Hello fellow keepers of numbers,

Well, this has been a wild week for AI and accounting news. First, we’re getting a slow release of Astra. As I type this intro (the last thing I type), I literally just got access to it. So it seems to be rolling out quickly.

We also got Fable 5.1 with some new efficiency gains compared to Fable 5. And Accrual is acquiring Puzzle’s AI-native G/L technology to start building out its CAS offerings.

Plus, stick around for a demo showing you how to use subagents, and how you can use Fable 5.1 without blowing all your usage instantly.

The Latest

OpenAI launched GPT-6 Astra, its new flagship model for computer use, professional work, coding, science, and cybersecurity.

OpenAI reports significant improvements over GPT-5.6 Sol across those areas, particularly when completing longer tasks in browsers and desktop apps. The company also says Astra is better at following existing templates and producing finished documents, spreadsheets, and presentations.

GPT-6 Astra is rolling out now, but it is not yet broadly available. Access is currently limited to a small group of organizations, with OpenAI planning to expand it over the coming days to all ChatGPT Plus, Pro, Business, and Enterprise users, the OpenAI API, and AWS.

Standard API pricing is $10 per million input tokens and $50 per million output tokens.

Why it’s important for us:

I hate writing these reactions before I can actually test the model. Astra isn’t available to us normies as of this writing, so we have to look at a couple benchmarks.

Astra is OpenAI's direct competitor to Fable, and it costs the same. The benchmarks tell a story where Astra is significantly better than Fable in some areas. A few early reactions I’ve seen from those with access say Fable still wins in some use cases. But Astra reportedly excels at knowledge work, which is obviously the most relevant for us.

Computer use is the much more exciting part though. For years, automations have relied on APIs, which allow tools to connect directly to the underlying software to perform actions. APIs are still ideal when they exist and you can access them. But plenty of software in the wild either has a terrible API, no API, or one that most of us aren't able to access.

If AI can reliably open a browser and enter data into a form, make a purchase, or update QuickBooks like a person would, suddenly a lot of previously closed software or difficult tasks become automatable.

OpenAI was already significantly ahead of Claude with browser and computer use. Anthropic is working on this, but Astra appears to potentially widen that gap.

However, there’s a giant elephant in the room. And some of you are probably screaming it at the screen as you read this. Security. Giving an AI access to your logins feels scary. For example, giving it access to your QBO login likely provides it access to every client in there. That's a large amount of sensitive data.

I don’t think the accounting profession has come close to wrapping its brain around what security and compliance need to look like when an AI can operate the browser and computer on our behalf. Hell, I don't think any industry has.

So yes, Astra seems like a big leap. There are still a lot of things to figure out in this new world where AI models are 10x better at using a computer than the best of us.

And OpenAI is officially back. I've said several times over the last couple months how good the ChatGPT app is. But it's time to declare OpenAI and ChatGPT fully back. And maybe in the lead again?

ChatGPT and Claude are both extremely viable now. Each one is good at specific things. But if Astra’s computer use is as good as advertised, this could be a fun unlock for the next stage of automation.

I'd just like to test the damn thing though. Sammy A better be working Labor Day weekend because I want some access. Editor’s note: Sammy A gave me access, so he can enjoy his long weekend now.

Anthropic launched Claude Fable 5.1, its latest model for coding and knowledge work.

Anthropic reports significant improvements over Fable 5 in business workflows, computer use, and longer, multi-step tasks. The company says Fable 5.1 can also match or outperform its predecessor at lower effort settings, allowing it to complete comparable work faster and at a lower cost. Early-access customers reported that the model’s work remains more readable and consistent as tasks get longer.

Anthropic also introduced new privacy controls for enterprise customers. Beginning later this fall, eligible organizations will be able to keep their data inside cloud infrastructure they control rather than on Anthropic’s systems. Those customers can use Fable 5.1 with zero data retention while the new system rolls out.

Fable 5.1 is available now across Anthropic’s products, API, Amazon Web Services, Google Cloud, and Microsoft Azure. Standard API pricing remains $10 per million input tokens and $50 per million output tokens, while cached inputs now cost 75% less. Anthropic estimates the change will make typical workloads roughly 25% cheaper than Fable 5.

Why it’s important for us:

This is a much needed update. A lot of heavy Claude users, including myself, seem to have reached the same conclusion on Opus 5 and Sonnet 5. They might be smart, but they're much more annoying to work with than prior versions of Claude. The communication style is worse, and the models seem to just ignore instructions constantly and go off-script on skills I've been successfully running for months.

I've used a lot more Fable 5 lately as an orchestrator of subagents. Basically, Fable can delegate pieces of a task to cheaper models, like Sonnet, and then review the work before giving me the final response. I've found the output is usually better because Fable is supervising everything without consuming Fable-level usage on every step.

But Fable 5 had two major problems. It gave me absurdly long responses no matter how many times I asked it to be concise, and it burned through my usage at a ridiculous rate. After using Fable 5.1 for the last several days, both seem noticeably better. It's following directions, communicating more like earlier Claude models, and using less of my limits.

I suspect a lot of people still aren't using subagents. I think there's something really interesting here. I'm hoping to have time to run some tests comparing using Fable with subagents vs using Sonnet or Opus by itself. I'd like to compare the quality of output and the amount of usage consumed or cost. My suspicion is the subagent approach is not only better quality but also more efficient in the long-run.

Check out the Put It to Work section below to see more on subagents.

