If you look only at actual usage, Fable 5 is hard to call a successful flagship model.

As the highest-positioned and most expensive model in Anthropic’s lineup, Fable 5 did not achieve usage scale matching its flagship status after launch. Ramp data shows that Fable 5 accounted for only about 6% of Anthropic’s enterprise token consumption.

Combined with enterprise compliance constraints from the data-retention policy, many customers ultimately chose the cheaper and easier-to-procure Opus 5. Surprisingly, Anthropic did not slow the pace of flagship model updates as a result.

Just now, Anthropic officially released Claude Fable 5.1 and Claude Mythos 5.1. The new models have already started rolling out on Claude, Claude Code, the API, and platforms including AWS, Google Cloud, and Microsoft Azure.

According to Anthropic, Fable 5.1 and Mythos 5.1 actually use the same underlying model; the main difference lies in safety restrictions. Fable 5.1 is available to general users and enterprises, while Mythos 5.1 has more relaxed permissions for cybersecurity and life sciences and is currently offered only through a trusted access program to vetted institutions and researchers.

Rather than simply chasing new benchmarks, Anthropic this time clearly focused more on real-world deployment issues exposed after Fable 5 launched. Pricing, data retention, safety-mechanism false positives, and the cost of long-running agents all became key adjustment areas for Fable 5.1.

Read more

Let AI help you make a PPT, write a Word report, or tidy up an Excel spreadsheet—everyone’s trying it. But anyone who’s actually done it knows the crash rate is high.

The problem isn’t that the model isn’t smart enough. It’s that the AI can’t see what it just made. It writes a pile of OOXML XML, but it doesn’t know whether a title overflowed, whether two shapes overlap, or whether a chart rendered crooked. It can read the DOM, but it can’t tell layout problems apart.

That’s “running blind”—the AI makes a PPT with its eyes closed, hands it to you, and you open it to find the title flying off the page, a table sitting on top of an image, and colors that look like a palette got knocked over.

Today this project fixes that at the root.

What is OfficeCLI

In one sentence: the world’s first Office suite designed for AI agents.

Over 29.6k stars on GitHub, Apache 2.0 open source, written in C#, a single executable, .NET runtime embedded, no Office install required, zero dependencies, runs on every platform.

Read more

Anthropic recently shared on its blog Warp’s approach to letting an Agent evolve itself.

The core idea is simple: two layers of Agent Skills plus a loop of human feedback form a closed system that lets the Agent keep correcting itself as it is used.

This mechanism now runs across Warp’s entire open-source repository. Agents that write spec docs, review code, and triage issues each carry their own improvement loop.

The overall approach is quite similar to “I Now Completely Stop Reading AI-Written Code.” That article can be used as a contrast and as a practical demo.

Why self-evolution?

It started when Warp’s internal code-review Agent went wrong. Engineers complained that its comments were useless and the quality was uneven.

The team’s first instinct was to hand-edit prompts and add more context to AGENTS.md. That can work in the short term, but it does not scale for long-running, large-context work.

Because human feedback to the Agent disappears as soon as the session ends.

The same mistake may have to be taught again every week. For the Agent to actually get better, feedback has to persist so the same errors stop repeating.

Read more

Grok Bot can now connect directly to X.

It’s not just “it can read X now.” Link your X account inside Grok Bot, and the system automatically creates a developer account for you. Paid users also get a free allotment of X API calls.

Search posts, read timelines, check who @’d you, summarize what’s happening on the platform—all in one click.

You might think: what’s the big deal?

That’s because you may not realize how expensive it is now to use the X API properly.

Read more

OpenClaw 2.0 Is Here.

Just now, OpenClaw officially shipped the largest update in its history: v2026.8.1.

In the official blog post, the OpenClaw Foundation admitted that this release was somewhat “unexpected.” The team originally only wanted to simplify installation and rebuild the browser client. The work went much deeper than planned.

From Memory, Skills, Automations, Browser, Native App, Plugin, and Security, to Cloud Worker, multi-user collaboration, and the Agent permission system, nearly every core module was touched.

After this round of changes, OpenClaw is starting to shift from a more geek- and developer-oriented Agent Runtime toward an Agent product that is easier to install, better suited to long-running use, and closer to how ordinary users actually work.

