Hello fellow keepers of numbers,

Some big news this week, and for once it’s about us. The AI Accountant is now The Appreciable Asset.

New name, new logo, new colors. Everything else is exactly what you subscribed to receive. The same takes, same demos, and same weekly read on the AI news that matters for our profession.

On a personal note for a second: I started this newsletter because the AI news cycle is a firehose, and none of it is written for us. And if I’m honest, I also started it just to make myself stay on top of everything. I love writing and learning new things, and if I can bring others along with me, even better.

I hope everyone reading this each week feels like they learned a little bit, feels less intimidated about all the new AI updates, and feels enough excitement to jump into a new tool and start testing things. I’d like to think those who do are better off for having invested in improving their mind, the appreciable asset.

In this week’s newsletter, we explore the new ChatGPT app with the newly released GPT-5.6. Claude Cowork is also coming to mobile and web. Plus, stick around for a demo of the brand new in-app browser in Claude Code compared with the in-app browser in the ChatGPT app.

The Latest

OpenAI launched ChatGPT Work, an agent that runs multi-step projects across a user's apps and files, and merged its Codex app into a rebuilt ChatGPT desktop app.

The new desktop app puts Chat, Work, and Codex in one place, available worldwide on Mac and Windows across every plan, including Free. Codex stays the coding agent for developers, and the older desktop app becomes the ChatGPT app.

ChatGPT Work itself is also rolled out in the app. It’s coming to web and mobile, starting with Pro, Enterprise, and Edu plans and reaching Plus and Business within days. It turns one request into finished spreadsheets, slides, and docs, connects to tools like Slack, Teams, Google Drive, and email through plugins, and can run on a schedule through Scheduled Tasks. On desktop, it also uses a built-in browser and can control local files and apps directly.

The agent runs on GPT-5.6, which OpenAI made generally available the same day across ChatGPT, Codex, and the API. The Sol, Terra, and Luna models had been previewed in late June under a limited, government-requested release that kept them from most users until now. Pricing runs per 1M tokens at $5 input and $30 output for Sol, $2.50 and $15 for Terra, and $1 and $6 for Luna.

OpenAI says its own finance team uses GPT-5.6 inside of ChatGPT Work to cut month-end close and forecasting from days to hours, having it gather source data, reconcile it in Excel or Sheets, build slides, and verify the results.

Why it’s important for us:

This was a big announcement with a lot of separate updates, so let’s break them down. Fair warning, this is a lot. But I think it’s all important.

GPT-5.6:

We already knew about these models from the announcement a few weeks ago. They’re now deployed. Early testing says that Sol is really good. Terra is fine. Luna is probably not very usable unless it’s a subagent (similar to Haiku in Claude).

ChatGPT app:

This is the biggest update by far. They’ve officially merged what was previously the ChatGPT app and the Codex app. They’re now one combined app.

If you’ve never used Codex, you now get one app with the chat you’re used to plus the agentic AI that’s comparable to Claude Code and Cowork. The app underneath is excellent. Far better than the Claude desktop app. More on this below in the features section.

But there’s a major problem with the app for anyone coming from regular ChatGPT. Chat is now a small tab on the left sidebar in the ChatGPT app. When you click it, a small window pops open to send your prompt. As of now, projects inside of chat seem to be gone (at least inside the desktop app). Existing projects don’t seem to be migrated into the app, and you can’t create new ones.

Projects are now entirely inside of ChatGPT Work or ChatGPT Codex (more below). And they’re not the same projects you’d be used to. They’re essentially just a folder on your computer. Custom instructions now have to live inside an AGENTS.md.

This is a major miss by OpenAI. They’ve tried to create an app to simplify life for non-developers, and instead they’ve raised the learning curve very significantly. If your main takeaway from this announcement and the new app is that you’re confused, you’re not alone.

ChatGPT Work:

This is OpenAI’s direct response to Claude Cowork. It’s another big miss.

First of all, ChatGPT Work is the same as what the Codex app was previously, except the AI model doesn’t present code to you while it’s thinking. It’s essentially just a small difference in the interface to make things seem less scary. That’s fine, I guess.

