Affiliate Disclosure: This article contains affiliate links. If you purchase through these links, I may earn a small commission at no extra cost to you. I only recommend products and services I genuinely believe in.

The race to build the best AI agent for everyday work is moving fast. OpenAI just dropped major updates to Codex that push it far beyond its coding roots. Google has been quietly rolling out useful changes to Gemini and NotebookLM. And somewhere in the middle of all this, a viral image prompt is making everyone’s photos look absolutely terrible on purpose.

This week was packed with real, practical updates that matter to people who use AI for actual work, not just tech experiments. The updates touch research tools, file creation, voice input, and the ongoing battle between OpenAI and Anthropic for the title of best AI work assistant. Some of these changes are small but satisfying. Others could genuinely shift how people manage their day.

Here is a clear breakdown of what changed, what it means, and which tools are worth paying attention to right now.

Key Takeaways

• Codex now handles knowledge work, not just coding, putting it directly in competition with Claude’s Co-work tab
• Gemini can now create downloadable files in formats like PDF, Word, Excel, and PowerPoint directly from chat
• NotebookLM’s new auto-label feature sorts research sources by topic automatically, a big deal for students and researchers
• GPT-5.5 prompting guidance from OpenAI confirms that shorter, simpler prompts now work better across major AI models
• Gemini’s voice input on mobile finally works properly after a frustrating history of cutting off mid-sentence
• Spotify’s AI music labeling signals growing platform pressure to flag AI-generated content clearly
• Viral image prompt turns any photo into a clumsy, scribbly sketch, with GPT Image 2 producing the funniest results

What Codex Actually Does Now (It’s More Than Code)

Most people still think of Codex as a coding tool. That view is now outdated. OpenAI’s recent update to Codex gives it what you might call “co-work powers,” pulling it much closer to what Anthropic’s Claude has offered in its Co-work tab for a while.

To understand what changed, it helps to know how Claude’s desktop app is set up. It has three separate tabs: a regular chat tab, a Co-work tab built for knowledge workers, and a Claude Code tab aimed at developers. The Co-work tab lets users draft documents, answer emails, set up automations, and connect to third-party tools, all without needing any technical background.

Codex takes a different approach. Instead of separate tabs, everything lives in one chat interface. Instantly get the right expert for any task you face. Just ask a general question, and it responds like an intelligent chatbot, giving you immediate answers. Then, tell it to write code, and it quickly switches into a powerful coding assistant for you. Need to sort emails or draft a report? It simply becomes your dedicated Co-work assistant, ready to help you now. No switching, no setup. That flexibility is a real advantage for people who don’t want to think about which mode to use.

One standout Codex feature that Co-work doesn’t have yet is a live task list. As the conversation moves forward, Codex tracks things that didn’t get finished and surfaces them as clickable follow-up items. For example, after asking it to scan and sort a full email inbox, it builds a list of emails that need action. Clicking any item on that list immediately starts the work. That kind of proactive tracking saves time and cuts down on the mental overhead of remembering what still needs doing.

➤ Write Smarter Code and Deploy Faster From One Place

Affiliate link — I may earn a small commission at no extra cost to you.

Claude vs. Codex: Too Early to Pick a Winner

The online debate about whether Codex or Claude Co-work is better has been loud this week. Both tools now cover similar ground. Both offer connectors (Codex calls them plugins), automations, and the ability to work through complex multi-step tasks without needing technical skills.

That said, there are real differences between competing AI coding platforms that go beyond surface features. Claude Co-work has had more time in the field and a more established user base. Codex is newer in this space but backed by OpenAI’s deep infrastructure and the full weight of its model upgrades.

The honest answer right now is that neither has clearly won. Codex feels fresh and the task-tracking feature genuinely impresses. Co-work feels more polished in certain workflows, especially for document-heavy tasks. A proper head-to-head test over several days of real work is the only way to know which one fits better for any specific use case.

What matters most is that competition between these two is good news for users. Each update from one side tends to push the other to move faster.

