Monthly AI Report: Issue No. 6 "April 2026, the month AI thought before it acted"

Published: End of April 2026 / note "Mumima | Tool Way"

Now that we are used to having AI by our side, the AI this month showed us a slightly different face.

A month of "evolution and stability"

In March, AI simply "increased." New models, new features, new services. They were there when I opened them, they were there before I knew it. April, which came after that tumultuous month, felt like a deep breath.

The number of new changes was small. But the depth of the changes was different.

The "AI by our side" became an "AI that thinks before it acts" this month. That is the one-sentence summary of April. GPT Image 2 plans its composition before generating, and Claude Mythos autonomously discovered bugs that humans hadn't noticed. Both indicate a shift from "AI that is used" to "AI that thinks."

Actually, I haven't been able to keep up with the latest LLMs this month. This article by Yui-maru is detailed, so I'll link it here.

🗒️ My favorite AI this month: Gemini's Gems

I kept writing "NotebookLM is my favorite" until last month, but that changed this month.

Gemini's Gems have leveled up as a partner for image creation. The trigger was a small feature addition. It is now possible to register images in the Gems knowledge section along with instructions.

This was more effective than I thought. When drawing illustrations, if I have Gem read reference images for the character's face, style, and worldview, I don't have to explain them at length in the prompt every time. When drawing with NanoBanana2 (Gemini 3.1 Flash Image), it feels like Gem acts as a "bridge." Reproducibility has improved.

It's as if the tool has become accustomed to my work, or rather, instead of me adjusting to the tool, the tool has come closer to me. That was the kind of change it was.

🚪 Main Topic

GPT Image 2 (ChatGPT Images 2.0)

Released on April 21st. This was the most impressive thing in the field of image generation this month.

There is one thing I noticed after using it. The results are better if you don't write the prompt in too much detail.

This is consistent with the design of this model. GPT Image 2 is structured to "think" before generating. It plans the composition and performs a self-check before it starts drawing. Therefore, if you specify too much detail, you will overwrite that thinking. The "give it space" usage I learned with NanoBanana2 worked just as well here.

Text accuracy has improved significantly. In previous models, there was a classic failure where "if you make a menu, the dish names are full of typos," but GPT Image 2 can now draw them almost accurately. There are also evaluations that the accuracy of CJK characters, such as Japanese and Chinese, is "surprisingly good."

Claude Mythos Preview

Announced by Anthropic on April 7th. This was the most notable event in the LLM field this month.

A limited release for an unprecedented reason: "too powerful for public release". Available only to 12 companies, including AWS, Apple, Microsoft, and Google, as part of a defense-only project called "Project Glasswing".

What makes it so powerful? Its cybersecurity capabilities.

It autonomously discovered a bug that had been lurking in the OpenBSD TCP implementation since 1999 (which security researchers around the world had overlooked for 27 years). It did the same for a vulnerability in the FFmpeg H.264 codec that had existed since 2010 (which automated testing tools had failed to detect after over 500 million scans).

What is noteworthy here is this explanation from Anthropic: "This cyber capability was not explicitly trained. It "emerged naturally" from general improvements in code, reasoning, and autonomy."

This is a phenomenon that can only be called "emergence." It became capable of doing things it was never taught. This was the biggest event of the month, signaling that AI evolution has begun to exceed human expectations.

The paradox of being too good to use—this symbolizes the AI of April.

OpenClaw

An open-source AI agent framework that continued to be a hot topic this month, with over 340,000 stars on GitHub. It is a system where AI autonomously executes tasks through messaging apps (Signal, Telegram, Discord, etc.).

To be honest, I have little personal interest in it.

The reason is that "the risks are still too visible to use it as a tool." There have been cases where entire inboxes were deleted, cases where it entered a loop resulting in high charges, and cases where malicious skills were mixed into the marketplace. There is still a gap between "anyone can use it" and "it can be used safely." It seems it will take a little more time to mature as a tool.

📋 AI Updates (Discontinued/Updated)

🔴 Discontinuation Notice

DALL-E 3 → Discontinued on May 12th
With the release of GPT Image 2, DALL-E 3 will be retired on May 12th. DALL-E 2 will also be discontinued at the same time. OpenAI's image generation will be consolidated into GPT Image 2. Sora ended in March, and now DALL-E 3. The generational shift is proceeding quietly.

🟢 Version Upgrade

Gemini 3 Deep Think → Major Upgrade
Google's top-tier reasoning mode, 'Deep Think,' has been enhanced. It is available for Ultra subscribers. The goal is to elevate it to a level where it can be used for practical scientific and engineering tasks, not just abstract theory.

Gemini App for macOS → Official Release
A native Gemini app for Mac has arrived. Although it follows the Windows version, this is a welcome update for Apple users.

📰 News Flash

Google I/O 2026 (May 19-20), Finally Approaching
It is just next month. Judging by the flow of the April Gemini updates, the focus will likely be on further integration of Gemini and the full-scale rollout of Flow. How Google responds to Anthropic's moves is a key point of interest.

Claude Mythos and Project Glasswing Cause Ripples in Japan
Following the announcement of Mythos, national-level discussions have emerged in Japan. On April 20, Mythos was mentioned at a joint meeting of the Liberal Democratic Party's National Cybersecurity Strategy Headquarters. On April 24, the Financial Services Agency held an emergency meeting with the heads of the Bank of Japan and major banks. The phrase 'the crisis that is here now' perfectly captures the atmosphere of this month.

🧭 Summary: The Map Revealed Over 6 Months

Six months ago, AI was our 'assistant.' This month, AI found something that humans hadn't noticed for 27 years. It wasn't just by our side; it noticed things we hadn't even realized yet, before we did.

In the May issue (No. 7), we will look at the 'results of Google I/O 2026' and the 'world where AI has started to get ahead of us.'

#MonthlyAIReport #GenerativeAI #AIUtilization #ClaudeMythos #GPTImage2 #GeminiGem #OpenClaw #NanoBanana2