Grok 4.5 Dominates CursorBench, Achieving Superior Performance to GPT-5.5 at Half the Cost
Key Takeaways
- ▸Grok 4.5 delivers the best cost-to-performance ratio on CursorBench at $1.51/task with 66.7% accuracy—roughly 50% less expensive than GPT-5.5 despite superior performance
- ▸Claude Fable 5 posts the highest accuracy (70.5%) but at $17.32/task, while Grok 4.5 offers comparable mid-tier performance for a fraction of the cost across all three settings
- ▸Grok 4.5's benchmark results are tainted by training data contamination: an earlier Cursor codebase snapshot was accidentally included, creating an unquantified but acknowledged advantage
Summary
CursorBench 3.2, a comprehensive benchmark for code editing tasks, reveals that xAI's Grok 4.5 achieves best-in-class cost efficiency while maintaining competitive performance against leading AI models. On tasks evaluating instruction-following, advanced tool use, codebase understanding, bug-finding, and code review, Grok 4.5 at high settings scored 66.7%—exceeding OpenAI's GPT-5.5 (58.4%) and rivaling Claude Opus—while costing just $1.51 per task compared to GPT-5.5's $2.05. The cost advantage persists across all Grok settings: medium at $1.54 for 65.4% accuracy, and low at $1.22 for 63.5% accuracy.
The benchmark data demonstrates stark cost-performance tradeoffs across the model landscape. While Claude Fable 5 achieved the highest raw score at 70.5%, it commands $17.32 per task—more than 11x Grok 4.5's cost. Claude Opus (max: 62.3%, $5.77) and Sonnet (max: 61.5%, $6.45) similarly show diminishing returns on spending, whereas Grok's tiered approach offers architects meaningful flexibility at every price point.
Critically, however, xAI disclosed that Grok 4.5 received an unfair advantage: an earlier snapshot of the Cursor codebase was accidentally included in its training data, a contamination absent from competitor models. While the exact score impact is unknown, xAI has removed the data for future models. This caveat significantly qualifies the headline claim of superiority and suggests Grok's benchmark performance may not reflect real-world competitive advantages.
- The benchmark reveals a clear market segmentation: premium accuracy (Claude, OpenAI) at high cost vs. pragmatic efficiency (Grok) at lower cost
Editorial Opinion
Grok 4.5's cost efficiency is genuinely compelling and represents real progress toward economical AI-assisted development, but the training data contamination issue substantially undermines the headline narrative. xAI's transparency about the Cursor codebase inclusion is commendable, yet without quantifying the exact score impact, claiming superiority over competitors is premature. Developers should view this benchmark as validating Grok 4.5's price-to-performance value proposition while awaiting neutral third-party evaluations before treating it as the definitive performance leader.