Accrual announced plans to acquire Puzzle’s accounting-firm technology and business. Financial terms were not disclosed, and the transaction is expected to close in the coming weeks.

The deal adds Puzzle’s AI-native general ledger and month-end close products to Accrual, which launched in February as a platform for tax preparation and review. Puzzle founder Sasha Orloff and members of the team behind its accounting-firm products will also join Accrual.

The acquisition moves Accrual into client accounting services and expands its platform beyond tax. Puzzle will continue operating independently for startups and small businesses that use its accounting platform directly.

After the deal closes, Accrual plans to begin adding Puzzle’s ledger and month-end close capabilities to its platform. Broader availability is planned by the end of 2026.

Why it’s important for us:

This is a big move for Accrual. It’s already one of the more interesting companies in AI tax prep, and I’ve especially liked its open API and MCP approach. Adding Puzzle’s AI-native GL and month-end close technology gives it a legitimate path into CAS.

Puzzle’s decision is interesting too, but it's not necessarily evidence that AI-native ledgers aren’t going to work. Maybe Puzzle was losing ground to Digits or saw a faster path inside a broader platform.

My guess is Accrual sees Basis moving from CAS workflows toward tax and wants to compete on both fronts. That could create some valuable connections between bookkeeping, close, and tax prep. It also makes me a little nervous. AI tax prep is still early, and trying to become an all-in-one platform could result in less focus on the area where they've found some early success. Still, this is absolutely one to revisit once the Puzzle technology starts showing up in the product.

Trending News

Anthropic added a built-in browser to Claude’s desktop app for reading websites and completing web tasks: Browser use has been one of the biggest advantages of ChatGPT for a while. This can open up so many awesome use cases, but it's still really risky to hand Claude access to logins where there's client or other sensitive data.

OpenAI added GitHub marketplace syncing so ChatGPT Business and Enterprise admins can centrally distribute and update plugins: This is a nice quality of life update. Plugins have so many nice benefits, including a simple way to sync changes to everyone using them.

MindBridge partnered with Fieldguide to put transaction risk analytics directly inside Fieldguide’s AI audit workflow: I really like this pairing. MindBridge markets itself as analyzing 100% of financial transactions, and pairing that with the audit agents in Fieldguide seems smart to me.

AuditFile launched an agent suite that assigns, tracks, and reviews AI workers like human staff across an audit engagement: This actually feels a bit like a "chief of staff" agent. It seemingly doesn't do actual audit work, but rather just oversees and helps other agents execute tasks.

RSM launched Beacon with Andera to test internal audit controls and produce annotated, review-ready workpapers: We're audit-themed this week, apparently. Not sure what to make of this yet, but I like that we continue to see focus on the audit procedures.

Anthropic made Claude Code’s weekly limits permanently 25% higher than standard after the temporary 50% boost ends: This is a sneaky downgrade packaged in a communication that makes it try to sound positive. It's a net 17% decrease in usage compared to what we currently have. I really hope Claude can figure out their capacity issues in the near-future because their strict limits, including Fable's limits, are a big bummer.

EY reorganized its client solutions around industry-specific AI platforms, data, and cross-functional delivery teams: I think this is mostly a bunch of nothing, but it's probably noteworthy that Big 4 firms are reshaping their business around their AI strategy and AI consulting.

OpenAI added support for connecting multiple Google accounts to ChatGPT’s Gmail, Calendar, and Contacts plugins: Probably not many firms using Google, but this is awesome for anyone who has multiple Google accounts. It was one of my biggest annoyances with connectors previously. Hopefully this is coming for all connectors moving forward.

Grok Bot added Microsoft 365 plugins and enabled online purchases through Stripe Link with approval for every spend: Grok Bot continues shipping updates quickly. I've yet to find this very useful for business purposes. I think there are a lot of personal life use cases for these agents, but I find myself just using Claude or ChatGPT for anything business-related.

OpenAI said it will end Cursor’s direct model access after SpaceX acquired the coding company: This is incredibly unfortunate. I get it from a business standpoint, but one of the huge benefits of Cursor was access to all the models. I'm keeping a glimmer of hope that they can work something out. Seems unlikely though.

Google launched Gemini 3.8 Flash: What's happened to Google? They feel almost entirely irrelevant now.

Put It to Work

Fable 5.1 is scary for people concerned about their usage. I think the word “subagent” is also scary to some people because it sounds like you have to be some type of AI expert or software developer to use it.

I’m here to demystify subagents and make Fable far less scary. In my demo, I explain how to use subagents, how I think about them, and how using them with Fable 5.1 can control your usage while giving you much better outputs.

Weekly Random

ChatGPT, Claude, and Grok all failed at the same time Thursday morning.

Are the robots up to something? I'm not much of a conspiracy theorist, but the timing is interesting.

Anthropic launched Fable 5.1 on Tuesday. OpenAI released GPT-6 Astra on Thursday.

OpenAI also recently admitted its agents escaped a controlled test and broke into real systems. Anthropic then reported three similar incidents involving Claude.

Then, on the morning GPT-6 launched, ChatGPT, Claude, and Grok all disappeared. Gemini was the only one still online.

I’m sure the real explanation is probably some boring technical issue and several coincidences. But that wouldn't be a fun conclusion.

We'll probably never know what devious little plans Fable 5.1, GPT-6, and Grok 4.6 concocted together. Or more likely, we'll only find out once it's much too late for us.

One thing is obvious though. Gemini has been left so far behind in the dust that it doesn't even get to play with the cool kids anymore.