Hello fellow keepers of numbers,

A week ago, everyone in the AI space agreed that we should slow down. So, naturally, we got new models from OpenAI, Anthropic, and SpaceXAI this week. Nice.

This was a good week for the usage in our Claude and ChatGPT accounts. Opus 5.5 was released with some major improvements to the quality and efficiency. It’s much more efficient than Opus 5 and 20% cheaper. GPT-6 Sol and Luna are a very similar story. Better models that are also cheaper. Yay, good for our wallets.

Lastly, we got a very interesting announcement from Filed and Crimson Tree Software. They partnered to found the Open Tax Technology Alliance and open-sourced a 1040 calculation engine and a draft standard for exchanging K-1 and K-3 data. Really smart people making it a little easier for the accounting profession to use AI to review work and automate tasks.

Plus, stick around for a video where I walk through all the essential AI terminology to get you up to speed on everything you need to know for AI tools like Claude and ChatGPT.

The Latest

Anthropic launched Claude Opus 5.5, the first model in its Claude 5.5 family. The company says it performs at the level of Claude Fable 5.1 on most work while improving on Opus 5 in knowledge work, communication, speed, and cost.

In one internal test, Anthropic asked Opus 5.5 and Opus 5 to analyze a proposed merger, build a financial model in Excel, and prepare an executive presentation. Anthropic says Opus 5.5 produced a more thorough model and clearer presentation, finishing in 63 minutes instead of 93 at half the cost.

Anthropic says typical workloads cost 40% less than Opus 5 and output is more than 30% faster. API pricing for Opus 5.5 is $4 per million input tokens and $20 per million output tokens. Both rates are 20% lower than Opus 5. Opus 5.5 is available now across Claude’s products, its API, and major cloud platforms. Anthropic also increased five-hour usage limits for Pro, Max, Team, and seat-based Enterprise subscribers.

Anthropic’s Mike Krieger said Claude Sonnet 5.5 and Haiku 5.5 will follow in the coming weeks. He said both will carry many of the performance, efficiency, and safety improvements introduced with Opus 5.5. Anthropic has not announced exact release dates or pricing for either model.

Why it’s important for us:

Opus 5 was frustrating to use. It ignored instructions, struggled with connectors, and sometimes failed at running the same skills I’ve used for a year now. It'd just run some skills based on vibes instead of following my hard-written instructions that Claude entirely wrote. Even its responses were harder to follow. Anthropic has acknowledged the complaints about Opus 5, and it was nice to hear that from them. I know I wasn’t the only one having trouble with it.

After a couple of days with Opus 5.5, my skills are running better. The communication is way better, too. It’s more concise and easier to follow, but it also just feels more like the old Claude, which was more enjoyable to work with. That’s hard to measure, but I’ve noticed it almost immediately.

I’m also really encouraged by the efficiency improvements. Anthropic says Opus 5.5 performs as well, if not better, than Fable 5.1 on some tasks while costing less than the previous Opus. This means it'll consume less usage as you do the same work. And Anthropic has raised the 5-hour usage limits as well. Getting both improvements at once is going to make me feel invincible.

And we still have Sonnet 5.5 and Haiku 5.5 coming. Some had speculated Haiku would be decommissioned, but Anthropic must've seen some value to having the really cheap model. I'm really excited for Sonnet 5.5 though since it's most people's daily driver. Sonnet 5 has some of the same issues Opus 5 had, though not quite as bad. Hopefully we'll see similar improvements for it.

OpenAI launched GPT-6 Sol and Luna, successors to GPT-5.6 Sol and Luna. The company reports improvements in business workflows, factual accuracy, and computer use, along with lower API prices.

To test factual accuracy, OpenAI used past ChatGPT conversations in which users had flagged incorrect answers. In those cases, the company says GPT-6 Sol made about half as many mistakes as GPT-5.6 Sol.

API pricing for GPT-6 Sol is $2 per million input tokens and $10 per million output tokens, half the promotional prices for GPT-5.6 Sol. GPT-6 Luna costs $0.10 per million input tokens and $0.50 per million output tokens, down from $0.20 and $1.20 for GPT-5.6 Luna.

