Today we are discussing the latest product from Anthropic. It is somehow more controversial than just about anything I've seen from them. And I think that the reasons why for that, or at least the multiple different reasons why for that, can tell us a lot about kind of where we are as a field right now in the larger space of artificial intelligence. Welcome back to the AI Daily Brief. In 2026, the one thing that's clear to everyone is that things are moving very fast. Even for an industry where it already felt like things were going quickly, we've ratcheted it up another notch. As part of that, everyone is grappling with a series of different issues. Everything from the very positive, how do I take advantage of all these new superpowers that I've been given? To the exciting, but it's a challenge kind of questions like, how do we redesign our organization around these new capabilities? To the much more existential questions of what does it mean that the work that I've always done is no longer the work that I will be doing. In many ways, it feels to me like all of those debates came home to roost around a single product this week, which was Anthropic's new code review feature.
Now, this is not a particularly complicated product to explain. Claude writes, "When a PR opens, Claude dispatches a team of agents to hunt for bugs." Code review is a key part of the development life cycle, so it stands to reason that AI would be trying to add new efficiency to it. And certainly, Anthropic is not the only company thinking in these directions. Cognition recently released Devin Review, which they call a reimagined interface for understanding complex PRs. In their announcement tweet, they wrote, "Code review tools today don't actually make it easier to read code. Devin Review builds your comprehension and helps you stop slobbing." Now they go through a whole bunch of ways in which the product is different, and it got pretty good response. A thousand people bookmarked that tweet, and three quarters of a million people viewed it. That is of course nothing compared to the nearly 14 million who viewed the Claude post, which speaks not only to the relative size of Anthropic but to the controversy surrounding this new product. So what actually was controversial? On the surface of it, this seems like it would be highly value additive. While they are biased and incentivized to say so, certainly it seems like all the folks inside Anthropic who are using it have had really positive experiences with it.
Alex Albert, who does Claude and Dev Relations, says this has been a game changer for our internal engineering research teams. Rare to see a product get this much praise from some of the top engineers I know. Boris Cherny, the creator of Claude Code, points out we built it for ourselves first. Code output per Anthropic engineer is up 200% this year, and reviews were the bottleneck. Personally, I've been using it for a few weeks and have found it catches many real bugs that I would not have noticed otherwise. Jared Sumner writes, "Been using this in Bun's repo." Bun, JavaScript being a company that joined Anthropic recently. Jared continues, "This in my opinion is the best product in the code review category today. It regularly catches extremely subtle bugs and rarely makes mistakes." Cloud Codes writes, "Code review is so, so good. One of those things I can't remember how I lived without." What's more, the discussion of code review and the inevitable changes to it is something that the larger agentic engineering community has been talking about recently, irrespective of this Claude release.
Of late, Swix wrote, "This is the final boss of agentic engineering, killing the code review." At this point, multiple people are already weighing how to remove the human code review bottleneck from agents becoming fully productive. I'm not personally there yet, but I tend to be three to six months behind these people, and yeah, it's definitely coming. Now, he points to a guest essay shared on Leighton Space by entrepreneur Ankit Jain called "How to Kill the Code Review." The subheader, which encapsulates the thesis pretty clearly, is: Human-written code died in 2025. Code reviews will die in 2026. I won't read the whole thing, but a couple of excerpts. Humans already couldn't keep up with code reviews when humans wrote code at human speed. Every engineering org I've talked to has the same dirty secret: PRs sitting for days, rubber-stamp approvals, and reviewers skimming 500-line diffs because they have their own work to do. We tell ourselves it is a quality gate, but teams have shipped without line-by-line review for decades. Code review wasn't even ubiquitous until around 2012 to 2014. One veteran engineer told me, "There just aren't enough of us around to remember." And even with reviews, things break. We have learned to build systems that handle failure because we accept that review alone wasn't enough. This shows in terms of feature flags, rollouts, and instant rollbacks.
