- Published on
The New Grok 4.5 Is Out. Elon Musk Says It Competes With Last Year's Claude Opus
- Authors
- Name
- DeepHorus
SpaceXAI released Grok 4.5 on July 8, positioning it as a cost-effective alternative to leading models from Anthropic and OpenAI. Priced at $2 per million input tokens and $6 per million output, it is less than half the cost of comparable models like Claude Opus 4.8 and GPT 5.6 Sol. Elon Musk, CEO of SpaceXAI, declared on X that the new model is 'roughly comparable to Opus 4.7, but much faster,' explicitly prioritizing speed and price over top-tier benchmark performance.
Grok 4.5 is SpaceXAI's first public model since the SpaceX-xAI merger closed in February and amid a pending $60 billion deal to acquire Cursor. The model targets coders, engineers, and knowledge workers, including lawyers and finance teams. Access is not yet available in the EU, with European availability expected in mid-July.
According to SpaceXAI's published benchmarks, Grok 4.5 scored 53% on DeepSWE 1.1, a measure of software bug-fixing reliability, trailing behind Claude Opus 4.8 (59%), GPT 5.5 (67%), and Claude Fable 5 (70%). On SWE Bench Pro, which evaluates software engineering problem resolution, Grok 4.5 achieved 64.7%, outperforming GPT 5.5 (58.6%) but still behind Opus 4.8 (69.2%) and Fable 5 (80.4%). The benchmarks compare against GPT 5.5 rather than the newly launched GPT 5.6 Sol.
SpaceXAI trained Grok 4.5 using tens of thousands of Nvidia GB300 GPUs in the Colossus supercomputer, which has over 200,000 total GPUs. This massive compute investment yields a model that is competitive but not state-of-the-art. The pricing strategy marks a shift: Grok 4.5 is designed for cost-sensitive enterprises needing efficient AI at scale.
The model's efficiency stands out. On SWE Bench Pro, Grok 4.5 used an average of 15,954 output tokens per task, compared to Opus 4.8's 67,020 tokens—a 4.2x difference. This token efficiency compounds with lower per-token pricing, enabling more iterations for the same budget. The model also delivers 80 tokens per second, positioning it in fast-model territory.
Musk framed the tradeoff as deliberate, pointing to real-world use by engineers at Tesla and SpaceX as evidence of practical utility. He emphasized that speed and cost savings matter more than raw benchmark scores for many applications. The training data included developer sessions from Cursor, incorporating debugging traces and real code edits.
Looking ahead, SpaceXAI's strategy of leveraging immense compute for cost-efficient models may reshape enterprise AI adoption. However, the model's limited availability outside the EU and mixed benchmark results suggest it will target niche use cases initially. As Musk noted, Grok 4.5 is not the best, but it aims to be the most practical for budget-conscious developers and firms.
