Grok 4.5 wins on speed and cost in a four-model coding build-off

TryAI ran a head-to-head test of four frontier models — xAI’s newly launched Grok 4.5, GPT-5.5, Claude Opus 4.8, and Claude Fable 5 — giving each identical one-shot prompts to generate self-contained HTML apps: a 3D Rubik’s Cube, a particle gravity sandbox, and a Breakout clone. The two Claude models were the only ones to render a working animated cube on the first try, the hardest stateful task; Grok blanked initially but recovered on its single allowed retry, while GPT-5.5 managed only a single dark face. The gravity sandbox and Breakout rounds were near-ties, with GPT-5.5 taking the prettiest sandbox and all four shipping playable brick-breakers.

The more consequential result came from the performance harness. Grok 4.5 hit first token in under half a second, streamed at roughly 110 tokens per second — about double the rest of the field — and was the cheapest to run, though its verbosity and a spiky ~9s p95 latency muddied its median wall-clock. GPT-5.5 was fastest on short answers, Opus sat in the middle, and Fable 5 was both slowest and priciest. A bonus hand-authored SVG round went to Fable for comic execution, while Opus shipped a duplicate-attribute bug that breaks strict parsers.

The takeaway: Grok is the speed-and-value pick for high-volume codegen where latency and cost compound, the Claude models are the most reliable builders at a premium, and GPT-5.5 is the stylish but less consistent option. Worth noting the piece is published by TryAI, which sells pay-as-you-go access to all four models, so it doubles as marketing — but the raw generated apps and methodology are exposed for readers to reproduce.