xAI shipped Grok 4.5 on July 9, and the interesting part is not another Opus class benchmark. It is how the model gets there. Elon called it Opus class but faster, more token efficient, and cheaper. For anyone running agents at volume, token efficiency is the number that quietly decides your bill.
The efficiency story
On SWE Bench Pro, Grok 4.5 resolves tasks with about 15,954 output tokens on average. Opus 4.8 at max spends around 67,020 for the same work. That is roughly four times fewer tokens to reach a comparable result. Pricing lands at 2 dollars per million input and 6 per million output, and it serves at 80 tokens per second. The benchmark scores are solid, Terminal Bench 2.1 at 83.3 and SWE Bench Pro at 64.7, but tokens per task is the line that changes your math.
Why this matters for builders
Most of the industry spent two years chasing bigger models. The 2026 shift is toward doing the same job with fewer tokens, because in an agent loop you pay for every step, every retry, every tool call. A model that finishes in a quarter of the tokens is not just cheaper, it is faster and less likely to drift over a long run. My advice: do not read the leaderboard, read your own logs. Take one agent workflow, run it on Grok 4.5, and measure output tokens per completed task against what you use today. If the gap holds on your traffic, the savings compound every day. It is not in the EU yet.