The huge miss here is the simplicity and guardrails. I want to compare this to Claude Cowork because I think that’s OpenAI’s ultimate goal.

Claude Cowork takes the power of Claude Code and slaps some major guardrails on it. Users can be limited to certain connectors, must approve Claude actions, and, in some cases, are limited in what Cowork can grab from the web. ChatGPT Work doesn’t have nearly the same level of guardrails.

Claude Cowork projects give users the ability to edit project instructions, save and view their memories inside the project, link multiple folders and see exactly which are linked, and create scheduled tasks specific to each project. ChatGPT Work can do literally none of this natively in the interface. Instead, a user is expected to create and maintain an AGENTS.md for each folder, tell the automations specifically which project folders to use, ChatGPT can easily escape the folder and work elsewhere on the computer (because this is what OpenAI intends), and maintain project-specific memories manually in a separate subfolder under each project folder. 

Again, if this seems confusing, it’s because it is. If you’re already familiar with the Codex app, or even Claude Code, this probably isn’t a big deal. But for those just now learning, it’s massive. The learning curve for Claude Cowork is steep but doable. The learning curve for ChatGPT Work seems very significant.

Feature updates to the app:

Despite all the terrible decisions they made in this rollout, their app is still awesome. The ChatGPT app (fka Codex app) has so many amazing features that make it light-years ahead of the Claude desktop app.

The in-app browser is by far the biggest differentiator. It’s essentially a Chrome that you can open inside of ChatGPT, and the AI model can see and use the in-app browser. It’s really good. Like good enough that it’s worth using the ChatGPT app just for this. Who cares if you don’t have a connector? Just pop open the browser and tell ChatGPT to use it. Equally as good.

Computer use has significantly improved as of this update. I’ve yet to test this, but others are saying it’s a significant improvement when combining the new ChatGPT app with the new GPT-5.6 models.

Sites in the ChatGPT app (very similar to Artifacts in Claude) have improved. The capabilities of the new GPT-5.6 models also make this more interesting.

Lastly, you can now edit Markdown files that ChatGPT creates directly inside the app. No need for other apps just to edit Markdown files anymore. Big upgrade.

Anthropic brought Claude Cowork to web and mobile, after previously offering the task-running tool only on desktop. Beta access rolls out over the coming weeks, starting with Max subscribers and reaching more plans later.

Sessions now carry across devices, work continues after the laptop is closed, and scheduled tasks run with no device online. Claude sends decisions that need a human call to the user and does not ship work until it has been reviewed and approved.

On web and mobile, sessions run on Anthropic's servers and handle connectors and cloud files on their own. Local file access, browser use, and computer use require the Claude desktop app open on that machine.

Anthropic doubled Cowork usage limits through August 5 to mark the launch.

Why it’s important for us:

Anthropic hasn’t been trailing OpenAI in many areas over the last few months, but this was one of them.

OpenAI already lets you use the ChatGPT app (fka Codex) on your phone to work with your files, plugins, and skills in the same way you do at your desk. Claude’s version of this was Dispatch, which was extremely clunky. They shipped Dispatch fast and then pretty much abandoned it, and it wasn’t very fun to use.

Now the phone will actually work with Claude. I’m excited for this update, and I think it’s probably going to fly under the radar, just like the comparable update from OpenAI.

Your files, connectors, and skills travel with you now when you’re away from your desk. I don’t know if this will be a highly used feature from day one. It probably won’t. But over time, I think this becomes a much bigger deal as we grow used to the capabilities of powerful AI in our pockets that are connected with our computers.

Trending News

Claude Code added an in-app browser on desktop, so the agent can open and click through docs, designs, or any site without leaving the app: This one came out of nowhere, and right after I typed up how the in-app browser in the ChatGPT app is a major differentiator over the Claude desktop app. TBD how good Claude is at using this, but this is a major update. AI companies might just own the browser market now.

Wolters Kluwer expanded CCH Axcess Expert AI to read tax source documents and auto-populate CCH Axcess Tax, from W-2s and 1099s down to messy K-1s: Could be a nice win if it rolls out accurately, but I'm skeptical. CCH hasn't done much well in this space so far, and the announcement itself was mostly AI and automation buzzwords that added up to a bunch of nothing.

