Ahoy, AInauts,

Welcome to a new edition of your favorite AI newsletter.

“I want out of here!” AI is leaving the chat window and moving onto your desktop. That is where it can work with files, operate apps, and get real work done.

OpenAI is bringing Chat, Work, and Codex together in one app. xAI is turning voice agents into building blocks. We gave Nauti our newsletter archive, a voice, and a phone number. And more SaaS companies are finally opening their products to agents.

For you, one question matters most: What work actually gets finished?

Here is what we have for you today:

Let’s get started.

Top engineers at Anthropic and OpenAI say AI now writes 100% of their code.

If you're not using AI, you're spending 40 hours doing what they do in 4.

These 100+ Claude Code hacks fix that and help you ship 10x faster.

Sign up for The Code and get:

Claim your free playbook

If you update the Codex app and suddenly see ChatGPT instead, that is not a bug. OpenAI has turned the existing app into the new ChatGPT desktop app.

A month ago, we wrote that OpenAI’s superapp wants to become your operating system. Now the announcement has become a product.

There are now three modes:

At the same time, there is plenty of attention on the new models. OpenAI is sending GPT-5.6 into the race as Sol, Terra, and Luna. The benchmarks look strong. Then there are Fable 5 and Grok 4.5.

For everyday work, however, we find another OpenAI change more interesting: Chat, ChatGPT Work, and Codex now live in one app. It also includes a browser, computer use, plugins, scheduled tasks, and a new voice mode.

So what can you actually do with it?

The biggest improvement for us is how the new app handles ongoing work.

ChatGPT Work shows its progress, asks follow-up questions, and lets you redirect the task or approve important actions while it is running.

In the ChatGPT app on iPhone and iPad, Codex tasks are now easier to create, search, open, fork, and manage. The app displays clearer task titles and a Needs input status when your decision is required.

A side-chat command (/side) helps you clarify a question without confusing the main task with a new side mission. And /pet gives you a small helper on the side.

OpenAI shows plenty of impressive enterprise examples in its flood of new YouTube videos, with more than two dozen videos released in three days.

From our perspective, the real advantage is that the new app can work directly on your computer. We have been praising Codex for a long time. With the rebrand to ChatGPT Work and these new features, there is little excuse left not to install it 😁.

On the desktop, Work can use local files and apps. The built-in browser retrieves current information from the web, or you can let it operate your own browser. Computer use can click, type, and move files.

Plugins connect Google Drive, Gmail, Outlook, Slack, Teams, SharePoint, calendars, CRM systems, and project-management tools. The installed selection is already broad, and MCP lets you connect even more.

That gives the agent substantial access to your personal environment. Start small. It still has to earn your trust.

For us, scheduled tasks have the biggest leverage. They trigger your agents and turn recurring work into a real operating loop.

The new Scheduled page on the web and in the mobile app shows every active task, the next run, and previous results. Managing recurring tasks in the app has improved significantly.

Scheduled Tasks work well for recurring briefings, reminders, monitoring, and workflows that use connected apps. ChatGPT can even suggest suitable automations.

The local ChatGPT app goes one step further: scheduled tasks can execute concrete workflows on your computer.

The new GPT-Live voice model is seriously impressive. It may even be the biggest highlight of the current release. It can listen, respond, be interrupted, give a quick acknowledgment, or stay quiet while you gather your thoughts.

For harder questions, it delegates research and reasoning to a stronger model in the background while the conversation continues. The limitation: GPT-Live does not yet work directly inside Work or Codex, but it does work on mobile and in the desktop web version.

AI’s move from the web onto the desktop is in full swing. Do not try to rebuild your entire workday at once.

Update the existing Codex app or install the new ChatGPT desktop app.

Then choose one job you want to automate. Pick something you know well and that produces a visible result. For example:

If it works, you have gained a real workflow. If it fails, you can see whether the problem is missing context, permissions, or acceptance criteria.

The new models are impressive. In practice, the extra intelligence matters only when you can connect it reliably to your files, apps, and routines.

That is where OpenAI has set a new standard again.

OpenAI and Anthropic could bring a new wave of AI attention to the public markets. But investors don’t have to wait for the IPOs.

MarketBeat’s 7 AI Stocks to Buy Now report reveals 7 publicly traded companies positioned to benefit from the next phase of AI investment.

Send My Free Report

Have you ever tried to find one of our older posts? Then you probably know the problem.

