Opus 5.5 Tested: One Prompt Yields 100 Pages, Animations, and Music

After the release of Opus 5.5, users quickly shared hands-on results. A creator known as MiaAI Lab gave the model an open-ended instruction: produce 100 HTML files that are visually appealing, non-repetitive in design, and maximally creative. The model autonomously decided what to build, wrote the page structure, CSS, and interaction code for each, delivering a varied set of web pages, animations, and games. This illustrates the workflow Anthropic now strongly recommends: provide a complete task description, define clear completion criteria, and specify only the conditions under which the model should pause and ask for clarification.

Anthropic's usage guide exemplifies this with a payment-gateway migration: the prompt states the goal, the definition of done (all interfaces use the new client, old client removed, tests pass), and the sole escalation trigger (unexplained test failures). The company emphasizes that Opus 5.5 sustains progress over long, multi-step tasks better than its predecessor. Official materials cite an early tester who had the model review and fix roughly 200,000 lines of code in under three hours — a task that took Opus 5 more than 20 hours. In an internal benchmark, Opus 5.5 rewrote HAProxy from C to Rust in 9.5 hours, and the resulting code passed nearly all of the original project's regression tests.

However, extended autonomy does not guarantee the right output. Tech outlet Every gave Opus 5.5 ten minutes to produce a minute-by-minute schedule for a client training session. The model spent nearly five minutes thinking, generated training data, verified tables, and wrote attendee materials, but the requested schedule itself was not delivered when time expired. In another application-development test, the UI looked polished and automated checks passed, yet manual interaction revealed errors on core pages. These cases show that one-shot delegation remains unreliable for time-sensitive or correctness-critical deliverables.

On the writing front, Opus 5.5 shows a marked style shift. Users report the model's output is now more direct, natural, and free of the characteristic "Claude tone." A side-by-side comparison shared by X user @theojaffee shows Opus 5's verbose, em-dash-heavy prose versus Opus 5.5's concise, em-dash-free version. Anthropic's own release notes confirm the model communicates more naturally, fronts important information, and adheres better to user-specified writing rules.

For document understanding, LlamaIndex CEO Jerry Liu evaluated Opus 5.5 on ParseBench, a PDF table extraction benchmark. The model scored 93.9%, a gain of over seven percentage points over Opus 5, outperforming Fable, Gemini, and Astra. Weaknesses remain in chart, formatting, and layout comprehension.

Creative coding capabilities have also expanded. X user Kevin Ngo prompted Opus 5.5 to draw a frame-by-frame animation in JavaScript, producing a smooth visual piece. Another user asked for "bass music with javascript" in a single sentence; the model returned complete, runnable JS code that generated a listenable bass track. These examples highlight the model's growing ability to translate high-level creative intent directly into executable code across media types.