Hello fellow keepers of numbers,

Welcome to the end of the world. It’s called Fable. You know, those things that teach moral lessons? How ironic.

Just kidding, kind of. Anthropic released Claude Fable 5 this week, which is a relative of Mythos. It’s a beast. It’s expensive. And it’s really damn good.

Karbon also released their embedded AI coworker, Kai. They deserve a round of applause for offering both an AI coworker inside their app and also an MCP in case you want to use your data elsewhere. Also, Magnetic launched a tax prep agent for 1040s that’ll only cost you if it works.

Plus, I demo how you can use Cowork to autonomously complete tasks you assign it throughout the week.

THE LATEST

Anthropic released Claude Fable 5, making a Mythos-class model generally available for the first time. Fable 5 shares its underlying architecture with Claude Mythos 5, a version currently restricted to authorized cybersecurity users. Three classifier-based safeguards covering cybersecurity, biology and chemistry, and model distillation are applied to the general release. Anthropic says more than 95% of Fable sessions trigger none of them.

Fable 5 reaches state-of-the-art performance across software engineering, knowledge work, vision, and scientific research. Anthropic says the model performs three times better than Opus 4.8 on long-context tasks that use file-based memory.

Fable 5 is priced at $10 per million input tokens and $50 per million output tokens, roughly twice the cost of Opus 4.8. It is currently included in Pro, Max, Team, and seat-based Enterprise plans at no extra cost before shifting to usage credits. All Mythos-class traffic is subject to a new 30-day data retention policy.

Why it’s important for us:

There’s some good, and there’s a lot of bad.

The good:

This is the smartest model ever released. It's not close based on the benchmarks. It’s also been really impressive in my own testing and what I’ve heard from others.

Fable 5 shines at knowledge work. It’s great at data analysis after pulling together large amounts of data from a lot of sources. If you’re an advanced user, it’s also really good as an orchestrator. It can spin up subagents that use smaller models like Sonnet so it doesn’t chew through usage, and then Fable compiles all the data from the subagents. I’ve had quite a bit of success with these workflows the last few days.

However, for most, Fable 5 is overkill. For most of the day-to-day work we're doing, Sonnet and Opus are still the right call. Fable 5 will chew through your entire usage in about 25 seconds. And in 10 days (June 22nd), it’ll no longer be included in your usage at all. It’ll be a separate cost that charges you each time you use it. 

I suspect most people will just go back to Opus at that point. Fable 5 will likely be included in your usage again in the future, but it probably won’t be until some newer models ship or Fable competitors arrive from OpenAI or others. But it’s still a big deal that we have a whole separate tier of models above Opus now.

The bad:

(1) If you remember from early announcements, Mythos-class models are incredibly good at finding weaknesses in software. Humans can use the model to exploit those weaknesses. Fable 5 could theoretically give them the ability to do this (more on this in a second).

There’s a lot we can’t control on the cyber front. We’re relying on software vendors to a large extent. What we can do is use proper data hygiene and protect our passwords and API keys.

(2) This is the big one. Anthropic has killed zero data retention (ZDR) for these models. If you’re on Claude Enterprise and using ZDR or have ZDR enabled through the API, you won’t have access to Fable 5. If you want it, you have to enable retention.

Every request Fable gets is reviewed and classified by Anthropic’s systems to determine if it hits a sensitive area like biology, chemistry, or cybersecurity. If so, it re-routes that request to Opus.

Because they’re classifying all messages that are sent to Fable, they have to retain the data. They’ve outlined their policy and reasons here.

Anthropic has now set a precedent. They’re going to decide what happens to your data and what are “proper” uses of the tool you pay to use. This is going to be a very important area to watch as others release similarly powerful models.

Karbon launched Kai, an AI coworker built into its practice management platform, drawing on the client history, workflows, and communications already stored there to help practitioners scope services, manage repetitive tasks, surface client insights, and plan their day.

