Claude Opus 4.7 Promises Coding Dominance—With Caveats

Anthropic dropped Claude Opus 4.7 yesterday, and the AI developer community is already running stress tests. The model promises to be Anthropic's most capable coding assistant yet—handling long-running tasks with less supervision, following instructions more precisely, and even self-verifying its output. But as WorldofAI's testing reveals, "most capable" doesn't automatically mean "best choice for every use case."

The benchmarks are genuinely impressive. Opus 4.7 outperforms its predecessor (4.6), GPT-4, and Gemini 3.1 Pro on SWE-Bench Pro and Verified—the standardized tests that measure how well AI models can solve real-world software engineering tasks. It's also hitting state-of-the-art performance across finance and legal knowledge tasks, suggesting this isn't just a narrow coding improvement.

What's particularly interesting: Anthropic has closed the gap in web development. Previous Opus models lagged behind Google's offerings for UI generation, but 4.7 is now "on par with Gemini 3.1 Pro" for frontend work, according to the tester's evaluation. That's a meaningful shift if you've been choosing models based on what they're good at rather than who made them.

The Reasoning Upgrade (and Its Price)

One of the more technical improvements involves reasoning efficiency. Opus 4.7 essentially bumps every reasoning tier up a level—what used to require "medium" effort now happens at "low," and so on. In theory, this means better outputs for the same computational cost.

In practice? The model is absolutely devouring tokens. The tester hit rate limits on a single prompt when running Opus 4.7 at maximum reasoning levels. Anthropic responded by increasing usage limits for subscribers, which is less a fix and more an acknowledgment that this model just needs more resources to do its thing.

"This model uses a lot more tokens," the video notes. "It creates a trade-off. Higher quality versus reduced usable context."

The pricing structure hasn't changed ($5 per million input tokens, $25 per million output tokens), but if you're getting fewer tasks done within the same token budget, your effective cost per project increases. For hobbyists burning through free credits, this might not matter. For teams running production workloads, it's a real consideration.

Real-World Testing: The Good and the Janky

The video runs Opus 4.7 through several practical tests using Kilo CLI, an open-source AI coding agent. The results are... mixed in interesting ways.

The model generated a 3D physics simulation of an SUV in a mountain environment, breaking down the massive request into systems for physics engine, rendering, camera controls—all the architectural decisions you'd want an AI to handle competently. "Not the best obviously, but it is the best any model has generated for this test," the tester concludes.

The Minecraft clone is where things get ambitious. Opus 4.7 built a version with different ores, procedurally generated terrain, water physics, mobs, and an ore system. It's "definitely the best Minecraft clone I've seen a model generate," but also "a little buggy" with execution issues. This feels like the essence of current AI coding: wildly creative system design paired with implementation gaps that a human still needs to patch.

The macOS-styled operating system demo is genuinely impressive—accurately cloned UI components, working menu bars, functional apps (well, some of them). The Finder works. Spotlight search works. But iMessage, Mail, Maps, Photos, FaceTime? Not coded out. The Settings app is incomplete. It's a beautiful facade with half-finished rooms behind it.

Where It Actually Regressed

Here's the thing that complicates the "most powerful ever" narrative: Opus 4.7 is worse at SVG generation than its predecessor.

The tester requested a PS5 controller in SVG code. The result was... not good. Comparing it directly to output from Qwen (a 35-billion-parameter model) shows "a substantial decrease in quality." The model that's supposed to be Anthropic's flagship is getting beaten by smaller models on specific creative tasks.

"In certain cases, it's not able to creatively focus on SVG generations as it used to with the 4.6," the video observes.

This isn't surprising if you understand how these models are trained and tuned, but it's a useful reminder that progress isn't linear. Making a model better at complex reasoning might make it worse at tasks that benefit from a different kind of pattern matching.

The Instruction-Following Problem

Anthhropic explicitly warns that "prompts built for the Opus 4.6 may break" with 4.7 because the new model interprets instructions "much more literally." This is being framed as improved precision, but it's also a compatibility nightmare for anyone with established workflows.

If you've spent months fine-tuning your prompts to work around 4.6's quirks—adding extra context here, being deliberately vague there—you now get to start over. The model that follows instructions more precisely might paradoxically be harder to instruct effectively, at least until you figure out its new interpretation patterns.

What "Best" Actually Means

So is Claude Opus 4.7 the best coding model right now? It depends entirely on what you're optimizing for.

If you're tackling genuinely complex, long-horizon tasks that require sophisticated reasoning and can tolerate higher token costs—probably yes. If you're doing rapid prototyping where speed and token efficiency matter more than architectural elegance—maybe not. If you need SVG graphics or have finely-tuned prompts from the previous version—definitely not.

The tester's frontend work shows Opus 4.7 has developed a distinctive style—particular color palettes and typography choices that appear consistently across different landing page generations. That's either a feature (consistent aesthetic) or a limitation (harder to get variety) depending on your needs.

What's clear is that we're past the phase where one model dominates every use case. The question isn't "which model is best" but "which model is best for this specific thing I'm trying to do right now." And that answer might change by the time you read this sentence.

—Yuki Okonkwo