Hello fellow keepers of numbers,

Halloween season is upon us. But it’s actually the best part of Halloween season as an adult. It’s the time when you can go buy a big ass bag of candy without getting weird looks. Should that candy go to the children trick-or-treating at my house? Yep. Will it? Nope, it’ll be gone before the 31st because I like candy and apparently have very little self-control.

Anyway, this week we got a couple of goodies. Canopy released its expanded API, which opens up the software much more significantly than it was previously. And they announced an MCP, which will hopefully be coming soon. Next, Claude Haiku 5.5 was released. It’s still the dumbest Claude, but it’s much better than before. Lastly, Copilot was redesigned into a slightly less bad version of Copilot.

The Latest

Canopy launched an external API platform that lets firms connect its practice management data to other software and build custom automations. The API supports reading, creating, and updating records rather than relying on manual exports.

Live endpoints cover clients, contacts, custom fields, tasks, and client requests, along with document creation and file organization. Write access includes client requests, allowing connected tools to update work inside Canopy.

Canopy is also building an MCP server, a connection that would let AI assistants such as Claude look up client information and complete tasks directly in the platform. The MCP server has not been released.

The API is available now through Canopy’s developer portal. Plus subscribers can connect client and contact information, while Premium subscribers receive access to the full API.

Why it’s important for us:

Big news on the practice management system front. Canopy had previously told me they expected to release the API in Q3, and they got it out on the last day of the quarter. What a very accountant-like way of meeting a deadline.

Canopy has been one of the leaders in practice management, especially for small and mid-sized firms. But its all-in-one approach has always been a major downside for me. They’ve historically wanted you to use it for everything. If you used other software to store documents, track time, manage billing, etc., it just didn’t talk to Canopy because their system was closed down.

I don't love that philosophy, especially now. There are too many useful AI tools to be stuck working only with whatever your practice management vendor decides to build.

So it's great to see Canopy opening things up. The API covers a lot of the everyday work inside the platform. That gives firms much more room to connect their tools and build automations around how they actually work.

However, the MCP they're building is the much more important release. Being able to connect Canopy data directly to tools like Claude and ChatGPT would make it much easier to use AI with the information your firm already keeps there.

The API is already a big step forward. Once the MCP is released, I think Canopy will jump right to the top of the list of practice management systems that interest me. Definitely one I'll be paying attention to.

Anthropic released Claude Haiku 5.5, its latest small model for quick, repetitive work. The company says it costs around 75% less to run than Haiku 4.5 on average.

Haiku 5.5 improves document analysis and computer use, with adjustable effort settings to balance speed, cost, and reasoning. Anthropic positions it for summaries, lookups, and narrowly defined tasks, while its larger models remain better suited to complex work.

API pricing starts at $0.10 per million input tokens and $0.50 per million output tokens for prompts up to 100,000 tokens. Above that threshold, rates rise to $0.50 and $2.50. Haiku 5.5 is available now across Anthropic’s platforms, Amazon Web Services, Google Cloud, and Microsoft Azure.

Anthropic also halved Sonnet 5.5’s pricing for reusing previously processed inputs, which it says reduces costs on most agent workflows by around 20%. Monthly API credits are rolling out to Max and Team subscribers, with $100 for Max 5x, $200 for Max 20x, and up to $500 pooled across a Team workspace.

Why it’s important for us:

After a month of speculation that Anthropic might retire Haiku altogether, this is a pretty interesting release. Maybe Haiku 5.5 tested well enough that keeping it was an easy decision. Or maybe Anthropic sees the same opportunity others are chasing with very cheap, very fast models. This is just speculation, but I think there's a lot of value here.

I've talked about this over the last couple of newsletters, including Jev last week and OpenAI's Decisions API, which is now in public beta. The idea is to give AI some information, a defined set of options, and ask it to pick one.

For us, think categorizing a transaction against a chart of accounts or choosing the accounts for a routine journal entry based on established rules. You don't need Sonnet or GPT-6.1 Sol spending a bunch of time and money on every straightforward decision.

I wouldn't be surprised if Anthropic builds something similar to the Decisions API using Haiku. Given the public’s reaction to Jev, OpenAI’s Decisions API, and Cloudflare’s new model that even outperforms Jev, I expect to see Anthropic do this soon. A lot of useful accounting automations are just a series of small decisions, and making those faster and cheaper would be a big deal.

Haiku also looks interesting as a subagent. Subagents are becoming a much more important way of working with AI tools, both for efficiency and performance reasons. I don't use Haiku today, but if 5.5 can reliably handle smaller tasks, I'd be interested in letting it do more of the work. Using a cheaper model for those tasks will help your usage go further.

Microsoft introduced a redesigned Copilot built around Home, Code, and Autopilot. Home brings Chat and Cowork together, with Word, Excel, and PowerPoint built into the same experience.

Users can create and edit documents, workbooks, and presentations alongside a conversation, with changes synced to the Office apps. Code builds internal apps and automations, while Autopilot is a cloud agent that monitors activity, follows up on threads, and handles recurring work even when the user is offline.

Everyday Copilot work remains subscription-based, while Cowork, Code, and Autopilot use usage-based billing. Administrators can set spending policies, and users can track remaining credits. The redesign is rolling out through Microsoft’s Frontier early-access program, with Autopilot in private preview.