Alongside Kai, Karbon announced three other AI capabilities. An AI notetaker joins client meetings, captures transcripts, and delivers structured analysis directly in Karbon. Agentic period close checks automatically flag transactions requiring additional review. AI email triage categorizes messages, routes them, and surfaces critical communications.

Karbon also published a public MCP server, an open-standard connector that lets firms connect Karbon data to external AI tools directly.

Most features are live now. Kai is currently in early access.

Why it’s important for us:

I’m a fan of the approach Karbon has taken to AI and their software. Kai is a great example of why.

The concept of Kai isn’t new. It’s a tool very similar to Claude Cowork sitting inside an existing software that can help you complete tasks in that software. Canopy recently did this with Canopy Coworker.

The big difference here is that Karbon is coming at AI from all angles. If you want to use AI inside of Karbon, you have Kai. If you want to pull Karbon’s data into something like Claude Cowork or Codex, you can use their MCP. Use it where it makes sense for you. And I think this is important for firms because it’s going to lead to higher adoption of AI. Some might want to work inside Karbon, while others want to stay in Claude Cowork or Codex. Both work.

Karbon has methodically built a large list of integrations over the last several years that pull other data into your practice management system. It always seemed useful, but now they’re really reaping the benefits of those efforts. AI has access to a lot of data for your firm across many of your different apps and vendors.

I also want to look at the bigger picture for a second. The ambition of practice management vendors has always been to be your home base. Now, we’re seeing these same vendors transition to your operating system. Beyond being your home base, they also want to be the place where you do the work. Completely different job. Just another signal that the market is trending towards one major operating system for all your work (and not just unique to accounting). Maybe it’s Karbon. Maybe it’s Cowork. Maybe it’s Codex.

Magnetic launched an AI agent that prepares 1040 individual returns end to end, backed by an accuracy guarantee: if a Magnetic-prepared return contains an error that changes the final tax liability by $10 or more, the return is free. During Spring 2026, only 3% of customers claimed the guarantee.

Magnetic uses an orchestrator agent with specialist sub-agents, one per schedule, to handle document parsing, contextual analysis, and return completion across UltraTax, CCH Axcess, Lacerte, Drake, and ProConnect. The system supports Schedules A, B, C, D, and E, along with K-1 income, multistate returns, and basis tracking.

The workflow starts with the firm exporting a starting file from its tax software. Magnetic parses the client documents, flags anomalies, builds a bookmarked PDF, and completes the return. A U.S.-based EA or CPA then reviews before delivery. Magnetic says all processing happens on U.S. infrastructure, it does not train on customer data, and the service is IRC Section 7216 compliant.

Magnetic's three-day delivery guarantee covers 1040 returns only. No business or entity returns are supported. The Spring 2026 average turnaround was approximately 2.5 days. Pricing was not disclosed at launch.

Why it’s important for us:

Magnetic is going with outcome-based pricing. It’s a hot pricing strategy these days. Digits recently moved to outcome-based pricing as well.

There’s a lot of skepticism in the market around the value AI tools deliver. So the vendors who are confident in their product want to put their money where their mouth is. Does this mean we should trust them? I’m not sure yet. But it definitely makes it more interesting to consider.

Magnetic stands out because they’re taking a different approach from other AI tax prep vendors. They’re not handing you a software to do your returns faster. They’re taking the return prep out of your hands almost entirely. Their pitch is basically: we built this, we think it's great, so send us the data, we'll prep it with our AI tool, we'll review it, and you get back a finished return that just needs your sign-off.

This won’t work for every firm. I think it makes the most sense for a few different types of firms:

(1) If you already outsource tax prep, this is the same model you’re used to. If it’s faster, cheaper, or more accurate (or all the above), it makes sense to explore.

(2) If you’re a new firm or are newly evaluating tax prep as a service, this could work since you don’t have built-up infrastructure that makes change management difficult.

(3) If you’re a solopreneur or a small firm with too much work (or if you’re turning down work), this could be a good strategy to help you scale.

Magnetic likely doesn’t fit established firms that already have processes built around tax prep internally. For those firms, you’d probably be better off with one of the other AI tax prep vendors.