OpenClaw 2.0

According to the official blog, OpenClaw 2.0 involved 933 contributors, 569 of whom were first-time contributors. The release includes more than 16,000 pull requests. By the project’s own count, the PRs in this single version account for roughly half of all merged code in OpenClaw’s history.

To ship it, OpenClaw even paused updates for nearly seven weeks—something rare for the project.

The blog explains that the team had grown larger and development had sped up so much that the old architecture and release process could no longer support that growth. They had to rebuild both the product foundation and the release system at the same time.

In other words, many of the issues users felt in OpenClaw over the past six months were also visible to the development team.

Read more

There are now better alternatives to the Codex App’s handy browser plugins.

ego is essentially a Chromium-based browser. The interface and usage habits are no different from Chrome — bookmarks, extensions, and browsing history all work as usual.

Its main pitch is letting agents like Claude Code and Codex work directly in the browser environment you already use every day, without having to go through CAPTCHAs and two-factor authentication all over again.

Agents can create their own workspaces inside ego to run tasks. Each workspace is dedicated to one thing and doesn’t interrupt your existing browsing. You don’t need those agent-browser plugins that create a pile of related tab groups.

After using ego, I completely uninstalled my previous agent-browser and Playwright MCP.

Read more

Unbelievable.

Google just updated Gemini Omni 1.1 Flash, and it shot straight to No. 1 in the world on Arena for text-to-video.

Image-to-video: No. 2 globally, 1488 points.

But the ranking is just the appetizer.

The real breakthrough is that Google fixed, in one go, the most painful problems in AI video.

• Scene extension: the model can look back at up to 10 seconds of previous footage and keep shooting from there. In 10-second segments, total length up to 40 seconds.

• First and last frames: you specify the start frame and the end frame; it generates the continuous motion in between.

• 360p draft: up to 60% faster than 720p, at only one-third the cost. Then upscale the one you like all the way to 4K.

• 4K upscale: finished clips can be raised to 1080p or 4K.

• Video reference: up to 3 seconds of reference video can be mixed into the input, folding action, camera movement, and rhythm into the new shot.

Okay. AI video is finally more than “draw another card.”

It can pick up a scene, hit marks, and even has a bit of editorial thinking.

The key point: when the model continues a shot, the context it can see went from Veo’s 1 second to 10 seconds.

Read more

🚀 The “AI Disruption” publication has gained 10,536 subscribers, and you are all very important!

This publication has been running for a full two years.

🤖 1633+ AI articles published! Are you satisfied with our progress?

The past two weeks have been truly exciting, as Microsoft, Google, OpenAI, and Claude have all released a wealth of tools. Welcome to read our publication, which contains wonderful insights.

Since its establishment, the publication has consistently published at least two articles every day.

“AI Disruption” often appears on the ranking lists.

The publication “AI Disruption” is now accepting guest articles on an ongoing basis and has already published contributions from 24 authors.

Please refer to this guide for participation rules.

The publication has launched 11 featured columns:

• “GPT Introduction” contains 10 articles, now completed.

• “Development of Large Model Applications” includes 23 articles, which have also been completed.

• “Practical Application of AI Large Language Model Systems” contains 28 articles, completed.

• “AI Painting Creation Intro Course” includes 16 articles, completed.

• “OpenAI Live Day 12” contains 12 articles, completed.

• “LLaMA 3 Practical” contains 12 articles, completed.

• “Cursor in Action” contains 15 articles, completed.

• “AI Trend Research” contains 5 articles, completed.

• “Claude Code” contains 42 articles, completed.

• “Gemini CLI” contains 7 articles, completed.

• “Codex CLI” contains 9 articles, completed.

For more details, visit https://aidisruption.ai/p/must-know-ai-updates.

The publication has been recommended by 203 people.

Daily readership reaches at least 6,000.

Thanks to all the followers, especially those who have subscribed.

I believe many have learned a lot of AI techniques from these articles and applied them in their work.

20% off coupon link:

https://aidisruption.ai/subscribe?coupon=d3efcfbd

On a recent earnings call, Jensen Huang declared: NVIDIA has already achieved AGI!