SpaceXAI released Grok 4.5, its first model built with Cursor, matching Opus 4.8 and GPT-5.5 on coding at a fraction of the price: First big announcement since the Cursor acquisition. Grok used to be a bit of a joke, and now it's a legitimate competitor. I suspect it'll be tuned to run well inside Cursor, which is probably due for its own update soon, the way the Claude desktop app did previously and the ChatGPT app just did. Grok is finally one to watch.

Anthropic extended included Fable 5 access on paid Claude plans through July 12, after saying it would end and move to usage credits: Maybe they'll just keep extending our Fable 5 access five days at a time. Probably not, but maybe. I could actually see them making it permanent now that GPT-5.6 is out and looks competitive with Fable.

Xero launched Industry Benchmarks, turning anonymized customer data into peer comparisons across nine health metrics with an AI-recommended action scorecard: The thing to watch is that this starts to infringe on advisory for any firm that sells benchmarking in its reporting, and clients who use Xero's analytics will now see these numbers themselves. Overall a cool feature though.

OpenAI made its ChatGPT PowerPoint add-in generally available worldwide, building and editing decks from your notes, files, or an existing deck: Just a move from beta to generally available. OpenAI's still a little behind on PowerPoint. Hopefully the newer GPT models make it better at slides and design overall.

Anthropic showed a pattern that runs Sonnet 5 as the cheap worker and calls Fable 5 only for advice, keeping most tokens at the lower rate: We've talked about this setup before. A stronger model sits in the orchestrator seat and hands the grunt work to Sonnet 5 subagents, which is exactly what Sonnet 5 seems built for. Expect more of it, either automatically behind the scenes or easier to spin up yourself, with tasks finishing on less usage at the same accuracy or better.

OpenAI introduced GPT-Live, a voice model that lets ChatGPT listen and talk at the same time, handling interruptions and pauses: Natural voice makes AI actually usable in the car or between meetings, which is where a lot of thinking happens anyway.

Meta introduced Muse Spark 1.1, a reasoning model with computer use and MCP support, now available in a public preview: Another laggard catching up, like Grok. Not enough steam behind it to take seriously yet, but worth keeping on the radar. For now it's stuck in the middle. The high-end models are better and the open-source models are nearly as good and free.

Put It to Work

In this week’s demo, I did nothing. Literally. I put a prompt into the Claude app and ChatGPT app and let them both run and use their in-app browser to pull data out of 10-Ks and put them into Google Sheets.

I saved no time since I sat there and watched. But it was still pretty cool.

Weekly Random

Is chat dead?

Don’t click away just yet. I think I can convince you it might be.

It’s a weird thing to wonder about for the feature that started the AI we know and love (or some who really, really hate). Chat was the on-ramp to the fastest highway of all time that got us to where we are now. And this week, the two biggest AI companies in the world built the exit ramp.

The new ChatGPT app launched around Work and Codex. Chat, the thing after which OpenAI named their product, got demoted to a little floating window that doesn’t even connect to projects right now. OpenAI could’ve made it another toggle alongside Work and Codex, but they chose not to. Maybe they will after some of the backlash they’re getting, but the point still stands. They’re pushing users to agentic AI, not chat.

Just a few days ago, Claude launched an update to the desktop app that folded the Chat and Cowork tabs into one Home tab. Now, you start a new task and choose between a Chat or Cowork toggle.

Neither company fully removed chat functionality. It would’ve created an uproar. But both made some major UI changes to push users toward ChatGPT Work and Claude Cowork.

Personally, I haven’t used chat for probably 6 months. Anything that used to go in the chat box now goes into Cowork or Codex (now, ChatGPT Work). Even the quick things that used to be Google searches. It’s where I already work, and if a 30 second question turns into real work, I’m already in Cowork within a task that has the context it needs.

The only real knock on my method is that it uses slightly more usage than a plain chat might. But if it takes up 0.2% more of my usage window than chat would’ve, I’m fine with that tradeoff.

So is chat dead? For plenty of people, it’s not yet. But for me, it is. And I know for a lot of others it is as well. It might still have a little road left, but that exit ramp is approaching fast.