Morning edition — Monday 2 November 2026

Meituan's LongCat team has released LongCat-2.5, a native multimodal model for long-horizon agent tasks. It has 1.6 trillion total parameters, with about 48 billion active per token, and builds on LongCat-2.0 with stronger visual understanding and agentic coding.

The team says it trained on more long-horizon trajectories, used large-scale reinforcement learning on complex codebases and full-project builds, and turned real agent failures into verifiable training tasks. Images are represented as multi-level discrete tokens, so text and vision share one token format.

The LongCat API changelog lists the release on 2 November and says the model works with agent harnesses such as Claude Code, Codex and Hermes. The Economic Observer reports that training ran entirely on domestic Chinese compute, and that Meituan's own results place it between Claude Opus 4.8 and Claude Opus 5 on several benchmarks.

Paying users of Grok Bot, the always-on agent from Elon Musk's xAI that is also offered on some Cursor plans, reported on Sunday that their weekly usage meters jumped toward 100% without matching activity. A Cursor forum bug thread opened at 04:25 UTC on 1 November had more than 100 posts by Monday morning; users describe readings such as 34% to 73% overnight and agents refusing work once the meter hit the limit.

At 20:11 UTC an X account asked Musk for a usage reset. He replied 23 minutes later: "The cost of serving Grok Bot was unsustainably high relative to revenue." He did not say whether limits had been cut, whether the jump was a metering fault, or whether anyone would get usage back.

MadRobot, a tech blog, reports that a Cursor staff member called the jump something that "shouldn't happen" and that xAI's status page logged a billing and usage dashboard incident on Sunday morning. This journal could not confirm the staff post or the status entry in the copies it read.

Linus Torvalds released Linux 7.4-rc1 on Sunday and called it "a *BIG* merge window", with almost 20,000 commits against the usual 15,000 or so. He wrote that he is "*much* happier with the flood of AI triggered work going into the merge window than it showing up mainly as extra fixes during the -rc period".

He described the release as "full of fixes for issues found by AI (and more and more the patches have been suggested by AI too, although they often get rewritten)". He hopes this means AI fixes are becoming part of "normal development" rather than piling into the stabilisation phase, and called that "a much healthier model".

Torvalds gave two other reasons for the size: developers expect 7.4 to be the next long-term support release, and a third of the diff is AMD GPU register headers.

The research firm SemiAnalysis has published GDPx, a benchmark of agentic tasks in domains it expects to earn the first trillion dollars of AI revenue, including recursive self-improvement, chip and hardware design, cyber, biology, finance, legal, quant trading, software engineering, materials science and alignment.

It says "even the best model scores just 12%". In SemiAnalysis's results, Claude Opus 5.5 is the best all-round model; Gemini 4 Carbon is the best biology model and narrowly wins software engineering; GPT-6 Astra wins recursive self-improvement and materials science. It says the finance and legal tasks are so hard that nearly every model scores zero.

Tasks were built with 14 specialised data partners, mostly reinforcement-learning environment start-ups. SemiAnalysis argues that OpenAI's GDPval, released in September 2025, predates agents and no longer measures frontier models well.

Abu Dhabi's Crown Prince, Sheikh Khaled bin Mohamed bin Zayed, met Sam Altman and witnessed the signing of four agreements between Abu Dhabi entities and OpenAI, the Abu Dhabi Media Office said on 2 November.

ADNOC, G42's subsidiary Inception42 and OpenAI agreed to deploy OpenAI models across ADNOC's operations, using in-country inference. The Abu Dhabi government launched a "Machine-Readable Government" initiative that turns selected rules into machine-readable form so AI can help officials check permit and certificate compliance; its first pilot is live with the food-safety authority for food imports, and officials keep the final decision.

Mubadala, Inception42 and OpenAI signed a deal on AI for investment analysis and risk assessment, and Inception42 and OpenAI signed a five-year agreement to support AI adoption in UAE government, business and education. No financial terms were given.

President Lee Jae Myung told the National Assembly on Monday that the government will put 21.3 trillion won (about $16 billion), "nearly twice this year's amount", into AI and three megaprojects: semiconductors, physical AI and AI data centres. The speech opens parliament's review of the 820.9 trillion won 2027 budget, a record 12.8% rise first proposed in September; the deadline for passage is 2 December.

Seoul Economic Daily lists 4.7 trillion won for frontier AI development aimed at putting homegrown models among the world's best, 2 trillion won to expand chip production capacity, 2.1 trillion won for power, water and industrial-site infrastructure, and 500 billion won for physical-AI trials in eight sectors including manufacturing and defence. Research and development spending would reach a record 39.5 trillion won.

Lee tied the plan to a chip-led export boom, saying exports of $1 trillion a year are within reach, and called for a shift from "growth for the few" to "growth for all".

Energy North, a start-up planning "Project Ares", a 1-gigawatt data centre powered by gas and solar on a cattle station in Australia's Northern Territory, has scrapped its "first power" date of 28 June 2028, Reuters reported in an exclusive on 1 November.

The Territory's environment regulator suspended its seven-week deadline on 7 October to seek federal advice, and gave no new timeframe. Its ruling warned of ten potential impacts, from water systems to Aboriginal sacred sites, and expressed a "low level of confidence" in the proposed safeguards. "We're no longer working to it," Energy North's Daniel Hunter told Reuters, saying timing would follow once the regulator sets terms for an environmental impact statement.

Reuters calls it the biggest Australian project to put back its timing; it lists a delayed NextDC expansion in Melbourne and a project Goodman pulled in Sydney.