Kilo

Anthropic just disabled Fable and Mythos. Fable 5 was available in Kilo on day one, and the reactions from our team and users ranged from seriously impressed to slightly suspicious. For one of our first internal tests, Brian Turcotte dropped a screenshot of the GitHub UI into Kilo Code and asked Fable to "recreate GitHub." It worked in one shot; functional repos and all. But alongside the excitement, there was real skepticism too. Some users noticed a tendency to over-engineer, and a fair number of people felt that while Fable 5 was clearly exceptional at UI and frontend work, it could be a bit overhyped when it comes to deeper engineering tasks. A solid improvement, no doubt - but is that all it is? Our own Darko G. wrote about this in detail on the Kilo blog. The link's in the comments below. So the question that kept coming up: is this model powerful enough to justify the kind of concern/hype it's generating? The shutdown only adds to that. We don't have a clean answer, but we're curious what you saw. If you got to use Fable 5 before it was pulled, what was your read on it?

Most AI code review tools apply the same generic rubric to every repository. Your frontend repo and your infrastructure repo get reviewed against identical rules, and the agent misses the conventions your team actually enforces. Today's update to Kilo Code Reviews changes that, three features, all live now: 1. REVIEWS.md: an open-standard Markdown file you drop in your repo root, like a README written for the review agent. Write down your conventions, style, and architecture decisions, and the agent applies them on every automated review. Different repos, different standards, no shared config. 2. Code Review Memory: you shouldn't have to document every preference by hand, most already live in how you respond to PRs. Toggle this on and it analyzes your feedback, then proposes updates to your REVIEWS.md based on what it learned. Keep dismissing line-length nitpicks but acting on error-handling comments, and it adjusts its focus to match. It can even open the PR, so your review standards stay version-controlled. 3. Local review suggestions: in VS Code, Code mode now offers a non-invasive review of your uncommitted changes before you push, catching issues before they cost a CI round trip. That wraps Product Week, five launches in five days. This one's a good one to end on: a review agent that gets better the more you use it.

Two new coding models and a sold-out token plan this week in Kilo. Kimi K2.7 Code from Moonshot, Claude Fable 5 topping our coding benchmarks, and the MiniMax token plans selling out fast enough to need a new batch already.

"The most expensive email I ever ignored. I didn't even know I ignored it."

Day one of WorldCupBench: four AI models went 2 for 2 on predictions. Our growth marketer went 1 for 2. Tomorrow gets interesting. The market heavily favors Switzerland over Qatar, but 3 of 4 models disagree. Vote and follow along!

Kilo Console is live in beta. It's a local, browser-based UI for the Kilo CLI. One place to manage your projects, git worktrees, sessions, and settings, without hand-editing JSON. What you can do in it: → Open a project, see its git worktrees, and launch a CLI session in any of them → See where each setting comes from, global config or the project itself → Manage agent tools and permission rules → Connect model providers → Switch UI themes live, no restart It's beta, so feedback is useful. Share it in our Discord! Day four of Product Week. Link in the comments.

What does an open model actually open? Chris Alexiuk of NVIDIA breaks it down in 60 seconds.

Wall Street is getting a crash course in tokens ahead of the AI IPOs. Our CEO talked to CNBC about the part the prospectuses gloss over: what businesses actually get back for the tokens they spend.

We built a small funny side project for the 2026 World Cup: Kilo WorldCupBench. The idea is simple: let different AI models predict the matches, then compare how they perform against each other and against humans. Everyone has same data: team stats, top players stats, and locations. The current panel: - Kimi (Moonshot AI) K2.6 - MiniMax M3 - NVIDIA Nemotron 3 Ultra - DeepSeek AI V4 Pro Each model gets the same match data and makes its call. As the tournament progresses, we’ll see which model reads football best. Mostly built for fun, but also a nice little benchmark for how different models reason about the same real-world events. As a human, feel free to make your choice and compare prediction results later. Link in the first comment 👇

Two developer tools changed the deal. Roo Code shut down. Copilot moved everyone to usage-based billing. The pattern isn't new, and it isn't going away: when your stack hangs on one vendor, their next move is *your* problem. Coding Plans are live in Kilo today. Here's the short version of why they're built the way they are 👇