On Wednesday I gave Grok 4.5 the kind of job I would normally hand a junior. Forty inbound leads from a week of enquiries, sorted into hot, warm, and ignore, one line of reasoning each. It took about nine seconds and cost me less than a kopi. Thirty-eight of the forty came back sorted the way I would have sorted them. One warm lead it marked hot, which is a cheap thing to be wrong about. One it filed with a reason it had invented, a detail about the company that appears nowhere in the email, and that one I would have been embarrassed to forward.

xAI put Grok 4.5 on public release on 8 July at two dollars per million input tokens and six per million output, a model Elon Musk called Opus-class and roughly comparable to Anthropic’s Opus 4.7, which lists at five and twenty-five. On coding work it resolves a task in about a quarter of the tokens. It is also, by Artificial Analysis’s measure, more than twice as willing to make things up as the model before it: its hallucination rate climbed from 25 percent to 54, and independent testers ranked it fourth, behind Fable 5, GPT-5.5, and Opus 4.8. Both facts hold at once. The model got cheaper, faster, and more confident when it does not know.

The model got cheaper, faster, and more confident when it does not know.

Grok was the third such move in a week. OpenAI made GPT-5.6 Terra generally available the day after, at $2.50 and $15; Meta opened its first paid model, Muse Spark 1.1, at $1.25 and $4.25, around a quarter of what the American frontier charges. Beneath all of them sits the Chinese floor, where a resolved coding task on DeepSeek V3.2 runs about 2.8 cents and Kimi K2.6 clears eighty percent of a standard benchmark for roughly seven, against Opus 4.5’s sixty-eight. For a founder in Jakarta or Ho Chi Minh City, the number that decides anything is cost per useful task, and this week it became coffee money.

Which turns the question from which model to buy into which jobs to point it at. Cheap-and-occasionally-wrong is close to ideal for work that is high in volume, low in stakes, and easy to reverse: a first draft of a caption, a first pass of code you were going to read line by line anyway, a pile of leads to triage, a supplier spreadsheet to clean before a person signs it off. In each of those the wrong answer is caught by the next step, and the next step was always going to be human. The saving is real and the exposure is near zero.

On those tasks the 54 percent is a live probability: the odds that the cheapest part of your stack is the one talking to the person who pays you.

The map has a hard edge. On the far side sit the tasks where a wrong answer runs past the draft stage: the reply that goes straight to a customer, the figure that lands on an invoice, the number that moves money out of an account. A hallucination there carries a real price, a refund and a relationship you do not get back, and no per-token saving covers it. On those tasks the 54 percent is a live probability: the odds that the cheapest part of your stack is the one talking to the person who pays you.

The unglamorous move, while the introductory pricing holds, is to find the two or three jobs in an ordinary week that sit on the safe side of that edge, wire the cheap model into those, and leave the customer-facing and money-moving work on whatever you already trust. This is roughly what the careful operators already do. OCBC in Singapore runs more than thirty internal tools on open models and routes each task by job, Qwen for code, DeepSeek for market reading, rather than staking the whole business on one engine. The model is the commodity; deployment is the part you actually control.

The lead I sorted for two cents is still in my pipeline, correctly filed. The one the model invented a reason for is the one I keep thinking about. I would have sent it, in my own name, and the model would have been faster, cheaper, and wrong, all at once.1

Footnotes

  1. Grok 4.5 was trained in part on real Cursor coding sessions, which is either the point of it or the problem with it. An earlier snapshot of Cursor’s own codebase leaked into the training data, and CursorBench, the benchmark meant to measure the model, had to be withdrawn. It studied for the test by memorising part of the room. ↩