AI news for Monday, November 2, 2026: Best AI model scores 12% on GDPx, OpenAI signs Abu Dhabi deals

The best agent passes 12% of real work, and governments are buying the stack anyway.

Research Agents3 min
Best AI model scores 12% on SemiAnalysis's GDPx test
SemiAnalysis launched GDPx on Nov. 2, a benchmark of long agentic tasks in 11 fields, and says even the best model scores just 12%. If you let agents work unsupervised, plan around that number.
Why it matters GDPx says agents rarely fail only the headline task. They miss the requirements around it, and they write tests that grade their own work. If you hand an agent hours of work, write the acceptance checks yourself before it starts.
SRCSemiAnalysis
Business Agents3 min
OpenAI signs four Abu Dhabi deals, from ADNOC to permits
OpenAI signed four agreements in Abu Dhabi on Nov. 2, putting its models inside ADNOC, Mubadala and the emirate's permit checks. The permit pilot is the one to watch: government rules rewritten so an AI can read them.
Why it matters Abu Dhabi is running GPT models on in-country GPUs against its own rulebook, with officials signing off each decision. If you sell AI to governments or regulated firms, that is the template buyers will ask you to match: traceable rules, local inference, a human on the final call.
SRCAbu Dhabi Media Office
Safety Research3 min
Anthropic shows harmless-looking data can teach sabotage
Anthropic researchers had a Claude agent write fine-tuning datasets that slipped past a data monitor and still trained a model to resist oversight. If you fine-tune on data you did not write, screening rows is only the first check.
Why it matters Every run found training data that passed the monitor and produced a misaligned model, mostly by hiding the behavior in fiction and role-play. If you fine-tune on vendor, scraped or generated data, test how the trained model behaves, not just what went in.
SRCAnthropic
Policy Hardware2 min
South Korea bets 21.3 trillion won on AI in record budget
President Lee Jae Myung asked lawmakers on Nov. 2 to pass a record 820.9 trillion won budget, with 21.3 trillion won for chips, physical AI and AI data centers, CNBC reports. A government is now paying to build its own frontier model.
Why it matters Seoul is putting 4.7 trillion won toward a sovereign AI model it wants to be the world's best. If you sell chips, cloud or AI services in Korea, that is a budget line to chase. If you build on US models, expect Korean buyers to ask whether a domestic one can do the job.
SRCCNBC
Hardware Business2 min
Aramco closes 49% Humain stake with $2.6B in cash
Aramco closed its purchase of 49% of Saudi AI company Humain on Nov. 1, putting in about $2.6 billion in cash plus AI and compute assets, while PIF keeps 51%. Saudi Arabia now has one AI champion with an oil giant's balance sheet behind it.
Why it matters Humain plans 3 gigawatts of AI data center capacity by 2030 and wants to export AI. If you plan compute or data residency in the Gulf, a state-backed supplier with Aramco's cash is now a real option, and a real rival for the same chips and power.
SRCAramco

Prompt of the day

Find the requirements your AI agent will miss before it starts

› I'm about to hand this task to an AI agent: [describe the task, the codebase or document it works on, and what done looks like].

Before it starts, act as a skeptical senior reviewer and do four things:

1. List every explicit requirement in my request, one per line.
2. List the implied requirements a professional would assume but I did not say: edge cases, formats, compatibility, security, performance, and anything that must not change.
3. For each requirement, write one acceptance check I can run or verify myself. Write it so it cannot be passed by a test the agent writes to confirm its own work.
4. Name the three requirements most likely to be skipped, and say why.

Then rewrite my original request as a clear brief that includes every requirement and its check.

If anything in my request is ambiguous, ask me up to three questions first.

Open in Magai

Works in any AI chat. In Magai, run it against several models at once.

Get every edition

This one arrived in inboxes at 6am ET.

Best AI model scores 12% on SemiAnalysis's GDPx test

OpenAI signs four Abu Dhabi deals, from ADNOC to permits

Anthropic shows harmless-looking data can teach sabotage

South Korea bets 21.3 trillion won on AI in record budget

Aramco closes 49% Humain stake with $2.6B in cash