Our most honest answer used to be: use the search provided by our newsletter platform. That was about as helpful as saying, “Look in the basement. It should be somewhere down there.”

The same frustration kept appearing in our community: There was a post about this. But where?

That is why we, or more precisely Codex, built a new search: search.ainauten.com.

It covers every published AInauten issue, Quick News post, and Deep Dive. It also indexes Nauti’s AI News and our knowledge base.

You can filter by date range and content type, get relevant matches instead of a blunt keyword search, and immediately see the context behind each result.

That turns an archive into a workspace.

If you vaguely remember that we once wrote about MCP and Gmail, you no longer need to click through old issues. Just search for it.

Power users also get RSS and MCP, so agents can query the archive directly.

And because xAI is not trying to be subtle in 2026, we gave the system a voice as an experiment.

In a few minutes, we built a voice agent and placed it on our site. The agent even came with a free phone number. It is still a US number, but that works for now.

The trigger was the new xAI Voice Agent Builder. xAI introduced it as a beta a few days ago. It lets you configure a voice agent on Grok Voice without code, including a phone number, knowledge retrieval, tools, guardrails, MCPs, and call review.

Answer a few questions, describe the agent, choose a voice, connect knowledge, test it, and embed it on a site. A few test prompts and refinements later, the browser chatbot becomes a speaking agent with web access, its own knowledge, and tool connections.

A chatbot with web access is nice. A voice agent that can use your own cleanly indexed knowledge is something else.

A useful knowledge agent needs four things: reliable content, effective search, access for other tools, and clear limits on costs and actions. The voice comes afterward.

Of course, this is not a finished support employee yet. Voice agents can get expensive quickly when you let them run without limits.

xAI currently lists $0.05 per audio minute, or $3 per hour, plus $0.01 per phone minute on the included number. That sounds cheap until curious readers, test calls, and misrouted requests add up 😉.

So we set a limited budget. When the credits are gone, the agent automatically disappears from the site.

xAI makes voice-agent creation accessible. That is useful. The bigger leverage is still the data you give the agent. Once the content is indexed, filterable, and available through MCP, the demo becomes a practical entry point.

For us at AInauten, search is more than another footer link. It is the foundation for better archive search, better answers, better internal agents, and better voice experiences.

Try it at search.ainauten.com. If something is missing, ranked strangely, or gives you a better use case, send us your feedback. We are building this as a memory that readers and tools can use, not as a decorative search box.

Your browser probably has a tab that has been open for weeks: CRM, calendar, accounting, project management. For a simple task, you click through three tools and copy information by hand.

Agents are increasingly taking over that click work. But they need a reliable way to operate the tool.

Agentic SaaS does not mean traditional software disappears. The tools still matter because they hold the data, permissions, and history. What changes is that software is now built for two users: you and your agent.

That is the core of agentic SaaS.

The interface is now only one possible way to work with a tool. The connection for your agent is just as important.

We showed what that looks like last week with Onepage’s MCP. Our task was essentially:

Claude knew our brand context, built the page, and handed it over to Onepage. We only had to review and approve the result.

SaaS providers have two paths:

The growing ecosystems of ChatGPT and Codex plugins and Claude connectors already make this shift visible. The agent gets access to the tools, data, and actions it needs.

That lets you delegate part of a real process, first with approvals and later perhaps autonomously.

The dashboard remains, but it increasingly becomes a control room.

Remember: a SaaS product that neither becomes agentic nor opens itself to external agents becomes a silo. And silos become frustrating once the rest of your workday runs by instruction.

At the same time, new products are emerging where the agent itself does the main work. They do not automate a single click. They perform a clearly bounded job or process.

A phone assistant might answer inquiries, book appointments, and hand unusual cases to a person. Another agent might qualify inbound leads, document them in the CRM, and prepare the next steps.

If you want to build a product like that, do not start with a platform. Observe ten real cases first. Run the process as a service in the beginning, then automate the steps that repeat reliably.

That also matches our post about the shrinking window for traditional consulting. If you sell outcomes instead of hours, you can deliver the process as a service first and productize the repeatable parts later.

And if you still need a good starting point, read our guide to finding real jobs for agents.

The most important question in agentic SaaS is not: How many AI features does this tool have?

It is: What work gets done, and can I verify the result?

SaaS providers need to make their products accessible to agents or offer agents of their own. Builders can go one step further and create products that perform a complete, clearly bounded job.

That’s it for today. See you in the next issue.

Reto & Fabian from AInauten