Is Claude Sonnet 5 Anthropic's Worst Model Ever? WorldofAI's Harsh Verdict

WorldofAI opens favorably. Sonnet 5 is the Sonnet series' biggest update, its most agentic Sonnet model yet. Reduced hallucination, better agentic performance, improved tool use, and the capability to autonomously operate browsers and terminals at a level once reserved for much larger, more expensive models.

The numbers back it up. Sway's verified Agentic Coding score comes in at 63.2%, a huge jump from Sonnet 4.6 and within about 6 points of Opus 4.8. Terminal Bench 2.1 reaches 80.4%, closely rivaling Opus. HLE lands nearly level, and computer use plus GDP Evolve scores actually outperform Opus in places. Overall, Sonnet 5 ranks 5th and is already rolled out across every Claude plan.

The hands-on test builds a Mac OS clone. Launchpad, Safari, messenger, mail, photos, calendar, notes, reminders, maps, music, and even a functional FPS shooter all get generated — each app rendered in SVG, which WorldofAI notes as a nice visual touch.

But the concerns are just as clear. The build took roughly 40 minutes and burned heavy tokens in max mode — efficiency-wise, not good at all. The maps app looks noticeably worse than what Opus models produce. Overall verdict: functional, but not something to praise unreservedly.

The second half of the video turns. What surprised WorldofAI most was the new tokenizer's efficiency — it burns more tokens without delivering the results you'd expect from a newer model. That makes it hard to justify using this model over Opus 4.8, WorldofAI states directly.

On creative tasks like SVG generation, the design taste falls short of models like GLM — WorldofAI goes as far as ranking Sonnet 5 below GLM 5.2, and recommends against using it for general-purpose work at all. The video closes anticipating Fable 5's near-term return, but lands on a harsh verdict for Sonnet 5 itself: 'you'll get better results sticking with Opus 4.8.'