The next section in the core thrust of Ankit's argument is called "We Have to Give Up on Reading All the Code." He continues, "Teams with high AI adoption complete 21% more tasks and merge 98% more pull requests, but PR review time increases 91%, based on data from over 10,000 developers across 1,255 teams. Two things are scaling exponentially: the number of changes and the size of changes. We cannot consume this much code. On top of that, developers keep saying that AI-generated code requires more effort than reviewing code written by their colleagues. Teams produce more code and spend more time reviewing it. There is no way we win this fight with manual code reviews. Code review is a historical approval gate that no longer matches the shape of the work." Now, Boris wrote something about this as well. His more broadly themed piece from February of this year was called "The Software Development Life Cycle Is Dead." Boris writes, "AI agents didn't make the SDLC faster. They killed it. I keep hearing people talk about AI as a 10x developer tool. That framing is wrong. It assumes the workflow stays the same and the speed goes up. That's not what's happening. The entire life cycle, the one we've built careers around, the one that spawned a multi-billion-dollar tooling industry, is collapsing in on itself, and most people haven't noticed yet." Boris argues that the software development life cycle as we learned it is a relic. He writes, "Here is the classic software development life cycle most of us were taught," and apologies for those of you who are just listening, "but basically, it's a circular chart that goes from requirements to system design to implementation to testing to code review to deployment to monitoring, and then back to requirements and through the system again." Boris writes, "Every stage has its own tools, its own rituals, its own cottage industry: Jira for requirements, Figma for design, VS Code for implementation, Jest for testing, GitHub for code review, AWS for deployment, Datadog for monitoring. Each step is discrete, sequential, handoffs everywhere." Now, here's what actually happens when an engineer works with a coding agent. In this chart, there is one starting point, which is intent, which moves to the agent, and then the agent works in a circular fashion through code plus test plus deployment to the question of does it work? If the answer is no, it's back to the agent for more code, tests, and deployment. Back to the question of does it work. And then as soon as the answer to does it work is yes, the code gets shipped. Boris's point is this quote: "The stages collapsed. They didn't get faster. They merged. The agent doesn't know what step it's on because there are no steps. There's just intent, context, and iteration." Boris is talking about the entire development process. But to refocus it back to code review, which is the subject of this particular product, his section on code review is called "Give It Up." Boris writes, "The pull request flow needs to go. I was never a fan, but now it's just a relic of the past. I know that's uncomfortable. Code review is sacred. It's how you catch bugs, share knowledge, maintain standards. It's also an identity thing. We're engineers, and reviewing code is what engineers do. But clinging to the PR workflow in an agent-driven world isn't rigor. It's an identity crisis. Think about it. An agent generates 500 PRs a day. Your team can review maybe 10. The review queue backs up. This isn't a bottleneck worth optimizing. It's a fake bottleneck, one that only exists because we're forcing a human ritual onto a machine workflow."
All right, so the point here that I'm trying to make is that clearly there is something in the air and big questions and perhaps an inevitable change coming to the way that we think about code review. And yet still, I was genuinely surprised to see how much antipathy there was towards this code review announcement. There were a few reasons for that. One has to do with a sort of "who's going to watch the watchers" idea. Professor Bo Wang writes, "Did Claude Code write Claude Code Review?" Next question: Can Claude Code review Claude Code's code and make it better and even create a better Claude Code Review? Now, he's a little bit tongue-in-cheek, but the idea of whether the code review is likely to bring the same biases to the review that might have created the mistakes in the code in the first place if people wrote their code with Claude Code is, I think, maybe a more practical question that a lot of folks have. The much bigger part of the response came around cost. The big thing that really caught people's attention was around the pricing. In the pricing section of the Claude Code Review docs, it says code review is built based on token usage. Reviews average $15 to $25, scaling with PR size, codebase complexity, and how many issues require verification. And boy, were people shocked at this. Varilon writes, "The Claude Code Max $200-a-month plan is literally infinite tokens. You can just write the one prompt to do a PR review locally, save it as a skill, and you get unlimited reviews. $15 to $25 per review is nuts." Dagster Labs' Nick Schock writes, "$15 to $25 USD per review. My lord." Alex Kaplan says, "$20 for a PR?" head-blown emoji, exclamation-question-mark emoji. "DevinReview.com is free." So one part of this I think is just a sticker shock argument. If you've got most developers used to paying in the tens of dollars for coding tools and seeing review-type features bundled into a broader plan, then this amount obviously seems much larger. What's more, people are immediately doing the scale math. If a team opens up lots of PRs, $25 per review sounds like it could explode very quickly into hundreds of thousands per developer per month. Now, it doesn't really matter that Anthropic is explicitly targeting a deeper review experience using multiple specialized agents, i.e., probably not using it for every single time you have to review anything. But still, people are just extrapolating out from that number and coming up with some very big numbers on the other side. Another piece of this, though, is that I think it shows some chinks in the Anthropic and Opus armor right now. For a very long time, Anthropic was the only game in town when it came to coding. This has been well documented on this show to the point where we don't really need to discuss it. However, ever since the release of GPT-5, OpenAI has been explicitly attempting to close that gap and even get out ahead. And increasingly, there is some evidence that that effort has been successful. Weser writes, "I really don't understand why you would pay $25 for Claude to review a single PR when Opus 4.6 isn't even the best model for deep code review. GPT-5.4 