SpaceXAI releases Grok 4.5 at $2/$6 per MTok; claims parity with Opus 4.7, faster
SpaceXAI released Grok 4.5 on July 8, 2026, positioned as an Opus-class model built on xAI's new 1.5-trillion-parameter V9 foundation. Elon Musk claimed internal assessments show it is roughly comparable to Anthropic's Claude Opus 4.7 "but much faster." The model was trained on tens of thousands of NVIDIA GB300 GPUs at Colossus (SpaceX's Memphis supercomputer cluster) with supplemental coding data from Cursor, which SpaceX agreed to acquire for $60 billion in June 2026.
Grok 4.5 is priced at $2 per million input tokens and $6 per million output tokens—roughly 60% cheaper per input token than Claude Opus 4.8 ($5 input, $25 output) and GPT-5.6 Sol ($5 input, $30 output), both also released Thursday. Independent benchmarks show mixed results: Grok 4.5 scored 64.7% on SWE Bench Pro (a software engineering task resolver), beating GPT-5.5's 58.6% but trailing Opus 4.8 (69.2%) and Claude Fable 5 (80.4%). Musk's claim of Opus parity rests on SpaceX's internal evaluations, not public benchmarks.
For teams running high-volume coding workloads, the cost-performance tradeoff favors Grok 4.5; for frontier performance, Fable 5 leads. The release is SpaceXAI's first major product under the merged SpaceX-xAI organization (closed February 2026) and first visible fruit of the Cursor integration. Musk signaled an aggressive cadence: one new foundation model every month through end of 2026. EU access is blocked at launch, rolling out mid-July. Developers should validate performance claims on their workloads before migrating high-stakes tasks.
Sources
- Primary source
- crypto.news
“Grok 4.5 costs $2 per million input tokens and $6 per million output. Claude Opus 4.8 runs $5 input and $25 output.”
- tech.yahoo.com
“Elon Musk's SpaceXAI released Grok 4.5 on Wednesday. It targets coders, engineers, and knowledge workers. The company's pitch is that it's cheap for a western model.”
- basenor.com
“xAI dropped Grok 4.5 on July 8, 2026, built on a new foundation and trained on tens of thousands of NVIDIA GB300 GPUs”