Grok 4.5 vs. Claude Opus 4.8: Costs and what works, not the spec sheet

Can Grok 4.5 really match Opus for a quarter of the tokens?

xAI released Grok 4.5 on July 8. It was trained with Cursor session data added in training. xAI says the model roughly matches Claude Opus 4.8 on coding. It just uses about 4.2 times fewer output tokens to do it. xAI’s pitch is aimed straight at developers now watching their token spend.

Grok is also cheaper to run out of the box. It costs $2 per million input tokens and $6 per million output. Opus is $5 and $25. That is less than half the price on both.

xAI says Grok 4.5 matches Opus 4.8 on real coding for a fraction of the tokens. I gave both the same three jobs in one Rust repo. Both models wrote nearly identical code but Opus used 4.3 times more tokens.

The benchmarks mostly back xAI’s marketing. Terminal-Bench 2.1 measures how well a model handles real work at the command line. There, xAI puts Grok 4.5 at 83.3%, just above Opus 4.8’s 78.9%. SWE-Bench Pro is the harder one. It scores how well a model fixes real bugs pulled from open-source projects. On that test, Opus still wins. So the marketing claim is not “better.” It is “as good, for far less.”

Just by reading these benchmarks and looking at pricing, it appears as if xAI wants to take some of Anthropic’s market share. Anthropic is the current fan favorite when it comes to AI companies, so for this to become a possibility, the marketing claims will need to hold up.

So the marketing claim is not “better.” It is “as good, for far less.”

In pursuit of this answer, I put Grok 4.5 and Claude Opus 4.8 head-to-head on three jobs in the same codebase and tracked every token.

The tests

I used fd, the popular Rust file-finder from sharkdp. fd is a fast, friendlier replacement for the Unix find command, and I chose it because it is real production Rust with a deep, documented bug history to pull from.

Grok and Opus both ran inside Cursor in Agent mode. The tool and the prompts stayed identical, so the model was the only variable. I built a new folde…