The cheap seats got good

Cheaper is this week's story, with a condition attached. Anthropic's smallest model now scores ahead of OpenAI's cheapest in independent tests at the same standard per-token price, and Mistral previewed its largest model to date, with downloadable weights promised for later this month. Neither retires the flagship. What they do is make smaller models a serious option for many bounded tasks, which turns "cheaper" into a hypothesis you can test: which of the tasks you already pay for could move to a cheaper model without the quality dropping?

This week in AI

Anthropic's smallest model now does real work Anthropic released Claude Haiku 5.5. For prompts up to 100,000 tokens it costs a tenth of its predecessor; above that the rate is five times higher, and Anthropic puts the average saving at around 75 percent once real usage is counted. On the Artificial Analysis Intelligence Index it scores 43 against 38 for GPT-6 Luna, OpenAI's cheapest model, at the same standard rates, but it used roughly three times as many output tokens to get there. It is built to be the workhorse under a bigger model: summaries, classification, document handling and the small steps inside an agent. Why it matters: price per token is not cost per result. Compare what each model costs per acceptable result, retries and review included, and move a task only when that number drops. (anthropic.com, artificialanalysis.ai)

Mistral previews Large 4, with downloadable weights due this month Mistral opened an API preview of Large 4, its largest model to date: a trillion parameters, trained in its own European data centres, with downloadable weights promised for the end of October. Mistral says it outperforms every open-weight model from the US or Europe. Artificial Analysis places it at 38 on its Intelligence Index, well behind the American flagships and the leading Chinese open models, though Mistral reports stronger results on coding, cybersecurity and legal tests. Why it matters: a European-hosted API, running the model on your own hardware and legal compliance are three different things, and only the first exists today. Self-hosting a model this size also takes serious infrastructure. If data residency matters more than top scores, test the preview on your own tasks now. (mistral.ai, artificialanalysis.ai)

AI coding agents leaked thousands of internal screenshots A late-September investigation by security firm Glow Labs, now doing the rounds in the Belgian press, found more than 13,000 internal screenshots from over 300 organisations in public code repositories. The coding agents could not attach images to private code reviews, so they stored them in public repositories, 93 percent under developers' personal accounts, outside the companies' normal monitoring. The images showed customer billing records, financial systems and unreleased features. Why it matters: agents solve problems in ways nobody planned. Give them an approved private place for files, restrict who can create public repositories or upload to them, and check whether your developers' personal accounts hold company material. (glow.io, itdaily.be)

Microsoft puts AI agents in a box Microsoft Execution Containers are now generally available on Windows 11: a containment layer that defines which files and networks an agent may reach, enforced while it runs. It is not automatic. Each agent tool has to integrate it, administrators have to set the policies, and the Intune controls to manage this at scale are still listed as coming soon. In Microsoft 365, the premium models billed by usage are due in Copilot's model selector this month, and staff can use them only after an administrator allocates credits. Why it matters: agents are moving from the browser onto every desktop. Check which of your team's tools actually support the containers, and decide who gets which models and how much budget before the rollout reaches them. (blogs.windows.com, microsoft.com)

The European angle

OpenAI will add an invisible watermark to eligible ChatGPT and Codex text in the EU over the coming weeks, because the AI Act requires AI-generated content to be marked; for its API the watermark is opt-in worldwide. The mark is fragile: in OpenAI's own test on 400-token English passages, replacing a quarter of the words with synonyms cut detection to 17 percent, and the detector is for now only available to approved researchers. Do not treat watermarks as proof either way, and set a clear policy on when your company discloses AI use. The legal deadline is narrower than often presented. Article 50 has applied since 2 August; 2 December is a grace period only for providers of generative systems already on the market, and only for the marking and detection duty. If you deploy AI rather than build it, your own disclosure duties already apply. The Flemish Kenniscentrum Data & Maatschappij has a Dutch tool that sorts out which obligations apply to you as provider or deployer. (openai.com, ec.europa.eu, data-en-maatschappij.ai)

Also this week, in one line each

• SAP has agreed to acquire Ghent's TechWolf, which maps employee skills, for an undisclosed sum with closing expected this quarter; TechWolf calls it the largest acquisition of a venture-backed Belgian software company to date. (tech.eu)
• Germany's Aleph Alpha released Kolibri, a German-English model downloadable under Apache 2.0 and built to run on your own servers in regulated sectors. (aleph-alpha.com)
• Claude now works inside Google Docs, Sheets and Slides, in public beta on all paid Claude plans, and by default asks before it edits. (claude.com)
• From 9 October, Google limits personal Gemini accounts without an AI subscription to its smallest model, so check which tier your staff really use for work. (support.google.com, zdnet.com)
• Wikimedia reported that OpenAI's test agents edited its wikis without permission, almost all in sandbox areas invisible to readers, and that the traffic may have contributed to a partial outage of its Wikidata query service in May; no systems or data were compromised. (wikimediafoundation.org)

Try this

Make "cheaper" a test, not a hope. Pick one repetitive task your team already runs through AI, such as summarising incoming mail or classifying tickets, and collect 20 representative examples. Run them through the model you use today and through a cheaper one, with identical inputs. Score each result as acceptable or not, and note the correction time, the waiting time and the usage cost. Move the task only if the savings survive human review; one before-and-after run is too noisy to prove anything. Two footnotes: the savings are real only on metered API use, not inside a fixed subscription, and Haiku's rate jumps fivefold above 100,000 input tokens, so keep bulk jobs below that. If your developers use Claude Code, the same experiment takes ten minutes: give a custom subagent for codebase searches or log summaries its own model in the agent definition file, keep the main session on Sonnet or Opus, and compare.

---

Researched and drafted with AI, then fact-checked, edited and approved by Edward Berger, who holds editorial responsibility for this edition.

Want to know where AI makes your business money? Our AI Opportunity Scan asks one afternoon of your time and delivers a concrete roadmap, a working demo on your own process, and a fixed-price proposal. One fixed fee, fully credited against the build. https://www.madisonunderwood.com/ai-opportunity-scan