Agent costs are dropping faster than you think

If you priced out AI agent automation six months ago and shelved the idea because the numbers didn’t work, it’s worth running them again.

xAI just shipped Grok 4.5, their first model built specifically for coding and agent workflows. The benchmarks put it near the top tier. The pricing puts it at roughly a third of what comparable frontier models cost. That matters if you’re running agents that chew through millions of tokens a day processing invoices, triaging support tickets, or updating your CRM.

The numbers

Grok 4.5 costs $2 per million input tokens and $6 per million output tokens. Cached inputs drop to $0.50. Compare that to GPT-5.5 at $5/$30 or Anthropic’s Opus 4.8 at $5/$25. The gap is not subtle.

For context, a typical agent workflow that processes a customer email, checks inventory, drafts a response, and logs the interaction might use 50,000 to 100,000 tokens. At Grok 4.5 pricing, that’s under a dollar. At GPT-5.5 pricing, it’s several dollars. Multiply that by hundreds of interactions a week and the difference pays for the automation stack itself.

Artificial Analysis, an independent benchmarking outfit, measured Grok 4.5 at roughly 60% lower output token usage than Opus 4.8 on equivalent tasks. It produces less waffle to get the same result. That efficiency translates directly to your bill.

Why this changes the business case

The main reason most Australian businesses haven’t deployed AI agents for routine operations isn’t capability. The models have been capable enough for structured tasks like data entry, scheduling, invoice matching, and CRM updates for over a year. The blocker has been cost.

Running an agent that monitors your inbox, triages incoming requests, and routes them to the right person costs real money when every token is priced at flagship rates. Most businesses looked at the per-interaction cost and decided a human could do it cheaper, or at least that the savings didn’t justify the setup effort.

That calculation is shifting. When the same agent workflow costs a third of what it did six months ago, the break-even point moves from “maybe worth it at scale” to “probably worth it next quarter.”

The Cursor angle

One detail worth noting: Cursor, the AI coding tool used by millions of developers, partnered with xAI to train Grok 4.5. They didn’t just integrate it. They helped build it.

This matters for business owners because it signals where the engineering talent and training data are concentrating. Coding agents are the proving ground for business automation agents. The same patterns that let an AI write and debug software are the patterns that let it process your accounts payable or manage your project board. When the best coding tool in the market co-builds a model, that model’s strengths tend to show up in business automation use cases shortly after.

What to actually do with this

If you’ve been watching the agent space from the sidelines, the practical move is not to rush out and deploy everything at once. It’s to pick one repeatable, structured workflow and run a costed pilot.

Good candidates:

Email triage and routing
Invoice data extraction and entry
Support ticket classification and first-response drafting
CRM record updates from meeting notes

Pick one. Measure the token cost at current pricing. Compare it to the hourly cost of the person currently doing it. The numbers are more favourable than they were even three months ago, and they’re likely to keep improving.

The broader picture

GPT-5.6 Sol launches this week. Anthropic extended Fable 5 access for paid users. Google’s Gemini API added managed agent features with background execution. The competitive pressure across every major provider is driving prices down and capability up simultaneously.

For businesses that have been waiting for the economics to make sense, the window is opening. Not because any single model solves everything, but because the cost floor keeps dropping while the capability floor keeps rising. At some point the spreadsheet stops being a maybe and starts being an obvious yes.

We’re getting close to that point.