Overnight, the official merger of ChatGPT and Codex has been put on the agenda.

Codex was originally designed for developers, focusing on programming and engineering tasks. Now, OpenAI is preparing to push it into broader enterprise workflows. OpenAI announced that Codex will be integrated into ChatGPT within the coming weeks.

After the integration, users will no longer need to switch between the two products. Codex’s Agent capabilities will be incorporated into ChatGPT, covering desktop, mobile, browser, and commonly used enterprise tools such as Excel, Slack, and PowerPoint.

In OpenAI’s vision, ChatGPT will evolve from a conversational entry point into a unified interface for employees to invoke Agents to complete their work. Codex will transform from a code-writing tool into a work Agent capable of handling data analysis, sales preparation, marketing assets, product prototypes, investment research, engineering fixes, and security scans.

The more critical shift is that, starting today, OpenAI’s product strategy is fully benchmarking against Anthropic. By bringing Codex into ChatGPT, OpenAI is moving beyond organizing products solely around a general chat interface. Instead, it is beginning to follow Claude’s approach—placing code, files, enterprise systems, long-running tasks, and workflow Agents all within the same client. ChatGPT’s new form is increasingly resembling the Claude client.

Read more

Ready? Starting this June, we may witness the largest IPO wave in human history.

On May 20, SpaceX publicly filed its S-1 registration statement with the SEC, targeting a Nasdaq listing on June 12. The valuation exceeds $1.8 trillion, with an expected fundraising range of $40–80 billion. On the same day, reports emerged that OpenAI is preparing a confidential S-1 filing, with Goldman Sachs and Morgan Stanley as joint underwriters. The target listing is in September, with the latest funding round valuing the company at $852 billion.

Yesterday, Anthropic officially announced it had submitted its S-1 to the SEC. Just last week, it completed a $65 billion Series H round, rocketing its valuation from $38 billion in 2024 to $965 billion — surpassing OpenAI to become the world’s most valuable private AI company.

Together, these three companies could raise more in IPO proceeds this year than the total raised by all U.S. venture-backed IPOs since 2000. For context, according to Renaissance Capital, the entire U.S. IPO market raised just $28.4 billion in the first five months of this year.

Read more

Claude Code Just Dropped a Major Feature — Dynamic Workflows Now in Research Preview

In simple terms, this lets you command hundreds of AI Agents with a single sentence. While they work, your terminal stays completely responsive — you can keep chatting, grab a coffee, or do whatever you want.

The Problem It Solves

When using Claude Code, have you ever run into these situations?

• You ask it to audit API security across an entire project. It reads a few files and then starts summarizing.

• You request a large-scale migration. It handles a dozen files and then stops because the context window is full.

This is the ceiling of Subagent mode. Subagents are “little helpers” Claude spins up one by one in a conversation. Each finishes its task and reports back, and all results have to be stuffed back into Claude’s context. The brain only has so much room.

Dynamic Workflows take a completely different approach:
It turns the orchestration plan into a JavaScript script that is executed by a dedicated runtime. Intermediate results are stored in script variables and do not consume Claude’s context window. Only the final result is returned to you.

Read more

Software has long been the darling of the US stock market’s tech boom over the past decade or so. These companies grow quickly with high upside potential—if you pick the right ones, they can easily deliver substantial returns for investors. However, this very group of market favorites has faced an unprecedented crisis in the past six months. With the explosion of AI Agents, the narrative of AI disruption has grown extremely loud. The entire software industry has been brutally bloodied, with almost no one escaping unscathed. The software stock ETF IGV has dropped nearly 40%, and many star individual stocks have been cut in half.

That said, a quiet reversal seems to be underway recently. As software companies release their earnings reports one after another, we are seeing a handful of software stocks begin to surge. For example, Snowflake soared 36% after its earnings, Datadog rose 32%, and TEAM jumped 30%, among others. At the same time, however, the majority of software stocks continue to face market punishment—even strong results have failed to dispel the skepticism.

What investment logic lies behind this contradictory phenomenon? Over the past period, I have repeatedly expressed my optimism toward the current software sector. I have said that instead of chasing the already sky-high foundational layer plays, it is better to position in application-layer software companies that still have reasonable valuations and are about to benefit from numerous catalysts. This is also one of my most favored US stock investment opportunities for the second half of this year. As I promised earlier, I will conduct in-depth research on application-layer software companies.

