An open model is now within a few points of the frontier
For two years, running AI on your own hardware meant accepting a weaker model — you traded capability for control. Moonshot AI's Kimi K3 broke that trade: the first open-weights model of its size class, it scored 57 on the Artificial Analysis Intelligence Index — third overall at release, comparable to Claude Opus 4.8 and GPT-5.5 on the same index, and ahead of every other open model. The full weights are on Hugging Face, free to download and run.
The leaderboard isn't the point. The point is that "run it yourself" is no longer the fallback option — it's a version of "use the good model." If the reason you stayed off public AI tools was that the local ones weren't good enough, that reason is aging fast.
Read more: run it yourself → Or start with: is your data safe with AI? →Source: Artificial Analysis, Jul 2026 Weights: huggingface.co/moonshotai/Kimi-K3
A capable model now costs 15 cents per million input tokens
That's not a rounding error or a promo — it's DeepSeek's listed price for V4.1 Flash, a 1M-context model: $0.15 per million input tokens, $0.60 per million output at off-peak rates, with cache hits priced at fractions of a cent. And the floor keeps moving: DeepSeek made its steep discounts permanent in May, explicitly escalating a price war, and hosting providers have been cutting against each other all fall.
At those rates, heavy daily use is pocket change. That changes the buying question. It's no longer only "which tool is good," but "is a capable-enough one cheap enough to just run myself?" For a small business, "good enough, and yours" keeps getting more true every month.
Read more: how do I buy AI? →Source: DeepSeek API pricing China Daily, May 2026
AI-written code ships measurably more vulnerabilities
Researchers put numbers on it this year. On 20,000+ real GitHub issues, an LLM writing the fixes introduced roughly nine times as many new vulnerabilities as the human developers solving the same issues — and agentic setups produced more the more autonomy they were given. A separate study of 7,200 real agent pull requests found hard-coded credentials and injection flaws among the critical findings.
The nuance matters: on small, focused fixes the agents did fine — the risk grows with the size of the change and the amount of rope you give it. Which is the old rule with new numbers: AI drafts, you read it, and the more autonomy you grant, the harder you look.
Read more: can I trust AI agents to work on their own? →Source: SWE-bench security study, arXiv ACM: security issues in agent pull requests
