All editions
AI News6 items

AI News, Tuesday 27 October 2026

1. OpenAI ships GPT-6.1 Terra, a cheaper model that uses your computer.

Try · OpenAI released GPT-6.1 Terra, its mid-sized model, on 26 October in ChatGPT Work, Codex and the API (gpt-6.1-terra). It focuses on computer use: it reads the screen and acts in apps, for example to fill out forms or edit videos. API prices fall to $1 per million input tokens and $5 per million output tokens, half of GPT-5.6 Terra. On OpenAI’s own tests, it scores 50.3% on Terminal-Bench 4.0, against 21.5% for the old Terra. Why it matters: a cheaper model that clicks through apps fits the back-office automations you build for clients. Try it in Codex on one repetitive desktop task. Source · Source 2

2. Hob, a free AI desktop app that runs offline.

Try · Hob is a free, open-source desktop app from Frigade, shown on Hacker News on 26 October as an on-device alternative to Claude Cowork and ChatGPT. It picks the best open-weight model for your machine, such as Gemma or gpt-oss, and sets it up for you. After setup it works offline and can fill out PDFs. It runs on Apple silicon Macs, Windows and Linux; with about 160 stars, it is very young. Why it matters: for clients who refuse to send documents to a cloud AI, it shows local AI in a workshop for free. Download it from hob.chat and test it on one PDF form. Source

3. Anthropic files to go public: $16.3 billion in 6 months.

Know · Anthropic filed its IPO prospectus, the S-1, with the US SEC on 26 October, to list on Nasdaq under the ticker ANT. Revenue reached $16.3 billion in the first half of 2026, against $1.16 billion a year earlier. The net loss reads $80.7 billion, but $72.4 billion of it is an accounting charge on convertible notes; operating cash flow was positive at $3.9 billion. Nearly half of sales go through the Amazon, Google and Microsoft cloud marketplaces. Fortune reports a target debut in Thanksgiving week. Why it matters: these are the first official numbers on how fast companies buy Claude, a strong data point for your client pitches. Source · Source 2

4. Microsoft offers buyouts to 60,000 employees.

Know · Microsoft offered voluntary buyouts on 26 October to about 60,000 employees, more than a quarter of its 223,000 staff. It covers the US, UK, Canada and India, in teams such as Copilot, finance, HR, Xbox and Windows. Fortune reports a US minimum of 8 weeks of pay and departures on 31 December. Reuters links the move to cost control while Microsoft funds huge AI spending, but a Microsoft spokesperson told KUOW that AI is not driving it. Why it matters: your clients’ HR teams will ask what this says about AI and jobs, and the facts are more mixed than the headlines. Source · Source 2 · Source 3

5. OpenAI claims its AI disproved a core belief of computer science.

Know · OpenAI said on 26 October that its internal models produced 3 results in math, computer science and physics, with proofs checked in Lean, a language that verifies proofs by computer. The boldest claims to disprove the Exponential Time Hypothesis; the paper and Lean code are public on GitHub. The other two cover a 1991 conjecture in harmonic analysis and quantum chromodynamics, the theory of the strong nuclear force. A group led by Russell Impagliazzo, Ramamohan Paturi and Ryan Williams will review the work; no expert has confirmed it yet, and early Hacker News comments are mostly skeptical. Why it matters: if it holds, AI has overturned a belief many proofs rely on, but wait for the review before quoting it. Source · Source 2

6. Cloudflare’s AI service leaked other users’ data in 155 answers.

Know · Cloudflare disclosed on 26 October that its Workers AI service mixed other customers’ data into 155 answers from the GLM-5.3 model, between 27 August and 9 September. The model accepted 1,048,576 tokens, but its KV cache only had room for 681,280; a bug in the open-source SGLang server then let longer requests read leftover data from other users. Cloudflare resized the cache, patched the engine and found no sign of deliberate attack. Why it matters: it is a concrete case for data-isolation clauses with shared AI hosting, useful in your governance workshops. If you self-host models with SGLang, upgrade to v0.5.22 or later. Source

Glossary

  • Computer use: an AI model that sees the screen and controls the mouse and keyboard to work in apps like a person.
  • Terminal-Bench: a benchmark that scores AI models on real tasks done in a command-line terminal.
  • Open-weight model: a model whose trained parameters are published, so anyone can download it and run it on their own machine.
  • S-1: the registration document a company files with the US securities regulator before selling shares to the public.
  • Convertible notes: loans that investors can turn into company shares later; when the share value rises, the accounts record a loss on paper.
  • Exponential Time Hypothesis: a widely held belief that some logic puzzles cannot be solved much faster than by trying almost every possible answer.
  • Workers AI: Cloudflare’s service for running AI models on its shared servers through an API.
  • KV cache: memory where a model keeps intermediate results for the text it has already read, so it does not recompute them.