Claude Code's Ultra Plan Is Fast But Breaks

Anthropic just dropped Ultra Plan for Claude Code, and the hype machine is already spinning. The feature—which leaked earlier this month and went live literally days later—promises to supercharge how developers plan projects. But here's the thing: when Chase AI put it through a head-to-head test against standard plan mode, the results were... complicated.

The speed difference is genuinely wild. Ultra Plan generated a complete project architecture in 30 seconds. Standard plan mode? Five and a half minutes, and it crashed on the first attempt. That's not a marginal improvement—it's the kind of 10x speed-up that should make developers immediately abandon the old way of doing things.

Except Ultra Plan has a problem, and it's not a small one.

What Actually Changed

The mechanics are straightforward enough. You type "ultra plan" in your Claude Code terminal, and the planning session gets pushed to a cloud-based browser interface instead of running locally. You see a fancy rainbow gradient (because AI features apparently need visual flair now), get a mermaid diagram of your architecture, and can leave emoji reactions on specific parts of the plan. Once you approve it, the plan syncs back to your terminal for execution.

Anthropicâ€TMs documentation is surprisingly sparse about what's actually happening under the hood. The official story is basically: it runs in the cloud, you need a GitHub repo with at least one commit, and you can edit plans more easily in the browser. That's it. No mention of additional agents, enhanced reasoning, or architectural improvements—even though developers testing it report that it feels more capable somehow.

"It does feel like Ultra Plan does have a bit more power behind it," Chase AI notes in the video. "And it's kind of a shame that Anthropic doesn't go a little deeper into what it actually buys us."

The Test That Reveals the Cracks

For the comparison, Chase AI gave both modes the same prompt: build a premium kanban board web app from scratch, use the front-end design skill, here are the requirements, plan thoroughly. Simple enough to evaluate, complex enough to matter.

Ultra Plan delivered its architecture in those 30 seconds. Standard plan mode took over five minutes and needed a restart. But when the actual code got built, something became obvious: Ultra Plan completely ignored the instruction to use the front-end design skill.

The visual difference was immediately apparent. The standard plan mode version had proper typography using Google Fonts, better card shading, color-coded priority indicators, and those little UI flourishes that make an interface feel polished. Ultra Plan's version was... functional. It worked fine, but it looked like a first draft.

"Even though we explicitly put in the prompt use this skill, it totally ignored it," Chase AI points out. "And you see this reflected in things like the typography. Local plan is actually using stuff like Google fonts and it's not even mentioned at all in the ultra plan."

The back-end code quality was roughly equivalent between both versions—Ultra Plan had a few hundred more lines, used slightly different frameworks, but nothing that would make you choose one over the other based purely on architecture. Gary Tan apparently weighed in and said Ultra Plan was better, but the gap wasn't dramatic.

The Speed-vs-Skills Trade-off

Here's where it gets interesting: this might not actually be about Ultra Plan being worse. It might be about the test being too simple.

Ultra Plan's advantages could only shine on projects complex enough to justify the additional computational overhead. A kanban board—even a "premium" one—isn't that complex. It's the kind of project where standard plan mode is perfectly adequate, and where the failure to invoke specific skills becomes the dominant factor.

"This could be a situation where we used ultra plan for a project that was just way too simple," Chase AI suggests. "This may be something where we need to be using ultra plan on something that is infinitely more complex."

But that creates an awkward calculus for developers. If you're building something simple, standard plan mode works fine and actually respects your skill specifications. If you're building something complex, Ultra Plan might be better... except it still might ignore your skills, which becomes a bigger problem on complex projects where those skills matter more.

"Are we balancing speed? Do we want it to be really fast like we get with ultra plan or do we want to actually use the skills?" Chase AI asks. "And does the speed really matter if it's not actually calling in the skills correctly?"

What This Means For Your Workflow

The rushed rollout—leaked one day, shipped two days later with minimal documentation—suggests Anthropic might still be figuring out what Ultra Plan should be. That's not necessarily bad. Early access to evolving tools is part of how modern development works. But it does mean the current version has clear limitations that might not exist in a month.

For now, the decision tree is pretty straightforward:

Building something simple and want to use custom skills? Standard plan mode.
Building something massive and speed is critical? Maybe try Ultra Plan, but verify it's actually doing what you asked.
Building something in between? Honestly, it's unclear which is better.

The skill invocation issue isn't a one-time fluke, either. Chase AI mentions running into the same problem during additional testing outside the video. That's a pattern, not an accident.

The Bigger Picture

What's most striking about this whole situation is how it illustrates the gap between AI tool marketing and AI tool reality. The hype around Ultra Plan positioned it as a clear upgrade—faster, better, the obvious choice. The actual testing revealed something more nuanced: a tool that's legitimately faster but trades off functionality in ways that might or might not matter depending on your specific use case.

That's useful information, but it's not what the launch messaging conveyed. And Anthropic's sparse documentation means developers are left running their own tests to figure out when Ultra Plan is actually worth using.

"I wouldn't be surprised if this also gets a lot more upgrades and changes as we move forward," Chase AI says. "It almost feels like this was kind of just pushed out to get it out there."

Which brings us back to the fundamental question: is Ultra Plan worth using? The answer depends entirely on what you're building and whether you can afford to have your custom skills ignored. That's not the kind of clear recommendation anyone wanted, but it's the honest one.

—Tyler Nakamura