He then made a stunning prediction—

In the future, NVIDIA may need only 40,000 human employees to simultaneously direct 400,000 or even 4 million AI agents working around the clock.

What would it mean if NVIDIA were the first to achieve AGI?

NVIDIA already sits on the world’s largest compute foundation. Once AGI is realized, an astonishing “compute feeding back into R&D” moat would be fully opened.

Once AGI unlocks recursive self-improvement (RSI), NVIDIA would effectively incubate an army of millions of AI-agent super-scientists internally.

That army would feed back into the company itself—automating the design of the next generation of even more extreme GPU architectures, and even rewriting CUDA.

This “left foot stepping on the right foot” leapfrog would deliver NVIDIA unprecedented monopoly profits. Because the physical world’s supply chain is tightly constrained, the vast majority of competitors would be locked out.

In the end, NVIDIA would not only monopolize the overwhelming majority of the world’s effective “AI labor,” but also capture the richest monopoly dividends as humanity enters the silicon-based civilization stage.

Read more

Are you still reviewing code written by AI?

I’m not anymore. Neither is Lauren Tan.

She used to work at Cursor and is now at SpaceX AI. A few days ago, she did an interview describing how she develops now:

For a lot of agent-written code, she no longer reviews it line by line. She even lets agents auto-merge PRs.

When she wakes up in the morning, there may already be twenty-plus PRs in main. Last month, she merged close to a thousand PRs.

Why is she willing to do that?

Most people’s workflow is still this:

Hand the requirement to an agent. The agent changes the code. Then you start the project and check it yourself. If the page looks wrong, you screenshot it and throw it back. If the program errors, you copy the logs over. If performance is bad, you open DevTools yourself.

It looks like the agent is doing the work, but you never actually leave the loop.

So two agents at once is still manageable. Five and you start falling behind. Once a dozen agents are working at the same time, the human becomes the slowest part of the system.

The first thing Lauren did was hand verification over to the agent as well.

She has the agent start the app itself, operate the UI itself, capture Performance traces and heap snapshots itself, or open the iOS Simulator.

Read more

Can a two-month flagship iteration hold up under a magnifying glass? No extra parameters, no architecture change—just post-training to lift coding ability a notch. That is the report card GLM-5.3 just submitted.

First, the conclusion: this is a “post-training–driven” iteration

On August 18, 2026, Z.ai (Zhipu AI’s international brand) published GLM-5.3 on its official docs site, docs.z.ai. That is only about two months after the previous generation, GLM-5.2 (2026-06-16).

What matters most is not the numbers, but the official characterization of this iteration:

“Z.ai’s latest flagship model, delivering comprehensive advancements in complex software engineering and agent capabilities. It uses the same base model as GLM-5.2, with all improvements driven by post-training.”

In plain language: the base was not swapped, the parameter count did not move, the attention mechanism was not changed—every gain comes from RL + SFT post-training. In a field that currently worships “new architecture + new parameter count + new modalities,” that is a counter-narrative. The GLM-5.3 story is: same base, how far can post-training push the ceiling?

We already unpacked GLM-5.2’s architecture details—IndexShare, MTP, Sparse Attention, and so on—in the late-June piece, A Deep Dive into GLM-5.2. This article will not repeat them. It focuses on three things: how much coding ability actually rose in 5.3, how it rose, and how to wire it into Java / Spring AI engineering.

A precaution first: every benchmark number below comes from Z.ai’s official self-evaluation. The harness is Z.ai’s own eval environment, not a unified third-party harness. As of publication, no independent third-party reproduction was found. Read with a discount filter on.

Read more

August 17 at midnight, DeepSeek’s new API pricing officially took effect.

The entire V4 lineup now uses peak/off-peak time-of-use billing. On weekdays from 9:00–12:00 and 14:00–18:00, usage is billed at peak rates; in all other periods, prices are halved. Who would have thought large models would one day also need to “shift electricity use off-peak.”

Third-party calculations show that for V4 Flash, off-peak cache-miss input prices rose about 57%, off-peak output about 136%, and peak-period output prices are close to 4.7 times the old rate.

