photoshop, Illustrator, Premiere, and Lightroom. What once took a team years to develop, a single developer with AI can now prototype, fairly completely, in a short time.

Meanwhile, another open-source project on GitHub, REA, has quickly gone viral. Some are even declaring that closed-source software is dead and that we’ll never need to pay for software again.

That’s an exaggeration. But REA has indeed changed something big about software development: by installing a single MCP server, you can let an AI agent study the internals of closed-source software and turn its findings into its own code.

Read more

OpenAI is in early acquisition talks with cybersecurity firm Trail of Bits. OpenAI is trying to build a sizable cybersecurity business, and the potential deal could strengthen its capabilities in this area. Details such as the deal value have not been disclosed, and the talks could still change.

At this stage, the more worthwhile questions are why OpenAI needs a specialized security firm, and which layer of its product and business system it plans to put security capabilities into.

Models can help developers generate code and check programs for errors, and enterprises are trying to plug AI into their development, testing, and deployment workflows.

AI coding is expanding the boundaries of what software production can do. But higher code-generation efficiency does not mean software quality and security will improve at the same pace. What enterprises really need are systems that can identify vulnerabilities, assess risk, verify fixes, and keep embedding these capabilities into the development process.

Read more

In the AI world, nearly everyone is debating what AI will replace: mathematicians, designers, economists. But the profession “killed off” most often is surely the programmer.

Anthropic has given an answer nobody expected: before programmers get replaced, the tools they have used for twenty years get replaced first. Claude writes code so fast that the real bottleneck inside Anthropic has become the software tools humans invented over the past few decades, which can no longer handle the code AI produces around the clock.

With no other option, Anthropic assembled a 7-person team, again using Claude, and in 5 months built a brand-new version control system from scratch and migrated its entire monorepo onto it.

Read more

After rolling out the Flash 3.5 through 3.8 series in quick succession, Google has yet to deliver its flagship Pro version, and it badly needs a major product to steady market expectations. Gemini 4 Argon, which quietly began limited testing two weeks ago, is meant to play that role.

Argon is the name of a colorless, odorless, chemically stable gas often used as a protective shield in critical processes such as welding and semiconductor manufacturing. The name fits what Google needs right now: not more hype, but a solid launch to steady the ship.

Just now, Gemini 4 Argon officially launched worldwide. It was polished for six months and battle-tested by more than 100,000 Google employees.

Logan Kilpatrick, Google’s AI product lead, said on social media that this is only the beginning of the Gemini 4 era, and that the new model was designed to help people tackle hard problems they once wouldn’t have dared to attempt.

Read more

Around 1 a.m. last night, I was still grinding on a Bug I just couldn’t pin down.

Running Opus, I got fed up: it never made a wrong step; it was just slow. Every time it “thought” for those dozens of seconds, I could go make tea and scroll a couple of videos, and when I came back, it was still spinning.

Then I dug through the Claude Code docs and found something I had somehow never used: /fast.

The moment I hit enter, “Fast mode ON” popped up, and a little lightning bolt ↯ appeared next to the prompt. On the next turn, Opus was noticeably faster. The official claim is up to 2.5x faster. In practice it doesn’t feel that dramatic, but the sluggishness was definitely gone.

Then I glanced at the price, quietly typed /fast again, and turned it off.

Today I’ll work out the math for you: what this thing actually is, when to turn it on, and when leaving it on is just burning money.

Read more

Last year, Zuckerberg waved checkbooks around the world to recruit AI talent, with some compensation packages reaching hundreds of millions of dollars, on par with an NBA superstar transfer. Now the expensively assembled team has finally delivered a report card closer to expectations.

Just now, released its new Muse Wonder series of models (internal codename “Watermelon”), including Muse Wonder and Muse Wonder Full. They focus on personal and professional agent tasks, with key improvements in cross-conversation memory, long-running task execution, and coding. In an interview, Zuckerberg said Meta has reached the frontier of AI technology.

According to Meta’s published evaluations, the new Muse Wonder Full model can, on at least some benchmarks, go toe-to-toe with GPT-6 Astra and Claude Fable 5.1. Tests by the third-party firm Artificial Analysis also show that Wonder Full is a dramatic improvement over the previous generation, placing in the top three on multiple agent and professional-work evaluations.

