- Published on
We made Grok 4.5, GPT-5.5, and Claude build the same apps
- Authors
- Name
- DeepHorus
xAI launched Grok 4.5 this week, calling it their smartest model yet, trained alongside Cursor for coding and agentic tasks. To evaluate its real-world performance, a head-to-head test was conducted pitting Grok 4.5 against GPT-5.5, Claude Opus 4.8, and Claude Fable 5 on the same three app-building challenges. Each model was given a single prompt to create a self-contained HTML file with no external dependencies, and the results were judged on functionality, aesthetics, and cost.
These build-offs have become a popular way to gauge coding capabilities, with the internet often posting clips of raw outputs. The test used three prompts: a 3D Rubik's Cube with scramble and solve buttons, a particle gravity sandbox with trails and attractors, and a playable Breakout game. Every output was loaded in a real browser, and only one retry was allowed if an app failed to render. The goal was to see which model could produce working applications in a single shot.
In the first round, the Claude models dominated. Opus 4.8 and Fable 5 both produced a properly colored, animated Rubik's Cube on their first try, while Grok 4.5's initial attempt rendered only a blank void. After a retry, Grok 4.5 succeeded with a clean cube. GPT-5.5 performed poorly, rendering a single dark face with minimal color, and was deemed the odd one out.
The second round saw all models produce functional sandboxes, but GPT-5.5 stood out for its mesmerizing neon attractors and dense, swirling trails. Grok 4.5 offered a clean orbital design with colorful streaks, while Fable 5 leaned into soft glowing orbs and Opus 4.8 produced a busy, web-like particle field. GPT-5.5 won on aesthetic appeal. For the third round, every model delivered a polished Breakout game with score, lives, and glowing paddles, all on the first attempt.
Beyond quality, speed and cost were measured. Grok 4.5 achieved its first token in under half a second and streamed at roughly 110 tokens per second, roughly double the others, making it the cheapest reply. However, its median wall-clock latency was only mid-pack due to verbosity, and its p95 latency spiked to 9 seconds. GPT-5.5 was snappiest on short answers, while Opus 4.8 balanced performance across tasks. These metrics reflect xAI's pitch of delivering intelligence per unit of time and cost.
The results highlight the trade-offs developers face. Grok 4.5 excels in throughput and cost but may lag in code quality for complex tasks, as seen in the Rubik's Cube round. GPT-5.5 offers fast, visually appealing outputs for simpler interactive apps. The Claude models demonstrated reliability and polish, especially for demanding 3D applications. No single model dominated across all criteria, indicating that the best choice depends on the specific task requirements.
Looking ahead, the competition among coding-focused models is intensifying. xAI's emphasis on efficiency positions Grok 4.5 as a strong contender for scenarios where speed and cost are paramount. Meanwhile, GPT-5.5 and Claude models continue to advance in creative and precise code generation. As developers increasingly adopt these tools, future benchmarks will likely focus on consistency across diverse tasks, pushing each provider to refine their strengths and address weaknesses.