Today, let’s start from this earnings season and examine what kind of software companies can prove themselves and win Wall Street’s recognition—and which ones will continue to be punished. From these trading dynamics, can we identify the next software opportunity ready to explode? I will also share my personal, specific software stock selection methodology and portfolio allocation approach.

Read more

As IPO expectations continue to heat up, Anthropic’s product and model release cadence is also accelerating sharply.

Just now, Anthropic dropped two major announcements in quick succession: first, upgrading its flagship model to Claude Opus 4.8; second, completing a $65 billion Series H financing round. The post-money valuation reached $965 billion, putting it on the verge of the trillion-dollar mark.

For an AI company already valued near a trillion dollars, the market is no longer focused solely on benchmark scores. Instead, it wants to see whether the company can turn intelligence, tools, development environments, cloud platforms, and compute resources into a scalable, deliverable infrastructure.

In other words, Anthropic must evolve from a company that simply delivers strong models into one that aims to reshape how enterprises work with AI.

Read more

DeepSeek V4 is fast, cheap, and has strong reasoning capabilities.

However, it has one fatal flaw: it does not support multimodal input.
The lack of multimodal support significantly affects its usage in Claude Code and frequently interrupts normal workflows.

For example, if you screenshot an error message or drag in a UI design draft, the DeepSeek model simply gets stuck. There are far too many important scenarios like this.

So I built a small tool that integrates seamlessly and delivers a native Claude Code experience. Video demo below:

Read more

After using Grok Build for a full day, here’s my conclusion:

It has the best interactive experience among all the CLI coding tools I’ve tried — bar none.

That said, if you’re hoping it will fully replace Claude Code or Codex, you might be underestimating it.

Musk’s Little Touch

After installing Grok Build, you’ll notice one of Musk’s clever little ideas: the terminal aliases are grok and agent.

That means once you type agent, you’re immediately dropped into Grok Build’s interactive control interface.

This small detail also reveals Musk’s bigger ambition:
Whether it succeeds or not, he’s claiming the “Agent” slot first.

Read more

Musk Has Played His Two Trump Cards to fix xAI’s Programming Weakness.

On May 14, xAI released the early Beta of Grok Build, officially positioning it as a “programming agent and CLI (command-line tool)”, initially available only to SuperGrok Heavy subscribers.

By May 25, xAI officially published the Grok Build announcement on its website, expanding access from SuperGrok Heavy to all SuperGrok and X Premium Plus users. At this point, it has evolved from a high-barrier, limited Beta into a tool that more paying users can actually use.

Musk had previously publicly admitted that xAI lagged in programming scenarios. According to Bloomberg, xAI executives internally pushed the team to make Grok competitive with Claude across various tasks. Grok Build is the first product from this catch-up effort.

However, shortly after launch, users raised a sharp issue: while the product interface is good, the underlying model isn’t strong enough.

“As long as xAI releases a truly SOTA (state-of-the-art) model, Grok Build could overnight go head-to-head with Codex and Claude Code.”

Musk then replied on X, revealing xAI’s next-generation base model:

“The Grok V9 1.5T we recently completed is performing extremely well — and this is before we’ve even added Cursor data for continued training.”

Read more

AGENTS.md is now familiar to everyone — it tells the AI how to build a project.

But when you ask an AI to create a landing page, dashboard, or admin panel, the visuals never quite match your product style — wrong button colors, wrong fonts, wrong spacing, wrong animations.

Google Stitch has introduced a new concept called DESIGN.md — the sibling file to AGENTS.md:

A pure-text design system document. Once an AI Agent reads it, it can generate consistent, on-brand UI — no need to export from Figma, no JSON schema, no special tools required.

The VoltAgent team has created an awesome collection that extracts 73 mainstream websites’ DESIGN.md files, which you can copy directly into your own projects.

Read more

I previously provided a systematic introduction to Claude Code’s .claude directory.

This time, I’ll continue by sharing a detailed breakdown of Codex’s .codex configuration files. This is based on the latest version of Codex, including explanations of the main configuration files, detailed field descriptions, and some of the configurations I am currently using myself.