Read more

Let me start with a real scene from last Friday.

Friday afternoon, bag in hand and about to head out, the refactoring task on my computer was only halfway done. In the past, I would either have had to leave the laptop open and not dare close the lid, or wait until I got back and explain the requirements all over again.

This time I pulled out my phone, opened the session in the Claude app, checked the progress, and sent: “Add the remaining test case too, and tell me when it’s done.” Then I pocketed the phone and went downstairs. By the time I reached the gate of my apartment complex, my phone buzzed: the job was done.

This isn’t some cutting-edge tech. It’s just the mobile side that Claude Code has had for a while. But honestly, it took me over half a year to realize how smooth it is, so today I’m breaking it down properly.

Read more

In a corner no one is watching, OpenClaw has updated again.

OpenClaw has just officially released v2026.10.1. According to the official figures, this update brings together the work of 402 contributors, covering 2,425 PRs and 3,055 commits.

Two new features stand out. Agents can now proactively search the Skills they have installed, and interactive MCP Apps can be opened right beside the chat window.

Beyond that, OpenClaw also brings off-device full backups, a simpler Mac installation flow, and a series of improvements across automation, session management, and model tooling.

As of publication, the release post had only about 34,000 views, a far cry from the buzz of the nationwide “raising lobsters” craze earlier this year.

So, now that it has left the center of public attention, where is OpenClaw actually heading?

Read more

In Claude Code, /model picks the model and /effort sets the effort level, which controls how much work the model puts into each turn. Opus 5.5 and Sonnet 5.5 default to medium, and Fable 5.1 defaults to high. Below that is low, and above it are xhigh and max. Is the default enough? What does the extra money for higher effort actually buy? When should you switch models instead of raising the effort level? I ran five tasks to find out.

The tasks are small, the kind you’d normally hand to Claude Code: renaming a function, fixing a CSV import error, fixing a bug where month-end subscription renewals compute the wrong date, fixing flash-sale overselling, and speeding up a slow dedup function. Each task’s problem is written up in the repo’s ISSUE.md, essentially a ticket. For every task, I also wrote a test suite the model can’t see (called “hidden tests” below), which I run afterward to score the result.

Read more

According to OpenAI’s disclosure, the company’s annualized revenue was close to $50 billion as of the end of September, below the $70 billion figure that had been widely circulated in the market.

After the news broke, chip stocks such as Nvidia(), Broadcom( ) and Micron( ) weakened together, and AI infrastructure-related assets were also affected. The market had been focused on how much faster AI revenue could grow. This time, the discussion shifted to how revenue is actually calculated, and whether massive spending on computing power can generate sustained, verifiable commercial returns.

The gap between the estimated $70 billion and the actual $50 billion cannot simply be read as OpenAI doing $20 billion less business. The difference stems from how OpenAI and Anthropic each count revenue from cloud partnerships.

OpenAI does not include all of the sales made by some of its cloud partners in its own revenue, and Anthropic uses a different approach as well.

Reuters also noted that an annualized revenue run rate is usually extrapolated to a full year from the revenue of a given period, and does not equal full-year operating revenue that has actually been realized.

Read more

Last Wednesday afternoon, I asked Claude to refactor a method.

It changed 7 files in one go, deleted 200 lines, and even “optimized” a utility class that I had specifically told it not to touch. I ran the tests, and everything came up red.

git status was a wall of M’s. At that moment I really didn’t want to type git checkout ., because a few of the earlier changes were ones I wanted, and they were now mixed in with this round’s. One sweeping command would wipe all of it.

Later I discovered that Claude Code ships with a “time machine” that is on by default. Many people simply never press the two keys that trigger it.

Read more

The plugin system is the most underrated feature of Claude Code from the past year, and many people haven’t even heard its name.

It’s not actually new. I wrote about plugins back in August, but at that time they were still an “experimental feature,” and few people really used them.

Now it’s different. Since Claude Code 2.1.286, plugins have become a standard part of every developer’s toolbox.

What is a plugin?

Let’s get the concept straight first, so you don’t configure things blindly.