TRENDING NEWS

Digits launched Agentic Close, a system that turns month-end close into a continuous process, booking, reconciling, and reviewing activity as it hits the ledger: Second month-end close launch in a couple weeks, right after Ramp Stack. We definitely have a theme for the last month or two.

Fieldguide launched Field Orchestrator, an AI agent that runs substantive audit testing end to end, from sample selection through workpapers, with human review built in: AI should be great as an auditor in theory because a lot of it is very procedural. How much value is something like this long-term vs using a Cowork or Codex to assist with samples and procedures? Not sure yet.

OpenAI plans to merge Codex and ChatGPT into a single superapp within weeks: I can’t remember the last time I’ve used chat in either ChatGPT or Claude. Claude Code and Codex are my daily drivers, even for the quick questions. Seems like OpenAI sees that broadly in the market too, which is why they’re eyeing this shift. Hopefully the merge is seamless.

Gusto launched Cofounder, an AI teammate for small business owners with 20+ prebuilt automations across payroll, onboarding, and compliance: This is interesting, but I’m not sure how much value this provides vs using the Gusto connector in Cowork or Codex.

OpenAI rolled out saved Codex rate-limit resets, letting paid users bank a reset and use it later: This is really nice, actually. It lets you wait to reset your usage until a time when it gives you maximum effect for your 5-hour usage window.

Anthropic CEO Dario Amodei published an essay calling for mandatory safety testing of frontier models, with regulators empowered to block unsafe releases, including Anthropic's own: Amodei has had a few viral blogs in the last few years. He’s surprisingly candid about progress, his vision, and even his uncertainty. He’s sounding the proverbial alarm bells to alert regulators and decision-makers to move more quickly.

SpaceX priced the largest IPO ever, raising roughly $75 billion at a ~$1.75 trillion valuation, with AI as the thesis: The IPO was priced at $135/share. It opened at $150/share and is currently (as of this writing) $161/share. That’s about a $2.11T market cap.

Apple unveiled a rebuilt Siri at WWDC that reads on-screen content and takes actions across your apps: Apple promised their “Apple Intelligence” at WWDC last year and it never came, so we have a right to be skeptical. This time, they demoed working AI features, so it seems more real. This has the potential to be incredibly impactful for both personal and business.

PUT IT TO WORK

I set up Cowork to do my work for me, and I’m going to go off on vacation for the next few months. Not really, I’m just going to lie on my couch all day and watch TV.

I’ve demoed how you can set up Cowork projects and create scheduled tasks linked with a task tracker to complete work for you autonomously. Cowork is great, but sitting inside tasks waiting for things to finish is so old school now. Just log work for it in a task tracker and then review it whenever you have time. Maybe later that day, maybe the next day, maybe next week. In the meantime, just keep logging new tasks for it.

How to get Cowork to do your work autonomously - Watch Video

WEEKLY RANDOM

Indulge me for a second about sports. If you hate them, this is your chance to stop reading.

This is a particularly fun time of year for sports fans because we have the NBA finals, the Stanley Cup, and now the World Cup all at once. Plus, we’re in the midst of baseball season. The dead time of year is in the not too distant future, so we’ll soon be depressed. Let’s enjoy our moment for a second.

The Knicks completed the greatest comeback in NBA finals history, and one of the all-time great comebacks in sports on Wednesday night. Shoutout to myself and all the others who chose not to turn that game off at halftime. I’m not sure if my own perseverance in watching the game or the Knicks’ coming back after being down 29 was more impressive that night.

The Stanley Cup has been absolutely wild. We’ve had 2 overtime games, including one where the Hurricanes scored 4 goals straight in the 3rd period to force OT.

And the World Cup kicked off yesterday. The US plays Paraguay tonight to kick off the home country’s first match. I always love seeing an entire country get behind a team because the passion is hard to match.

So congrats to us sports fans, and sorry to all the spouses out there (like mine) who have to deal with the TV constantly being on something they don’t care to watch.