Hello fellow keepers of numbers,

Happy 4th of July weekend for those in the U.S. And best of luck to those of us with small children and dogs…

This week, Claude Sonnet 5 finally launched after 4.5 months without an update to Sonnet. It’s receiving mixed reviews, but should still be a little better than Sonnet 4.6. Microsoft joined the AI consulting race, including staffing projects from the Big Four talent pool. And speaking of Big Four, Deloitte introduced AI agents into their Omnia audit platform.

Plus, stick around for a demo showing how to build a live dashboard connected to your time and budget data.

Also, some fun news: the newsletter is rebranding. Same newsletter, same takes, same demos, new name and logo. More on that next Friday.

THE LATEST

Anthropic launched Claude Sonnet 5, calling it the most capable Sonnet model it has built and pricing it well below its top-tier Opus model. The company says it's a real step up from the last Sonnet on reasoning, coding, and everyday work tasks.

Early testers said the model finishes multi-step jobs on its own more often, checking its own work along the way instead of stopping halfway through. Anthropic also says it hallucinates and flatters the user less than the previous version.

Sonnet 5 is now the default model on Claude's Free and Pro plans and is available to Max, Team, and Enterprise users, plus Claude Code and the API. API pricing starts at $2 per million input tokens and $10 per million output tokens through August 31, 2026, then rises to $3 and $15, the same rate Sonnet 4.6 charges.

Why it’s important for us:

Sonnet 5 is here. It had actually been about 4.5 months since the last Sonnet release, which is like three decades in AI time. It’s nice to see some updates, but the reaction has been fairly mixed.

Pricing on Sonnet 5 is an interesting wrinkle. And not because it directly costs more than Sonnet 4.6. Several people have tested it on a variety of tasks and found that Sonnet 5 actually costs more per task than Opus 4.8 in some cases to get roughly the same answer.

If that seems confusing, think about this. Let’s take something niche, like R&D tax credits. Hand an R&D project to a new associate, and they'll burn hours just learning what the credit is. Hand it to an experienced R&D tax manager, and they're immediately chugging away at the work. The manager costs much more per hour than the associate, but they’re far more efficient. So, net less cost per project than the associate.

Same idea with the models. The cheap one isn't automatically the cheap answer, so it's worth being deliberate about which one you point at a given job.

Now I'm speculating, but if I read between the lines on what Anthropic is doing, Sonnet 5 looks less like a smart model and more like a better worker. The upgrades they're pointing at are agentic. Better running tasks end-to-end, handling multi-step jobs, holding its own inside a workflow.

The way we talk to these models could shift so you always talk to the smartest one (Fable or Opus), and it runs the show. Instead of doing everything itself, it can split up the job and use subagents running on cheaper models like Sonnet 5 to complete tasks, and then pull the results back together to send the user the output.

Microsoft already shipped something similar as part of Copilot’s Researcher, and it’s the default there now. The mode, called Critique, has GPT write the draft and Claude review it for accuracy and citations before it reaches you. Microsoft says the setup scored about 14% higher on research benchmarks.

I’m obviously just guessing, but the Sonnet 5 capabilities and updates seem to line up with that logic. So maybe we see a shift in how Claude Cowork and Claude Code complete tasks behind the scenes using Sonnet 5 subagents. Still, Sonnet 5 seems a little smarter than Sonnet 4.6. So, even in the meantime, there are benefits to this launch.

Microsoft launched Microsoft Frontier Company, a new business built to help large customers actually put its AI tools to work instead of just buying licenses. The company is backing it with $2.5 billion and 6,000 engineering and industry experts who'll sit inside client organizations to build, deploy, and keep improving AI systems tied to measurable results.

Microsoft named the London Stock Exchange Group, Unilever, Land O'Lakes, and Accenture as early customers. The company also says it'll lean on outside systems integrators, including Accenture, Capgemini, EY, KPMG, and PwC, to help staff the work.

Microsoft's Judson Althoff, who runs the company's commercial business, pushed back on comparisons to the "forward-deployed engineer" teams other AI companies have built, calling Frontier Company "the largest, most capable, outcome-driven engineering organization in the industry." The move follows similar bets from Amazon, which committed $1 billion to its own AI deployment unit days earlier, and joint ventures OpenAI and Anthropic have each launched for enterprise AI work.

Why it’s important for us:

None of these updates should be shocking anymore. We’ve now seen this same news from OpenAI, Anthropic, the Big Four, and, just this week, Amazon. Interestingly, Microsoft is staffing some of this work through the Big Four firms and other consulting giants. Pretty obvious sign that there’s massive demand for this.

I continue to include these updates as important stories because I think AI implementation can be a significant service line for many firms. If we think about how accounting work has crept outward over the years, I don’t think it’s a stretch to assume this should be part of most firms’ offerings.

It started with “keep our books” and “do our tax return.” Then “you’re already in our books, help us understand them.” Then “you’re already helping us understand our books, so tell us how to make more money.” At a certain point, you’re the client’s CFO and tax advisor on top of the standard bookkeeping and tax return services.

Everyone is struggling through the AI adoption curve right now. Businesses will need help using AI with their books and financials, understanding ROI, and making purchasing decisions for new AI products. The biggest companies on the planet aren’t pouring money into this service offering for nothing.