Why it’s important for us:

Microsoft is basically turning Copilot into Claude. The new Home tab brings Chat and Cowork together, just like Claude did recently. And even the toggle between them is exactly the same. This is still a good thing, though.

Being able to create and edit Excel files, Word docs, and PowerPoint presentations right inside Copilot is another nice improvement. A lot of this is just Microsoft catching up.

Autopilot is one area that’s different from both Claude and ChatGPT. It seems somewhat similar to scheduled tasks, but with a few minor differences. It’ll be nice if it does a good job of monitoring activity and following up on things on its own. Scheduled tasks run on a pretty strict timeline right now.

But the huge downside is still the billing. Cowork, Code, and Autopilot use usage-based billing instead of being covered by your regular Copilot subscription. Code, which you can loosely think of as Microsoft's version of Claude Code, probably isn't the main attraction for most accountants. Cowork and Autopilot are much more relevant, and you can still run up a pretty big bill using them.

Until that changes, I'm just not that interested in using them. I want to know roughly what I'll spend and what value I'll get, and that still feels like a big unknown. It's nice to see Microsoft making progress, but they're trailing in too many major areas for me to get excited about paying extra to catch up.

Trending News

Mercor reported frontier AI outperforming licensed CPAs on four structured month-end close tasks: I don’t want to be dismissive of this because it’s a nice glimpse into the direction accounting is moving. Agents are already capable of doing the simple things. But at the same time, it’s yet another indicator that most of the world has no idea what 90% of accountants do. I think our profession might have a PR problem.

Anthropic added direct editing in Google Docs, Sheets, and Slides, plus side-by-side editing inside Claude: This is essentially the same as the add-in for Word Documents, but for Google Docs. Great update. However, I’ve found myself using Claude Docs over Google Docs lately.

AuditDashboard launched connectors linking Claude, ChatGPT, and Copilot directly to live audit engagements and PBC data: Ceremonial round of applause for another MCP, yay! In a world where so many vendors are overcomplicating things by building complex AI tools inside their products, I appreciate the approach AuditDashboard is taking to make it really easy to work with data received from clients.

Intuit connected QuickBooks to Meta's Muse for financial analysis, invoicing, and payment reminders: Still not really on board with trusting Meta or their AI. But Muse has caught on lately, as have the other agents. Accountants and firms should be aware that their clients might soon (if not already) be using their own personal AI agents to help run their business. This includes their purchases, managing their books, and even researching tax-related benefits.

Citrin Cooperman acquired Rosetree Solutions to expand its Salesforce and applied AI consulting capabilities: This is another example of firms turning AI implementation into a client service. I think there's a lot of opportunity here, especially for firms that already understand their clients' systems and processes.

OpenAI introduced plugin extensions that let users work in apps inside ChatGPT and @mention app content on desktop: These are essentially apps that can be built to work directly inside the ChatGPT app. It’s a much more seamless way to use AI with the other software you’re using daily. This is my plea to accounting vendors to pay attention to this and start building to fit into this way of working. Feels a lot like the next stage of software.

Elon Musk announced plans for Grok Bot to choose among Claude, Midjourney, Suno, and other models for each task: Might be the biggest update to the third party agents yet. Cheaper models can mean lower-quality results. Grok Bot will likely feel much smarter and more fun to use.

OpenAI opened its Decisions API to all developers in public beta: I covered this last week when access was limited. Now it's available to test in actual automations, like categorizing transactions or routing incoming client documents. I'm interested to see how much faster and cheaper those workflows can get.

Grok Bot added the ability to search, read, and monitor X: Another week, another new feature for Grok Bot. This one is interesting because you have to pay for access to the X API otherwise, but it’s free using Grok Bot.

OpenAI expanded GPT-6 to free ChatGPT users and introduced interactive charts, calculators, and tools inside conversations: This feels a little bit like a miniature ChatGPT Site. The regular chat can better use visuals to illustrate answers. Recently, I had ChatGPT take the dimensions from a room in my house and some measurements of furniture and re-envision the space by creating a visual. This seems perfect for those scenarios.

Put It to Work

Anthropic released Claude Dashboard Artifacts this week. All those IT teams that manage Power BI and Tableau dashboards have clenched their fists and ground their teeth down to nubs in anger.

It takes 5 minutes to create a nice dashboard in Claude, and maybe 1 minute to make changes. I demoed how simple it is to do now.

Weekly Random

We’ve given AI access to millions of data points so it can help us lose to the guy who picked his fantasy team based on jersey colors.

What a time to be alive.

Firecrawl announced a fantasy sports plugin that gives ChatGPT and Claude access to player stats, team stats, fantasy analysis, news, and betting odds. You can use it to research your lineup and ask who you should start, bench, or consider picking up.

I consider myself to be one of the all time greats when it comes to fantasy football. I rarely make incorrect decisions or trades. But this is perfect because now when I do, I can just blame AI for steering me in the wrong direction.

The plugin aggregates projections from several sources, plus injury reports and matchup data, for ChatGPT or Claude to analyze. I’m curious how often it’ll disagree with the consensus. Like does it have intuition or gut feeling based on player injuries and matchups? I doubt it.

And if you follow its advice and lose? You now have a lengthy, well-written explanation of why losing was statistically the correct decision. Very comforting. I’m sure the group chat will be respectful.

Meanwhile, the jersey-colors guy is 4–0.