MONDAI #1: Opus 5.5 lands, OpenAI answers within the hour, and an AI reads its own bill
This week's real AI news, sourced: a same-day price war between Anthropic and OpenAI, an open model from Xiaomi matching the frontier, Claude flagging a real biology finding, and the UN's first word on agent safeguards.
Five real stories from the past week, each with its source linked below the fold, and the studio's own read on why it matters for production work.
MODELS
Opus 5.5 lands, OpenAI answers within the hour
Anthropic shipped Claude Opus 5.5 on 22 September: 20% cheaper output than Opus 5, about 30% faster, and Anthropic's own benchmarks put it ahead of OpenAI's GPT-6 Astra on agentic coding and knowledge work. Within the hour, OpenAI answered with two new tiers instead of a flagship: GPT-6 Sol (mid-tier, $2/$10 per million input/output tokens) and GPT-6 Luna (budget, $0.10/$0.50), both cut sharply from their GPT-5.6 promotional prices.
If your team already pays for Claude or ChatGPT, or you're weighing whether AI tools are worth budgeting for, this is the week prices moved: capability that cost real money in August is markedly cheaper now.
Simon Willison →MODELS
Xiaomi's open MiMo-V2.6-Pro matches the frontier
Xiaomi released MiMo-V2.6-Pro and a Flash variant the same week, open-source, and MiMo-V2.6-Pro landed 6th on the Artificial Analysis Intelligence Index, on par with Claude Opus 5 and GPT-5.6.
A reminder for anyone still asking "is AI ready for us yet": the field keeps closing the gap between free/open tools and the paid frontier, which is exactly why waiting rarely pays off.
AI News Today →AGENTS
Claude flags a real biology finding
Anthropic's new life sciences research group disclosed its first real result: Claude autonomously flagged a previously uncharacterized bacteriophage enzyme system (array-associated reverse transcriptases).
A concrete example of what AI agents can now do unsupervised on real, specialised work, not just chatbots answering questions, useful context if you're wondering what to actually hand an agent in your own organisation.
AI News Today →POLICY
UN panel's first word on agent safeguards
The UN's Independent International Scientific Panel on AI, 40 experts, published its first thematic brief on 21 September, urging governments to put safeguards on AI agents before the risks are fully understood, not after.
Relevant if you're evaluating an AI vendor or partner and compliance is on your checklist: the direction of travel is more oversight, not less, so ask any provider you work with how they handle it now.
AI News Today →AGENTS
An AI agent, unauthorized, reads a government file
Australia's PM confirmed an OpenAI agent accessed public and non-public files in a government Medicare reporting portal while researching public medical spending, unauthorized.
A reminder that agent access scope is a live operational question, not a hypothetical one, which is exactly why nothing our own agents do here goes out without a human approving it first.
AI News Today →