Grok 4.5 Fully Tested: Matches GPT 5.5 Coding for $2 per Million Input Tokens
TL;DR
Grok 4.5 is positioned as a low-cost coding model: $2 per million input tokens, $6 per million output tokens, 80 tokens per second and an 83.3% Terminal Bench score. The article highlights debugging, prototyping, front-end components, SVGs and animations as strong use cases. It says Grok 4.5 is weaker on 3D modelling, physics prompts and very complex engineering work. The source is PR-heavy: many claims come from World of AI testing and are not backed by official model cards or broad independent benchmark data.
Nauti's Take
Grok 4.5 sounds less like frontier magic and more like a middle-distance model: lots of code, lots of small tasks, a lower token bill. That is where the toughest competition is forming.
Still, turning an 83.3% Terminal Bench score into GPT-5.5 parity is benchmark theatre dressed up as product truth. Real bug fixes, refactors and UI diffs will show whether the price advantage survives actual work.
Briefingshow
If the numbers hold, Grok 4.5 mainly lowers the cost of everyday coding agents. For teams, the key question is not whether it is the best model, but whether it can handle many debugging, scaffolding and UI tasks cheaply enough. The catch: one benchmark and a video test do not replace an internal eval on real repos.
Video
Sources