Use a plain script when inputs are structured and the rules don't change; reach for an LLM pipeline when you deal with messy text, photos, or decisions that need judgment. Kimi K3.1 Pro just made the second option cheaper, which shifts where that line sits.
Moonshot launched Kimi K3.1 Pro on their API at $2 per million input tokens and $10 per million output - half the price of Claude Opus 5.5 and matching GPT-6.1 Sol. Artificial Analysis scored it 55, behind only Opus 5.5 at 58. Open weights are promised by November 4, with a Highspeed tier that runs 2-3x faster for double the price.
What it changes for small business pipelines
If the task is parse a fixed CSV and push it into Google Sheets, a 20-line Python script still wins - no model, no latency, no token cost. But for invoice OCR cleanup, lead classification from free-text WhatsApp messages, or a tender analysis pipeline I built, price per million tokens is a real line in the monthly budget. A frontier-grade model at half the Opus price - and soon self-hostable - means you can batch more documents, keep a larger context, or run a verification pass that was too expensive a week ago. The catch: Chinese API, data residency rules differ, and the headline benchmarks are Moonshot's own, so test on your real data before switching production.
Why it matters: Cheaper strong models don't replace scripts, they expand the zone where an LLM-backed workflow beats hand-rolled logic. If you shelved an idea six months ago because tokens were too expensive, re-price it this week.
Source: Simon Willison