DeepSeek is not the only one raising prices. OpenAI, Google, and Anthropic have already raised API prices or shrunk free quotas. For two years, model vendors kept driving prices down; in the second half of 2026, that price curve has started turning collectively upward.

The reason is not complicated. Growth in inference demand has outpaced growth in compute supply, and agents turn a single call into dozens of calls. When the bill changes from “one question, one answer” to “one task running for hours,” the low prices previously sustained by subsidies can no longer hold.

Capabilities keep going up — can prices still go down? That has become a required question before large models truly enter production environments.

Just now, Alibaba delivered a satisfying answer: it released Qwen3.8-Flash and simultaneously open-sourced Qwen3.8-Flash-Next (they are the same model; the former is the API product name, the latter the official open-source name).

Read more

Making chips is hard; making a competitive first-generation product is even harder. The answer OpenAI has delivered this time is, at least when it comes to large-model inference, already quite extraordinary.

Today, OpenAI officially released the first batch of measured performance data for its first custom inference chip, Jalapeño (Mexican chili pepper).

On the public benchmark InferenceX based on its own open-source model GPT-OSS 120B, Jalapeño outperforms existing hardware systems in both peak throughput per kilowatt and token latency.
It’s worth noting that for most chips, throughput and latency are often a trade-off where you can only choose one. Beyond the clear optimization for OpenAI’s large models, the chip also performs better on the 670B DeepSeek R1 and the 1T Kimi K2.5, with the competitors being the strongest commercial AI systems on the market, such as NVIDIA’s GB200/GB300.
Altman commented bluntly on X: “We built a chip that’s ridiculously fast.”

From data centers to phones, AI chips are becoming a contested territory for all tech giants.

Read more

Today, Claude officially announced that chat and Claude Cowork now share a single memory. Going forward, when you hand a task to Cowork, it can start from projects, manager preferences, or client backgrounds it already knows from chat—no need to explain everything from scratch each time.

This update may sound like it merely saves a few prompts, but what it actually changes is the default starting point for tasks. Background left in chat yesterday can flow directly into Cowork’s work today. It reduces repeated explanations, yet it can also carry old information into the next task. Before enabling the new feature, it’s safer to delete three categories of content first.

First, clearly see what it has actually remembered

Claude states that memory is stored as a list of topics in Settings. Users can view, edit, or delete items one by one. It updates automatically with chat; when you encounter something worth keeping long-term, you can simply say “remember this.”

Read more

Last Friday night at 7 PM, I had a major refactoring task on my hands that I estimated would take two or three hours. In the past, I could only leave my computer running and wait. Now I just did one thing:

claude --cloud "Migrate the payment module under src/legacy to the new architecture, keep all tests green"

Then I closed my laptop and went downstairs for dinner.
When I came back and checked my phone, the work was done, and the PR had already been submitted. This is what we’re talking about today: Claude Code Cloud Sessions.

Let’s be clear first: It’s not the same as Remote Control

A lot of people mix these two up—I did too at the beginning.

• Remote Control: Claude Code runs on your local computer, and you “watch” it remotely from your phone. Your computer cannot be turned off.

• Cloud Sessions: The task runs directly on Anthropic’s cloud VMs. You can shut down your computer however you like—the task keeps running.

One is a remote control; the other is outsourcing the work entirely. Completely different things.

Read more

Today, the 10-year U.S. Treasury yield fell back to around 4.70%, and oil prices also dropped. Based on experience from the past few years, these two signals are generally not bad for tech stocks.

The result, however, was completely the opposite.

The Philadelphia Semiconductor Index continued to fall. Nvidia dropped 2.9% and has now closed lower for a seventh consecutive trading day; Micron fell 5.8%, and SanDisk dropped 6.45%. Memory and AI hardware remain the hardest-hit areas.

This shows that what the market is worried about now is no longer just “whether interest rates are high.”

A more troublesome issue is another one: the U.S. government needs to borrow a large amount of money, while the AI industry has also entered a peak period of capital expenditure and financing. Both sides are competing for long-term funds.

Money has not suddenly increased, but the number of borrowers has.

Read more

“AI Disruption” Publication 10,500 Subscriptions 20% Discount Offer Link.

