SpaceX Corp shipped Grok 4.5 this week — a 1.5 trillion parameter model trained on tens of thousands of GB300 GPUs, and the first model built in direct collaboration with the Cursor IDE. Two numbers carry the headline: Grok 4.5 finishes a SWE-Bench Pro task on roughly 16,000 tokens where Claude Opus 4.8 burns about 67,000, and it lists input at $2 per million tokens versus $5 for Opus 4.8 and $5 for OpenAI's GPT 5.5. That's not a benchmark delta — it's a budget line item.
For the past year, the agentic coding loop has been priced per turn. Every patch, diff, tool trace, and re-feed compounds into the context window. Token cost is the silent tax on leaving a coding agent running. A model that needs four times fewer tokens on the same task changes the math on what an "agentic loop" actually costs to operate.
The receipt: pricing per million tokens
The price table is the cleanest part of the release. SpaceX published the four-way comparison directly, and it deserves the obvious presentation:
Model