Deloitte unveiled a unified agentic intelligence network inside Omnia, its global audit and assurance platform, bringing new and existing AI agents together to coordinate entire workflows. The rollout reaches nearly 85,000 Audit & Assurance professionals worldwide.

The agents flag potential risk factors, handle preliminary work like data extraction and evidence analysis, draft documentation, and help evaluate regulatory compliance, producing a first pass for staff to review. Deloitte built the system internally, under a framework it says embeds governance and compliance controls throughout.

The release also adds "tutor" agents that give staff on-demand micro-training tied to the work in front of them. Deloitte says the update builds on more than a decade of investment in Omnia, part of an ongoing push to expand AI across its audit and assurance work.

Why it’s important for us:

We continue to see a lot of audit updates, and they all sound pretty much the same. It’s clear that audit was a major focus for the large firms and software vendors over the last 6-12 months. It makes sense because it’s very procedural and there’s a heavy focus on sampling and statistical analysis. Perfect for AI.

Deloitte is the newest top accounting firm to announce agentic AI in their audit software. It’s hard to know exactly what these updates look like if you’re not working for the specific firms. But I suspect a lot of this is marketing material more so than game-changing new tech.

Still, the audit space is really interesting right now. Audit work has, for the most part, been eaten up by the largest firms. Most smaller firms don’t even have an audit practice. But the barrier to entry is lowering significantly.

AI is simplifying the procedural grind of an audit. A year from now (or even today for some tech-forward firms), it might take far less time to complete an audit, and firms might need fewer audit experts to feel comfortable taking on new audit work. The cost of running an audit could drop significantly, which will make smaller firms much more competitive in acquiring the work.

I also think the quality of audits is going to improve drastically. At scale, especially on public company audits, we’re getting close to being able to test nearly 100% of the data instead of sampling. Obviously, this assumes you can access 100% of the data, which is certainly still a problem that needs solving.

TRENDING NEWS

Anthropic redeployed Claude Fable 5 globally after the US lifted the export controls that had knocked it offline: Fable 5 is back after a couple weeks in the void, now with a safety classifier Anthropic says blocks the jailbreak over 99% of the time. Cruel joke to hand it back right before a long holiday weekend. Shame on them.

Google brought Gemini Spark to macOS in beta, adding custom MCP support and connected apps like Dropbox and Canva: This is Google's answer to Claude Cowork. Looks like it'll be fairly limited to start, but as someone living in Google Workspace, I'm still excited to test it.

Microsoft made Claude Opus 4.8 and Haiku 4.5 generally available in Foundry on Azure: Boring on the surface, useful in practice. If your firm's data already lives in Azure and M365, you can now have Claude work on it directly, choosing it over Copilot or OpenAI as the model doing the work. Microsoft's been leaning into Claude hard over the last year.

OpenAI offered the US government a 5% stake worth about $42.6 billion and wants Anthropic, Google, and Meta to match it: Not sure what to make of this one. On the surface, it reads like a political play, cozying up to the administration for favorable treatment. But it could also lead to an ROI that the government redistributes to people AI displaces, which could obviously be beneficial.

Anthropic expanded Claude Code Artifacts to Pro and Max plans, a month after launching them for Team and Enterprise: Artifacts have easily been my favorite Claude feature since Cowork launched. People are starting to talk about them, but they're still wildly slept on right now.

Notion added an HTML block that lets you build interactive pages and ask AI to turn content into explainers, prototypes, or diagrams: Notion keeps staying at the front of the pack on baking AI into its software. Great update that turns Notion into much more of a visual tool, handy for making pages interactive and spinning up charts, dashboards, and other visuals.

AppZen reported that AI-generated fakes are now the top method for faking receipts in expense fraud, up from basically zero a year ago to about 71% of flagged receipts: It's getting harder to spot AI fakes by eye, and receipts look like an even tougher case than most. I think we'll end up leaning on the tech that AI companies are building to tag AI-generated images and video to catch this.

PUT IT TO WORK

I made the claim above that Claude Artifacts are undervalued right now. Probably my favorite release since Cowork. Because of my low self-esteem, I felt the need to prove it in this week’s demo.

In this demo, I create a Claue Artifact that’s intended to be used as a dashboard to track budget vs actual for projects. And as a bonus, I had Claude create a capacity planning tab as well.

It needs a lot more love before I’d be happy using it internally, but it took 15 minutes and probably replaced some really ugly and annoying Power BI dashboards.

Claude Artifact for Budget vs Actual and Capacity Tracking - Watch Video

WEEKLY RANDOM

Anthropic launched Claude Science this week. You’ll never guess, but it’s an app built for scientists.

It pulls the whole mess of research, dozens of databases, code, clusters, and the published literature into one place so a researcher can actually do the work instead of using 15 different tools.

Scientists have been in the beta for a few months, and some of what they've already pulled off is awesome.

I've said this before, but I’m most excited about AI impacts on science and health. So much of the AI conversation right now is negative. AI is taking jobs, AI is slop, AI is ruining this or that. But this is the other side of it.

And it's not just Anthropic. OpenAI and Google are making progress here too, as are many other companies dedicated directly to AI for science and health.

Claude Science won't cure anything on its own. But this is the kind of AI news I wish got more attention because it's going to change lives.