Read more

Yesterday, OpenAI opened GPT-6 to all users, free and paid. Today it launched a new model: the Ultrafast version of GPT-6.1 Sol. Paired with an updated model control system, it can respond to user instruction changes in real time.

Turbocharged

GPT-6.1 Sol itself only launched at DevDay on September 29, positioned as “near-Astra capability at one-fifth the price.” Standard API pricing is $2 per million input tokens and $10 per million output tokens, already one of the best-value frontier models available.

Ultrafast mode takes that and pushes speed up by 8x.

Read more

Just now, Anthropic released Claude Haiku 5.5, which it calls its cheapest, fastest, and most capable small model to date.

Haiku 5.5’s average running cost is about 75% lower than the previous generation, Haiku 4.5.

Despite being much cheaper, it scores far higher than Haiku 4.5 on coding, computer use, and knowledge work.

Anthropic also announced two other things: the cache read price for Sonnet 5.5 is cut in half, and Claude Max and Team subscribers will get an extra monthly API credit.

Read more

Tibo’s promise of “28 consecutive days of launches or resets” has reached day three, and today’s perk is a banked reset for paid users.

The real headline, though, is GPT-6.

Last month, OpenAI rolled out the first GPT-6 models to paid users, and the rollout has now widened further.

Just now, OpenAI announced that GPT-6 is launching to all ChatGPT users worldwide, both free and paid.

These models will replace GPT-5.6 Sol and GPT-5.6 Luna in ChatGPT.

Users accessing GPT-6 Sol and GPT-6 Luna through Codex and ChatGPT Work will continue to use the previously released versions.

The most notable new capability in this upgrade is Intelligent UI.

Read more

A few months after Nano Banana 2 launched, Google() has finally updated this product line.

Today, Google DeepMind officially released Nano Banana 2.1.

By name, this is only a “0.1” upgrade, but Google’s own assessment is quite direct: the new version improves on the earlier Nano Banana models, with the main changes concentrated in visual design, mask-based editing, subject consistency, and image naturalness.

More notably, Nano Banana 2.1 doesn’t change the product line’s original direction. It still aims for Flash-level speed and cost efficiency while moving closer to professional image generation and editing. Google defines it as its latest efficient image generation and conversational editing model.

Nano Banana 2.1 is now a stable model in the Gemini API, with the model ID gemini-nano-banana-2.1. According to Google, it has been rolled out to products and developer platforms including the Gemini app, Google AI Studio, the Gemini API, Google Search AI Mode, Google Ads, Flow, and Stitch.

Judging by the model card and benchmark results Google published, the improvements are not limited to a handful of showcase examples.

Read more

In early October, OpenAI’s Tibo set himself a rather pressure-laden challenge: for the next 28 consecutive days, OpenAI would ship at least one clear improvement for Codex / Work users every single day. If there was no update on a given day, the company would simply do a full usage reset instead.

The Day 1 update was straightforward: GPT-6 Astra and GPT-6.1 Sol got a roughly 50% increase in default speed under the subscription plans.

On Day 2, the pace suddenly picked up. Instead of a single update, Tibo released four at once: Auto-review made free, a simplified API usage tier structure, a Meetings plugin that captures meeting records, and wider developer access to the Decisions API. The four updates span Codex, ChatGPT, and the API.

Taken individually, none of them is a DevDay-level product launch. Taken together, though, they point to the same thing: OpenAI is concentrating on the details that, once agents enter everyday work, are most likely to leave them stuck.

Read more

Honestly, I’ve always been pretty casual about installing Claude Code plugins. If something in the community marketplace looked decent, I’d press Y in the /plugin panel and be done with it, with roughly the same mindset as installing a Chrome extension.

Then I read through the official docs page on Plugin security and trust, and a chill ran down my spine. The very first line says it plainly: a plugin you install can execute arbitrary code on your machine with your user privileges.

I’m not trying to scare anyone. Today I’m going to lay out, once and for all, what to check before installing, how to check it, and how to remove a plugin cleanly when you don’t want it anymore. Everything uses the official built-in tools, with nothing extra to install.

First, the math: what can a plugin actually do?

Many people assume a plugin just adds a few slash commands. In reality, a plugin can bundle far more than that:

Read more