Multiple terminal windows of Claude no longer need you as the “messenger” — v2.1.224 lets them chat privately among themselves.

First, understand how AIs “chat privately” with each other

If you frequently use Claude Code, you’ve probably run into this scenario: you open three terminal windows—one modifying the backend API, one writing the frontend calls, and one running tests. After Window A finishes updating the interface, you have to manually copy and paste the change details into Window B, telling it “the interface is ready, time to integrate.”

You are the messenger between the AIs.

On August 7, 2026, Anthropic released Claude Code v2.1.224, which introduced the cross-session messaging feature, and officially announced it the next day. In simple terms: Claudes in different terminal windows can now send messages to each other on their own.

You only need to tell the Claude in Window A something like “notify the backend window of the API changes,” and Claude will write a summary itself and send it to the Claude in Window B via an internal messaging channel. Window B receives it and reads the message during the gap between tool calls, without interrupting the task currently being executed.

💡 Analogy: It’s like colleagues sending private Slack messages—no need for you to relay them; they handle it themselves.

Read more

In August 2026, two seemingly independent events occurred in the same week:

• On August 19, OpenAI officially open-sourced Codex Harness—the underlying execution engine that powers the ChatGPT desktop app, CLI, and VS Code extension. In the past, this system had only run millions of lines of code inside OpenAI.

• DeepSeek’s DeepSeek Harness (dsh) had already been released as a developer preview under the MIT license, with the official statement clearly stating: “Everything is a plugin.”

Many people simply interpret these two events as “OpenAI has finally open-sourced Codex” or “DeepSeek has dropped another bombshell.” But if you only watch the spectacle, you will miss the truly important signal—

The formula Agent = Model + Harness has already been widely recognized by the industry. Large models are the engine; the Harness is the systems engineering that turns the engine into a controllable car that can actually hit the road. Codex Harness and DeepSeek Harness happen to represent two completely different manufacturing philosophies for this “car.”

What this article aims to do is not to pick sides for you, but to dissect the underlying logic of these two paths and tell you at what stage, for what kind of team, and for what kind of business you should bet on which side.

Read more

If we break down the U.S. stock market performance so far this year, we can see a very important change underway.

Over the past two years, the simplest and most effective trading strategy in the market has been to buy AI. Anything related to artificial intelligence—whether GPUs, servers, networking equipment, optical modules, or data centers—would attract capital willing to assign higher valuations.

But by August 2026, this “buy everything AI” trade is clearly losing effectiveness.

Goldman Sachs’ latest market observations offer a judgment that deserves close attention: the AI trade has not ended, but it has shifted from the previous broad-based rally into a phase that depends more heavily on fundamentals, earnings expectations, and valuation mismatches.

In other words, the real money-making opportunities going forward may no longer come from simply hunting for “AI concept stocks,” but from identifying companies whose share prices have lagged significantly behind earnings expectations while their fundamentals continue to improve.

This is why Goldman Sachs currently sees particular tactical opportunities in the storage and data center sectors.

Even more noteworthy is that this shift in capital style is not limited to the AI space. European and Japanese banks, gold miners, copper mining stocks, and certain financial and hard-asset plays are also entering the field of view of institutional investors.

For U.S. equity investors, this may signal that a very important market regime shift is underway.

Read more

Grok Bot is now free to use.
No need for SuperGrok Heavy at $300 per month. SuperGrok Plus users can use it directly, and users without a subscription can try it for free.

So what’s the real difference between Grok Bot and other Agents?
Actually, there isn’t much. It’s just the model plus an Agent, with the addition of a computer that the Agent can control.

What makes it truly powerful and advanced is that it completely eliminates the need for MCP, Skills, and all kinds of plugin integrations. It is a fully autonomous Agent.

The smallest execution unit of Grok Bot is the Bot. Every Bot is directly connected to this always-on cloud computer. Each Bot is allocated a browser, a file system, and a terminal. Once you assign it a task, it logs in by itself and gets the job done.

This persistent computer runs in the background forever—it never shuts down and never resets. It logs in autonomously, uses tools autonomously, and interfaces directly with real systems. Expense reports go straight into the finance system; emails land directly in the inbox.

Read more