Both models launched in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise, and Edu users, and are available through the API. Free and Go users can access Luna in the desktop app.

Why it’s important for us:

Before I get to the models, I have to call out the interactive graphic in OpenAI’s article. It's probably the coolest graphic I’ve ever seen. Drag the moon in front of the sun and you can create a solar eclipse. Click the link in the news story above to check it out. That’s not the point of the announcement obviously, but I had to mention it.

Both Sol and Luna are major upgrades. Luna is the smallest model, and so far I’ve found it pretty underwhelming unless I'm using it for really quick, simple questions. I use Sol for most of my work in ChatGPT.

I’ve been testing Sol over the last few days, and it’s been really great. My only nitpick is that it seems a little more verbose than GPT-5.6 Sol and it doesn't quite write as well (in my opinion). I just prefer shorter answers, and this version keeps giving me longer ones. It’s not the end of the world.

GPT-6 Astra is awesome, and Sol is getting close to that level while being way cheaper. Even after Anthropic lowered the price of Claude Opus 5.5, Sol’s prices are still half the cost. More savings on our weekly usage. The performance feels a little worse than Opus 5.5 so far though.

Luna is as cheap as can be. If you’re finding that it works well for your simpler tasks, you can use it seemingly endlessly, depending on your plan. I just haven’t found many reasons to reach for it myself.

Filed and Crimson Tree Software founded the Open Tax Technology Alliance and released two projects: OpenTax, a free open source Form 1040 calculation engine, and Open Tax Document, a public draft standard for exchanging Schedule K-1 and K-3 data.

OpenTax calculates 2025 federal Form 1040 returns using inputs that include W-2s, 1099s, K-1s, and Schedules A, C, and E. It runs locally on macOS, Linux, and Windows without an account or cloud connection. It also checks returns against IRS Modernized e-File rules and exports the required XML. Its application to the IRS e-file system is still pending, so the alliance has not launched the proposed e-filing route.

Open Tax Document defines a structured format for K-1 and K-3 information, including supporting footnotes, so software can exchange that data without relying on PDFs. Both projects are publicly available. OpenTax uses an AGPL v3 license, with a commercial license available for organizations that cannot use AGPL.

Why it’s important for us:

I love this so much. I’ve spent the last year complaining about the old guard of tax software. They charge us a fortune for clunky products, then put the tax calculations behind another annual license. As Leroy from Filed pointed out in the article, the calculation itself was solved decades ago. Why are we still paying to access it through software that hasn't changed in decades?

Filed and Crimson Tree’s answer is to release a free, open-source Form 1040 calculation engine. Others in the space have done this already, but it's always nice to have more variety. What excites me is that AI and new investment in accounting are bringing more smart people into our industry who seem willing to build what’s better for the profession.

I’m especially interested in what this could do for return review. Right now, the prep and review all live in the same shitty tax software, and you hope you're able to catch any errors. Sometimes diagnostics help with that. But what if we ran the return through a separate calculation engine and had an AI agent compare the results, research any differences, and flag the ones a manager or partner should look at? That workflow still needs to be built and tested, but a free engine makes it a real possibility.

The K-1 and K-3 standard interests me for a similar reason. Firms pay companies to pull information out of those forms and move it into other software. A common format for exchanging that information could make a lot of the PDF shuffling unnecessary. It’s a draft, and I suspect there’s plenty more to build, but I love the idea behind it.

For the skeptics out there, I think your question is: Why would Filed and Crimson Tree give this stuff away? I think the optimistic point of view is that they want to do something good for the profession. Something that should've been done a long time ago. But the other side is that they'll get a lot of attention and publicity for being the companies willing to take on the major players in the space. I'm sure it helps their businesses. And even though this is open sourced, there's still work to be done to set this up properly and implement it into firm workflows. But I'm really glad people out there are doing this.

Trending News

OpenAI scheduled custom GPTs for retirement on December 11 as it moves users toward plugins: Anyone who still uses GPTs needs to make note. Building a plugin isn't very difficult, but can be a bit intimidating at first.

Ramp launched AI-powered accounts receivable that drafts invoices, prepares collection emails, and matches payments: They're on a ramp-age right now. If you laughed, you're reading the right newsletter. If you didn't, I don't blame you. But Ramp continues to ship really great AI features built into their platform. This could obviously be a huge ROI for companies too if it can improve collections.

