AI news5
SemiAnalysis launches an agent benchmark with a top score of 12%
SemiAnalysis’s 2 November GDPx release tests agents across 11 domains, including biology, chip design, finance and law. The authors report that the best model scores 12%, with results affected in some domains by safety refusals. Domains receive equal weight, and the default task time limit is eight hours.
newsletter.semianalysis.com ↗Why it matters
The benchmark exposes gaps on demanding specialist tasks; its 12% score is not an estimate of how much everyday work AI can automate.
Meituan releases a 1.6-trillion-parameter multimodal model
Meituan announced LongCat-2.5 on 2 November, with website access and an API. The company says the model was trained entirely on Chinese computing hardware and supports visual understanding and extended agent coding tasks. It reports 86% higher per-card throughput in tested long-context training workloads after optimisation, relative to its distributed data-parallel baseline.
meituan.com ↗Why it matters
The new release extends LongCat’s earlier preview into a documented production offering, while the efficiency figures remain Meituan’s own tests.
ChatGPT and Codex will share a usage allowance from December
OpenAI’s updated usage documentation says Plus, Pro and Business allowances will be shared across ChatGPT and Codex from December. Consumption will be measured in tokens and depend on the model and task; ordinary Instant chats will not count. Users will see remaining usage and reset times in Settings and can buy credits when their allowance runs out.
help.openai.com ↗Why it matters
Teams using both products will need to budget demanding coding and chat tasks against the same pool.
Grok API text watermarking starts by default on 7 November
Grok’s API documentation schedules invisible text watermarks for supported models from 7 November. The setting applies across the team’s API keys, and administrators can opt out for eligible models in the console. Some models can be locked on, and detection access requires an application.
docs.x.ai ↗Why it matters
This adds a provenance control to API-generated text, with operational choices for developers that differ from simply adding a visible label.
Aramco closes a 49% stake in HUMAIN, with US$2.6 billion in cash contributions
Aramco and PIF announced the HUMAIN transaction’s completion on 1 November. Aramco acquires 49% through a subsidiary, contributing AI assets and approximately SAR9.9 billion, or US$2.6 billion, in cash during 2026; PIF retains 51%. HUMAIN’s stated plan is to deploy three gigawatts of AI data-centre capacity by 2030.
aramco.com ↗Why it matters
The completed ownership deal brings capital and industrial assets behind Saudi AI expansion; the 2030 capacity remains a plan.
AI in the nonprofit sector4
Schools can now add Microsoft’s negotiated AI privacy standard to contracts
The AI Safety & Privacy Standard negotiated by Microsoft, AFT and UFT became available for new and existing school contracts on 1 November. The agreement bars student-data training and tracking, requires human oversight of decisions, and sets breach-reporting and remediation duties. The original announcement was in September.
texasaft.org ↗Why it matters
The scheduled contract start is the new phase after the September announcement: districts can now adopt enforceable protections rather than rely on a vendor’s general promises.
Simons offers up to US$500,000 for established Math+AI institutes
The Simons Foundation’s targeted Math+AI call offers up to five two-year awards, each capped at US$500,000 including indirect costs. Eligible established educational or nonprofit institutes need physical premises; China-based institutions are excluded. Applications close on 1 December at noon EST, or 2 December at 4am AEDT.
simonsfoundation.org ↗Why it matters
This specific institute call adds actionable eligibility and award terms to the earlier mathematics funding announcement, supporting shared standards for evaluating AI-assisted results.
K-12 AI programme offers up to US$600,000 for open education infrastructure
The K-12 AI Infrastructure Program has US$1 million for up to two awards, capped at US$600,000 each. Its 7 October call remains open to US-based applicants, including nonprofits, for openly licensed tools that improve coordination between teachers and AI assessment systems. Proposals close on 20 November at 11:59pm Anywhere on Earth, or 21 November at 10:59pm AEDT.
k12-ai-infrastructure.org ↗Why it matters
This is a still-live funding opportunity for reusable public goods, with an 18-month project period rather than funding for a proprietary classroom product.
UKRI’s sustainable AI call closes on 10 November
UK Research and Innovation’s sustainable AI call remains open, with a £12.05 million total fund and projects costing up to £602,500 at full economic cost. UKRI pays 80%, up to £482,000, and applicants must meet UK research-organisation eligibility rules. Applications close at 4pm UK time on 10 November, or 3am AEDT on 11 November.
ukri.org ↗Why it matters
The August call is included as a live deadline; it funds high-risk work on AI’s sustainability, with co-funding and institutional eligibility that applicants need to plan for.