In Codex, the .codex directory primarily serves two responsibilities:

1. Codex’s own local state directory
Default location: ~/.codex
Codex stores its local state under CODEX_HOME, which defaults to ~/.codex.

Common contents include config.toml, auth.json, history.jsonl, logs, and cache.

2. Project-level configuration directory
That is, .codex/config.toml inside the repository.

Codex searches for .codex/config.toml starting from the project root and moving up toward the current working directory. Configurations closer to the current directory have higher priority.

The main difference between the two:

~/.codex is Codex’s global workbench, while the .codex folder inside a project provides project-specific local configuration for Codex.

Recommended .codex Directory Structure

Here is the recommended structure for the new version’s local directory:

~/.codex/
├── config.toml              # Main user-level configuration
├── auth.json                # Login cache (appears when using file storage; sensitive file)
├── history.jsonl            # Command history (appears when history persistence is enabled)
├── log/                     # Codex log directory (default log location)
├── hooks.json               # User-level lifecycle hooks (optional)
├── themes/                  # TUI custom themes (optional)
├── model-catalogs/          # Custom model catalog files (optional, self-hosted)
└── state/cache/...          # Other caches or state files (may vary by version)

In this structure, only config.toml is the file you will edit most frequently.

Read more

Anthropic Is Finally Making Money!

While everyone is still debating whether “AI companies can actually make money,” Anthropic just ended the argument with its quarterly report.

Second-quarter revenue is projected at $10.9 billion, more than doubling from the previous quarter’s $4.8 billion.

Even more explosive: Anthropic turned profitable this quarter, with expected operating profit of $559 million.

This is the first time Anthropic has turned a profit since its founding — two full years ahead of its own guidance to investors.

On the same day, OpenAI’s Q1 financials were also leaked.

Adjusted operating margin: -122%

In plain English: For every $1 OpenAI makes, it loses $1.22.

Read more

Today, let’s talk about how to combine two MCP plugins to instantly double Claude Code’s combat effectiveness.

These two plugins—one for code understanding and one for context slimming—when used together, will make your Claude Code experience for large projects take off.

Claude Code’s Context Dilemma

If you’ve used Claude Code for large projects, you’ve probably experienced this:

At the beginning, the conversation flows smoothly. But after about 20 minutes, Claude starts “forgetting”—it forgets which file it was editing, what task you assigned, and even repeats code it has already written.

This isn’t because Claude suddenly got dumber. It’s because the context window is completely full.

Imagine this: Every time Claude calls an MCP tool to read code or run commands, the raw data returned gets dumped straight into the context window. One Playwright snapshot = 56 KB, 20 GitHub issues = 59 KB, one access log = 45 KB… After half an hour, 40% of the context is occupied by garbage data.

Even worse, for Claude to understand your project, it has to read files one by one. How could it possibly read an entire million-line codebase?

Read more

Gemini App has surpassed 900 million monthly active users, with a monthly token processing volume of 3.2 quadrillion (3,200 trillion), and Nano Banana has generated over 50 billion images...

At the Google I/O conference that just concluded in the early hours of this morning, Google CEO Sundar Pichai kicked things off by throwing out these impressive numbers.

Over the past year, AI has become the dominant theme across all industries. Gemini’s positioning within Google has evolved from a standalone app into the most important AI foundational capability across all Google products.

This keynote started with the models, then moved into Coding and Agent products.

Gemini Omni pushes Google’s video generation toward the “world model” direction, while Gemini 3.5 Flash is being positioned alongside AI programming tools as part of an Agent development platform.

These two capabilities will subsequently be integrated into Google’s full ecosystem, including Search, the Gemini App, Flow, Spark, Chrome, XR glasses, and e-commerce scenarios.

Read more

“At 1/10th the cost, its performance nearly matches models at the level of Claude Opus 4.7.”

This is Composer 2.5, the most powerful model Cursor unveiled in the early hours of today.

According to the official announcement, Composer 2.5 is significantly more intelligent, far better at handling long-running, sustained tasks, and much more reliable at following complex instructions.

Within the next week, Cursor will double the usage quota originally included with this model.

Compared to Composer 2, Composer 2.5 shows substantial improvements in both intelligence and behavioral performance.