Gemini Gets File Creation and a Much Better Voice Input

Google has been making steady, quiet improvements to Gemini. The most visible new feature is file creation directly from chat. Users can now ask Gemini to research a topic and produce a downloadable file in the format they need.

Supported file types include Markdown, PDF, Word documents, Excel spreadsheets, PowerPoint presentations, and rich text format. Gemini can also create files straight into Google Drive, producing Docs and Sheets that open immediately for further editing. Users can even describe changes to the document from within the Gemini chat after the file is created.

This works on mobile too, which makes it genuinely useful on the go. Creating a slide deck during a commute or drafting a research document from a phone is now a real option, not a clunky workaround.

The second Gemini update is smaller but arguably more satisfying to everyday users: voice input on mobile finally works properly. The old version would cut off recording at the first pause, making it almost useless for anyone who speaks naturally with brief gaps between thoughts. The new version keeps listening after a pause and gives users two buttons when they finish. One transcribes the message without sending. The other transcribes and sends it right away. It sounds minor, but for anyone who uses voice input regularly, this fix is a real relief.

➤ See Every Visitor, Every Click, Every Conversion in Real Time

Affiliate link — I may earn a small commission at no extra cost to you.

NotebookLM’s New Source Labels Are Bigger Than They Look

NotebookLM rolled out a feature this week that most people will scroll past without realizing how useful it is. The new auto-label button sorts all research sources in a notebook by topic, automatically grouping them into labeled categories.

Each label works like a filter. Users can click a label to see only the sources in that group, rename labels, add emoji to them, or move sources between categories. Sources can also appear under multiple labels, which matters when a single document covers more than one subject.

This feature is especially valuable for students. A single notebook can now hold sources for an entire semester. Each subject or topic within the class gets its own label. When a test is coming up on one subject, the student selects that label and works only with those sources, whether that means chatting through the material, creating an audio overview, or building flashcards. At the end of the semester, all labels can be selected together for a full review.

On top of that, the notebooks themselves are now available to all free and paid users inside Gemini, including on mobile. This brings a proper folder structure to research work, keeping individual project chats separate from the main chat history. For researchers, writers, or anyone managing multiple ongoing projects, this kind of organization for research notebooks cuts a real amount of friction out of daily work.

GPT-5.5 Wants Shorter Prompts Now

OpenAI released an official prompting guide for GPT-5.5, and the headline finding is a welcome one for non-technical users: shorter prompts now work better.

To get good AI results, you used to write long, detailed instructions. This meant providing plenty of context and specific rules for the model. GPT-5.5 has moved past that. OpenAI now recommends simply stating the goal and letting the model work out how to get there. The model is better at reading intent and filling in the gaps on its own.

This same pattern is showing up across Anthropic’s models too. Claude now handles shorter, goal-focused prompts much better than it did a year ago. Both platforms are quietly making the case that AI tools no longer need expert users to get expert results. That’s a meaningful shift for anyone who felt intimidated by the idea of “learning to prompt.”

For people interested in using AI coding tools more affordably, this trend is also good news, since fewer tokens spent on elaborate prompts means lower usage costs over time.

➤ Track Every Click and Conversion Without the Complexity

Affiliate link — I may earn a small commission at no extra cost to you.

The Viral Prompt That Makes Your Photos Look Terrible (On Purpose)

Not every AI story this week was about productivity. A prompt went viral for turning any image into what can only be described as a clumsy, scribbly, utterly pathetic drawing. Think kindergarten-level art done with a shaky hand.

Testing this prompt across multiple image models gave interesting results. GPT Image 2 handled it the best, producing sketches that were genuinely funny in how bad they looked. The model leaned into the brief and came out with something deliberately awful in the most charming way.

The broader image model comparison this week also produced a clearer verdict on GPT Image 2 versus competing tools. GPT Image 2 excels at generating images from a text prompt alone, producing consistently polished and creative results. When it comes to editing an existing photo, though, it struggles. Changing a facial expression or adjusting specific details in an uploaded image works much better in other image models. The practical takeaway: use GPT Image 2 to create from scratch, and use a dedicated image editor when the starting point is an existing photo.