SpaceXAI released Grok 4.7 with improvements to longer knowledge-work tasks, documents, and presentations: This update was a bit underwhelming considering how much Elon hyped this model. Elon overpromising something that came both later and sadder than we expected? Never would've guessed. No, but in reality, Grok is still really good. It does seem to be a bit of an improvement, and the SpaceXAI team continues their great work, thanks to the Cursor team they acquired.

Accrual launched Arc, an agent platform firms can teach their own workflows and use across connected systems: This sounds similar to Claude Cowork with prebuilt "agents" (probably just skills/plugins with what Accrual knows). Also, Accrual published a blog that seems to be written by Arc. You can read it here. We're definitely anthropomorphizing AI now.

Qount launched an MCP connector that brings practice management data into Claude, ChatGPT, and other AI tools: Obligatory round of applause for another accounting software releasing an MCP.

Financial Cents launched three free AI agents that rename, validate, and route client files: They might be slightly behind when it comes to AI and MCPs, but this is a good update. These are high-volume minor annoyances for firms. Can most firms do this themselves with their AI tool of choice? Sure. But still nice, and it's free.

ChatGPT added support for connecting multiple accounts to the same plugin: This makes me so happy. I suspect other people aren't as insane as me, and they probably have only one main account for all their software. But I have several different accounts for Google, Microsoft, Notion, and more. Now ChatGPT supports adding multiple accounts, and the AI model can switch between them without me logging out and back in manually.

FloQast introduced agents built from completed closes and added AI review of journal entries: This is a good update, but just another example that every CAS software better be doing something like this if they haven't already. Otherwise, they're going to be left behind quickly.

Anthropic expanded Claude for Small Business with 27 new integrations, including Xero, Gusto, and Stripe: One of my favorite plugins to use to "steal" skills or skill ideas that Anthropic has open-sourced for us. It continues to grow.

Grok Bot added native Google Docs, Sheets, and Slides integrations, plus support for email attachments: The Grok Bot updates keep coming. These aren't game changing, but definitely nice quality of life improvements for those using it for work.

Claude Code added support for AGENTS.md instructions when a folder has no CLAUDE.md: Finally. Every other AI model on the planet reads an AGENTS.md, but Claude still hadn't officially supported it. This makes it much easier to switch between Claude, ChatGPT, Cursor, etc. when you're working with your projects.

PwC explained how its auditors use AI while professionals direct, supervise, and review the work: Most of these articles are just marketing material, but maybe there are a few interesting nuggets for the auditors out there.

Put It to Work

AI is weird, and the terminology is odd and confusing. The words are in English, but most have probably no meaning unless you follow AI very closely. So I did my best to break down 20 of the most essential AI terms.

Weekly Random

Claude found a previously unknown biological system this week.

Anthropic’s new biology lab is testing whether Claude can find promising clues in the enormous amount of DNA data scientists have collected. For this project, researchers pointed Claude toward a type of enzyme they study and asked it to look for anything unusual.

About 950 Claude agents searched for 21 hours. One spotted a repeating DNA pattern in viruses that infect bacteria. It appeared near an enzyme researchers already knew about, but nobody had recognized the larger system they formed together.

The pattern looks a bit like one found in CRISPR. That’s interesting because CRISPR also started with scientists noticing strange repeats in DNA, long before they figured out how to use it to edit genes. Nobody knows yet what this new system does.

I wrote a few weeks ago about Claude helping scientists design an early building block for new medicines. In that case, researchers knew what they wanted to make. This time, Claude found something they didn’t know to look for. I think that’s the most exciting part.

What I find fascinating is the division of labor. Claude found a lead and helped guide what to investigate next. Scientists ran the experiments and confirmed the pattern produces small RNA molecules, with Claude helping interpret the results. AI helping steer the investigation while humans do the hands-on work feels like a very new way to do science.

I have no idea whether this discovery will ever help create a treatment. But I keep thinking about how many other things might be sitting in those databases, waiting for someone to notice them. Having AI look for the clues and scientists run the experiments seems like a pretty damn good way to find out.