Read more

Codex “Connections” is mainly divided into three categories:

Once set up, you can grab your iPhone and work anytime, anywhere. Even while out shopping with your girlfriend and waiting for her, you can vibe-code. It effectively connects your entire development network through Codex, sharing context across devices — it’s insanely useful.

Below, I’ll explain in detail how to set it up. If you’ve already connected, feel free to skip this section.

1. Control this Mac

This means “Use this Mac as the Codex host, allowing your phone or other devices to control it.”

Read more

Claude is showing strong determination to push into traditional industries.

Just recently, it open-sourced its financial skills, and now Anthropic has directly embedded Claude into the entire legal industry.

Anthropic has open-sourced github.com/anthropics/claude-for-legal on GitHub, including:

• 12 role-specific plugins: Corporate Counsel, Business Contracts, M&A, Privacy, Product Law, Intellectual Property, Litigation, Compliance, AI Governance, Law School Clinic, Law School Student, and a Builder Hub.

• 5 hosted Agent recipes: Contract Renewal Monitoring, Case Activity Tracking, Regulatory Update Scanning, Due Diligence Table Generation, Product Launch Monitoring.

• 20+ MCP connectors: Ironclad, DocuSign, iManage, NetDocuments, Box, Datasite, Everlaw, Relativity, CourtListener, Slack, Google Drive… basically connecting to almost every system a law firm uses on a daily basis.

Read more

Recently, there’s a must-read article: Perplexity has open-sourced their internal best practices for maintaining hundreds of Skills.

The biggest takeaway after reading it: The best practices for writing Skills are almost the complete opposite of the best practices for writing code.

Zen of Python vs Zen of Skills

The Perplexity team made a joke using PEP 20 — “The Zen of Python.” They created a comparison table showing that roughly half of the Python Zen principles should be completely reversed when writing Skills.

In one sentence: Writing a Skill is not writing software — it’s building context for the model. The constraints are completely different, and so are the design principles. If you approach writing Skills the way you write code, the result will be terrible.

I especially agree with the last one — if something can be explained clearly in a few words, the model can probably handle it on its own. Putting it in a Skill just wastes tokens.

Read more

Anthropic has officially launched Claude for Small Business.

This product directly transforms AI from a “chat window” into a “digital employee.” With 15 ready-to-use Agent workflows, it’s essentially handing small and medium-sized business owners a lean, highly capable virtual team.

I stayed up late last night going through every single tutorial, case study, and manual in the official resource library—dozens of them. I have to admit: AI has truly started doing the “dirty work.”

We used to struggle with writing better prompts to train large models. Now Claude has turned “Standard Operating Procedures (SOPs)” into intelligent agents. Quite intimidating, isn’t it?

Another weekend for some enjoyable learning. This time, I’ll do a deep breakdown of how this “full suite” helps bosses “slack off” effectively.

Read more

Anthropic has officially jumped into the fray itself. The repository is called claude-for-financial-services. Right out of the gate, they’ve covered the four most expensive and high-value segments on Wall Street: investment banking, equity research, private equity, and wealth management.

In simple terms, Anthropic has taken the daily work of Wall Street analysts and packaged it into a ready-to-install Claude plugin suite.

The official wording is quite telling: these Agents are designed to draft working materials for analysts (models, memos, research reports, statements). They do not make investment decisions, execute trades, assume risk, or approve account openings. Every output is left there for a human to review and sign off.

The boundary is drawn very cleanly. The financial industry is most sensitive about accountability. Being able to do the work without taking the blame is actually the most practical and realistic approach for B2B deployment.

The entire repository is structured in two layers:

• Agents (11 total): End-to-end workflow agents, such as Pitch Agent, Earnings Reviewer, GL Reconciler. Each is a self-contained plugin — install it, and it can run an entire pipeline.

• Vertical Plugins (7 industry-specific packs + 2 partner packs): Underlying skills, slash commands, and data connectors, packaged by financial sub-sector. If you don’t want the full Agents, you can just install these foundational capabilities.

Read more

Cloudflare has quietly launched a new feature called Agent Setup.

It has consolidated all major Coding Agents — including Claude Code, Codex, Cursor, GitHub Copilot, and Windsurf — bringing every single one of them under its roof.

The main purpose of Agent Setup is to directly connect Code Agents to the entire Cloudflare ecosystem through proxy configuration.

Agent Setup provides three layers of capabilities:

Read more