Cheap AI Agents Are Here. The Catch Is Real.
xAI's new Grok 4.5 makes running AI agents four times cheaper — but its hallucination rate doubled, so teams need to think carefully about which work they hand it.
- $6/M
- output token price for Grok 4.5
- 4.2x
- fewer tokens than Claude Opus 4.8
- 54%
- hallucination rate, up from 25%
- $2.49
- typical task cost vs $11.80 for Claude Code
On July 8, xAI launched Grok 4.5, a new AI model priced at $6 per million output tokens — the little chunks of text an AI generates. That sounds technical, but the business math is simple: running AI coding agents just got a lot cheaper. Compare that to Anthropic's Claude Opus 4.8 at $25 per million, or Claude Fable 5 at $50. For teams already paying for AI automation, this is a significant shift.
The price advantage gets bigger when you look at how many tokens Grok 4.5 actually uses. On standard coding tests, it completes tasks using about 15,954 output tokens — compared to 67,020 for Opus 4.8 doing the same job. That 4.2x efficiency gap means a typical coding task costs $2.49 instead of $11.80. For any team running dozens or hundreds of automated tasks a day, that difference adds up fast.
But there is a serious catch. While Grok 4.5's raw accuracy improved — from 35% to 52% correct answers — its hallucination rate also jumped. Hallucination means the model confidently gives a wrong answer, making up facts or code. That rate jumped from 25% to 54%. The model gets more things right overall, but when it gets something wrong, it now does so with greater confidence.
The practical takeaway is about choosing the right tool for the right job. Grok 4.5 makes sense for high-volume, low-stakes work — drafting, code scaffolding, first versions of anything. For work that goes to clients, involves legal or financial facts, or needs to be reliably accurate, stick with a frontier model and validate outputs carefully. The price of AI agents falling is good news; the new skill is knowing which jobs deserve the cheaper rate.