SpaceXAI Grok 4.5 launches at 60% cheaper pricing, but trails on benchmarks
In brief
- Grok 4.5 released Wednesday, SpaceXAI's first model since SpaceX-xAI merger closed in February
- Pricing: $2 input/$6 output per million tokens, 60% cheaper than Claude Opus 4.8 and GPT 5.6 Sol
- Benchmark performance trails Opus 4.8 and Anthropic frontier models, beats GPT 5.5 on SWE Bench Pro
- Model trained on Cursor AI developer session data using Colossus, SpaceX's Memphis supercomputer
Pricing and positioning
Grok 4.5 costs $2 per million input tokens and $6 per million output tokens, undercutting rivals substantially. Claude Opus 4.8 costs $5 input and $25 output per million tokens, while GPT 5.6 Sol runs $5 input and $30 output. That's a 60% savings on input versus Anthropic's flagship.
Musk's own framing was candid. He called Grok 4.5 "roughly comparable to Opus 4.7, but much faster." Opus 4.7 is Anthropic's previous flagship; Opus 4.8 has since succeeded it. The pitch, then, isn't that SpaceXAI built the best model. It's that it built a capable one at a price competitors can't match.
Benchmark results tell a mixed story
On DeepSWE 1.1, Grok 4.5 scored 53%, behind Claude Opus 4.8 at 59% and GPT 5.5 at 67%. Claude Fable 5, Anthropic's frontier model, topped the chart at 70%. On SWE Bench Pro, the picture shifted. Grok 4.5 posted 64.7%, enough to beat GPT 5.5's 58.6%. Opus 4.8 still leads at 69.2%, and Fable 5 sits at 80.4%.
Efficiency is where Grok 4.5 shines. On SWE Bench Pro tasks, Grok 4.5 used an average of 15,954 output tokens to complete each job. Opus 4.8 burned through 67,020 tokens for the same work, a 4.2x gap. The model also runs at 80 tokens per second, which is fast-model territory.
Training and deployment
SpaceXAI trained Grok 4.5 in collaboration with Cursor AI on tens of thousands of Nvidia GB300 GPUs inside Colossus, the Memphis supercomputer. Colossus has total capacity across more than 200,000 GPUs. That scale matters—Grok 4.5 was trained on developer session data from Cursor, including debugging traces and real code edits rather than synthetic datasets.
The model is available via API, on Hermes and Grok build with half a million tokens of context. SpaceX is in the process of acquiring Cursor.