Spotify Labels AI Music and Apple Accidentally Exposes Its Own Tools

Two smaller stories from this week are worth noting. Spotify is rolling out a “Verified by Spotify” label that will make it clear to listeners which music was generated by AI and which wasn’t. This kind of labeling is overdue across every major platform. Social media platforms in particular, where AI-generated images and audio spread fast among audiences who may not realize what they’re looking at, would benefit from the same approach.

The second story is more entertaining. Apple pushed an update that accidentally included a file called claw.md in the package, which pointed strongly to the company using something like Claude Code in its own development process. The file was pulled within hours via an emergency update. It confirmed what many in the industry had suspected: major tech companies are quietly using AI coding tools internally, even when they don’t say so publicly.

What the Next Few Weeks Mean for AI Users

The pace of updates right now is genuinely fast. Google IO is just weeks away, and the recent Gemini improvements suggest Google has more to show. OpenAI’s Codex is still fresh in its knowledge-work role and will likely keep adding features. Anthropic has shown it doesn’t sit still when competitors move quickly, as seen in recent Claude model upgrades.

The most practical takeaway from this week is simple: AI tools for everyday work are getting easier to use, more organized, and more capable of handling real tasks without requiring expert-level prompting. That’s good for anyone who has found these tools useful but frustrating. The friction is shrinking, and the results are getting better. The smart move right now is to try Codex alongside whatever tool is already part of the daily workflow and see which one fits better in practice.

Frequently Asked Questions

What is OpenAI Codex and how is it different from Claude Co-work?

Codex is OpenAI’s AI agent that handles chat, coding, and knowledge work all in one interface, without switching between modes. Claude Co-work is Anthropic’s dedicated tab for knowledge work tasks like email, documents, and automations. Both tools now cover similar ground, but Codex stands out for its live task-tracking feature that surfaces unfinished actions as clickable follow-up items during a chat session.

Is NotebookLM free to use with the new label feature?

The auto-label feature for source organization appears to be rolling out to paid Google AI subscribers first. The notebook organization system itself, including access within Gemini and mobile support, has been made available to all free and paid users in most regions. Europe and some other regions may receive certain features later due to local regulations.

Does Gemini’s file creation work on a phone?

Yes. Gemini can now create and download files including PDFs, Word documents, Excel sheets, and PowerPoint decks directly on mobile. It can also create Google Docs and Sheets that open straight in Google Drive. This makes it practical for anyone who needs to produce a document or slide deck while away from a desktop.

Why does OpenAI recommend shorter prompts for GPT-5.5?

GPT-5.5 is better at reading intent from a brief goal statement than earlier models were. OpenAI’s updated guidance says users should describe what they want to achieve and let the model determine the best path. This mirrors a similar trend with Anthropic’s Claude models, where over-explained prompts with too much instruction can actually limit the model’s ability to find a better solution on its own.

Which is better for image generation, GPT Image 2 or other image models?

GPT Image 2 produces strong results when generating images from a text prompt alone, with consistent quality and creative output. For editing an existing photo, such as changing a facial expression or adjusting specific details, other image tools perform much better. The practical rule is to use GPT Image 2 for original creation and a different tool when the task starts with an existing image.

What does Spotify’s AI music labeling actually do?

Spotify’s “Verified by Spotify” label will mark content to show users whether it was created by a human or generated by AI. The goal is to make it easy for listeners to know what they’re hearing. This addresses growing concern about AI-generated music appearing without disclosure, and many observers hope similar labeling will spread to video and social platforms.

How does NotebookLM’s auto-label feature help students?

Students can keep all sources for an entire course inside one notebook and assign labels to each topic or subject within the course. When studying for a specific test, they select the relevant label and work only with those sources through chat, audio overviews, or flashcard creation. At exam time, all labels can be selected together. It removes the need to juggle multiple notebooks or manually hunt for the right source files.