is the only model I trust for reviews right now." Shopify product builder Gil writes, "Imagine spending $15 to $25 on code review and you still have daily downtime and buggy releases. I'd be more confident in this feature if their production quality was higher." Fairman's Tibo writes, "In all our benchmarks, Claude reviews are always just the worst. But don't you worry, now you can pay between $15 and $25 per freaking PR and you'll have good reviews. Are you kidding me?" And even some of the first people who are testing it aren't necessarily coming away all that impressed. Daniel Sand tested Claude Code Review and said, "I'm always the first to get excited when Claude ships something new, but in this case, enabling code review is just not worth it." Now, none of this is to say that there aren't people who are taking the other side of this argument. Lindy founder Flo writes, "People's comments on the $15 to $25 per PR price tag remind me of Michael Bloomberg's answer to people bing at the $2,700 per month cost of the Bloomberg terminal: If you can't make $2,700 a month with our product, you've got bigger problems to deal with." OpenCode's Ree Sullivan writes, "A $15 to $25 PR review bot that catches an incident that would have cost the company $5 million in breached SLAs and reputation is a no-brainer." I think maybe ultimately the even more interesting dimension of the cost part of the conversation actually has to do with the implications for where things are going. I think increasingly, and especially as AI coding weaves itself deeper into how we do work, cost profiles which were somewhat ignorable before become not ignorable anymore. Another way of saying it is that AI inference costs start to look a little bit closer to labor costs than to software costs. Sourcegraph CEO Dan Adler writes, "I spend much of my week every week talking to large enterprise buyers. The appetite for tokens is insatiable. C-level FOMO is off the charts, and every spare dollar is going into Claude Code, Cursor, Amp, etc. Tens or even hundreds of millions of dollars in engineering organizations that cost billions in salaries seems reasonable. But if CTOs can't deliver headcount savings, we're going to see some real whiplash on token budgets in the next two to four quarters." Another way to put what Dan is saying is that something's got to give. The cost of agentic engineering can't keep rising without there being commensurate cost cuts somewhere else in the organization. Anonymous 40 account on Twitter writes, "This marks the beginning of the end of the subsidized inference era. It will only go higher." I think we are just beginning to grapple with what the full-bore cost of AI when fully utilized is going to look like and what it means for the structure of organizations. There is of course, however, another piece of this one that Boris got in that essay that I read before. As he put it, it's also an identity thing. We're engineers, and reviewing code is what engineers do. From some corners, you can almost feel the existential nature of the response. Look at how Montano puts it: "We need to admit defeat. We won't be reviewing code before it goes to production. Humans are already the bottleneck." Now, it wasn't strictly related to this release, but there's been a viral video going around Twitter/X from Mo.io with the caption on the video, "I was a 10x engineer, now I'm useless." It's actually less dramatic than the caption makes it sound, but it's a real, honest exploration of a lot of the feelings that many developers are having right now, as the fundamental nature of what they do as developers has changed underneath their feet almost overnight. And it does feel to me a bit like part of the response to the code review feels a bit like watching the last part of the sandcastle that they've spent their whole lives building washed away into the ocean. Why I think this part matters, regardless of whether you're an engineer, is that as I've frequently said on this show, if you want to understand what other types of knowledge workers are going to be feeling like in a year or two years, watching how developers handle these changes and how the broader shape of their field is shifting is the closest thing we have to peering into the future. There is a deep set of existential questions in this liminal moment, and I think how folks resolve them on a personal and professional and organizational level is going to create a template and a blueprint for how we deal with AI disruption in other areas. Now there is one more piece of the negative response to code review that I think is worth tracking as well, which is not just about cost but about pricing power, and this to me maybe represents another chink in Anthropic's armor, although I don't think it's limited to Anthropic alone. Garb writes, "Feels like the wild west days of pricing. The general store has you hooked on their supply. They know it, and they're telling you how much they're going to fleece you because they can." Eaz writes, "Lol, Anthropic just killed a $50 billion industry with a single feature again. Companies pay $50k a year to scan their code for CVE vulnerabilities. Anthropic code review does it for you in minutes for a fraction of the cost." Broadloom's Todd Saunders uses an analogy: "Anthropic is the new Amazon. Build on our platform, and once you get scale, we will build a basics version of your product and put you out of business instantly." That was a quote tweet from this from Verun Ganesh, who wrote, "At this point, it's pretty clear that if you are an app-layer company using Claude Code SDK, it is inevitable that Anthropic sees your usage and then develops that tool in-house." One of the potential reckonings in the AI space is going to be questions of power and consolidation around the very small number of neutron star companies that are just absorbing everything around them. It is worth pointing out, of course, that this particular product is not guaranteed to work. Certainly the teams at Cognition and OpenAI are using this as a bonanza for their own marketing, and maybe the market will force the price of AI code review down. Still, it's very clear, taking a step back, that the response around the code review product was about more than just price. It cut to the quick of the types of issues that are just going to be part of our everyday in the period that's coming up next. We will of course continue to track this. I will say I have the sense, like Swix, that perhaps while it takes three to six months for everyone to get there, it is very likely to me that this debate or conversation seems kind of quaint in retrospect. For now, that is going to do it for today's AI Daily Brief. Appreciate you listening or watching as always, and until next time, peace.
