OpenAI Signs Four Abu Dhabi Deals as a New Agentic Benchmark Tops Out at 12% and ChatGPT and Codex Move to One Shared Allowance
Monday’s AI news follows the money and the meters. OpenAI signed four agreements in Abu Dhabi, including a deal that makes ADNOC the first energy company to run OpenAI Frontier agents, while Aramco closed its $2.6 billion purchase of 49% of Saudi AI company HUMAIN. SemiAnalysis launched GDPx, a benchmark across 11 high-value domains where the best model scores just 12%, and OpenAI said ChatGPT and Codex will share one token-based allowance from December. South Korea’s chip exports jumped 259% as President Lee proposed 21.3 trillion won for AI, chips and data centers, and Meta reportedly built a humanoid prototype with China’s Fourier.
OpenAI Signs Four Agreements in Abu Dhabi, Making ADNOC the First Energy Company on OpenAI Frontier
Sam Altman’s visit to the UAE produced a cluster of deals on Monday. The Abu Dhabi Media Office statement on the Crown Prince’s meeting with Sam Altman says Sheikh Khaled bin Mohamed bin Zayed met the OpenAI chief executive and then witnessed the signing of four agreements between Abu Dhabi entities and OpenAI, spanning government, energy and investment. ADNOC, G42 (through its subsidiary Inception42) and OpenAI signed an energy deal; the Abu Dhabi Government and OpenAI launched what they call the world’s first AI-native “Machine-Readable Government” initiative; Mubadala, Inception42 and OpenAI agreed to apply AI to investment analysis, opportunity identification and risk assessment; and Inception42 and OpenAI signed a five-year agreement to drive frontier AI adoption across government, enterprise and education in the UAE.
The government initiative is the only piece described as already live. It translates a defined set of policies and regulations into machine-readable form so that AI systems can help officials with routine compliance assessments, such as those behind permits and certificates. Each assessment is traceable to the rule that governs it, and officials keep oversight and accountability for final decisions. The first pilot is in production with the Abu Dhabi Agriculture and Food Safety Authority, focused on food imports, and runs on OpenAI’s UAE inference residency, with model computation on GPUs inside UAE data centres.
The energy agreement carries the most detail. ADNOC’s press release on its OpenAI Frontier partnership, announced at ADIPEC, describes a multi-year deal that makes ADNOC the first energy company to deploy OpenAI Frontier, the platform for coordinating AI agents across a business. It starts with two applications: field master planning that merges separate subsurface, facilities and economic studies into one integrated decision, which ADNOC says can produce complex development plans in hours rather than months, and a system of interconnected agents that continuously forecast crude supply, demand, prices and margins across production, refining, logistics, sales and trading. The system will run on G42’s sovereign infrastructure in the UAE, with human experts keeping oversight of consequential decisions. The deal also covers cyber defence through OpenAI’s Daybreak Blue initiative and workforce training, building on ADNOC’s ChatGPT Enterprise rollout to more than 40,000 employees. The Next Web’s report on the ADNOC deal notes that the companies did not disclose the value of the agreement or when the system will be rolled out.
“The next frontier of AI is not simply doing things faster. It is making possible what was previously beyond reach.” — Dr. Sultan Ahmed Al Jaber, ADNOC Managing Director and Group CEO
“Abu Dhabi and the UAE are setting a bold, global standard for AI integration across both the public and private sectors to make sure the benefits of AI reach everyone.” — Sam Altman, OpenAI CEO, via the Abu Dhabi Media Office
The Machine-Readable Government pilot is the template worth studying, more than the oil deal. Turning rules into structured, citable logic and having AI check cases against them, with every result traceable and a human signing off, is something any regulated business can copy for procurement, onboarding or compliance reviews. Start with one narrow rulebook, measure how often officers overturn the AI, and expand only once that rate is stable.
SemiAnalysis Launches GDPx, an 11-Domain Agentic Benchmark Where the Best Model Scores 12%
A new benchmark wants to replace OpenAI’s GDPval as the standard yardstick on model release cards. The official GDPx leaderboard site lists version 1.0 at 206 tasks across 11 domains, built with 14 specialized data partners. Claude Opus 5.5 leads with a GDPx score of 12.0, followed by Gemini 4 Carbon at 10.9, GPT-6 Astra at 9.6, Muse Wonder Full at 9.0 and Claude Fable 5.1 at 8.5. Grok 4.7 scores 3.9 and GLM 5.3 sits at the bottom of the main comparison at 2.7.
SemiAnalysis’s launch article on measuring frontier models across the most valuable domains, by Max Kan, Andrew Megalaa and Dylan Patel, explains the choice of domains: chip design, cybersecurity, biology, finance, legal, quant trading, recursive self-improvement, hardware design, materials science, alignment and software engineering. The authors argue that the first trillion dollars of AI revenue will be concentrated in sectors that can deploy frontier intelligence at high return, so they deliberately skip industries such as government, real estate and insurance. Each domain was built with specialist reinforcement-learning environment startups, and every task ships with a realistic environment. One finance task, for example, drops the model into a reconstructed private-equity deal of roughly 4,000 anonymized files, where it must find its own context the way a human associate would.
The results are spiky. SemiAnalysis says Opus 5.5 is the best all-around model and wins quant trading, hardware design and chip design; Gemini 4 Carbon is the best biology model and narrowly wins software engineering; and GPT-6 Astra wins recursive self-improvement and materials science. Finance and legal tasks are so hard that basically every model scores exactly zero. The authors also say the overall scores of Opus 5.5 and GPT-6 Astra are dragged down by safety classifiers in cyber, biology and alignment, and that Muse Wonder Full is the best available cyber model by a meaningful margin. Their critique of GDPval is pointed: its tasks are precisely specified and non-interactive, with context spoon-fed in the prompt, and its human expert inter-rater agreement was just 71% on the gold subset.
“GDPx is a true frontier-level benchmark: even the best model scores just 12%.” — SemiAnalysis
The useful lesson is not who is on top but how uneven the leaders are. A model that wins chip design can trail in biology, and safety classifiers can quietly block legitimate security or research work. Before standardizing on one vendor, run your own tasks in a realistic environment with your real files, and keep a second model approved for the domains where your first choice refuses or stumbles.
OpenAI Will Put ChatGPT and Codex on One Token-Based Allowance From December
OpenAI is changing how paid plans are metered. OpenAI’s Help Center article on understanding ChatGPT and Codex usage limits, updated over the weekend, says that starting in December the usage allowance for Plus, Pro and Business plans will be shared across ChatGPT and Codex, as the company brings the agentic capabilities of Work into the main ChatGPT experience. Usage for both will be measured in tokens, and how much of the allowance a request consumes will depend on the model chosen and the task. More advanced models and complex tasks generally use more, while Instant chat will not draw from the allowance at all.
Combined remaining usage and reset times will appear under Settings, then Usage. Users who hit the limit can wait for a reset, upgrade their plan or buy credits to keep using eligible features. The page gives no token totals per plan, no exact start date and no conversion between today’s message caps and the new token budget, so heavy users cannot yet tell whether their effective limits will rise or fall. What is clear is that coding agents and long chat sessions will now draw from the same pool, which means a long Codex run could leave less room for research in ChatGPT later that day.
“We expect most users won't need to wait for their usage to reset more often than they do now.” — OpenAI Help Center
Shared token pools reward teams that know what their work costs. Before December, have your heaviest ChatGPT and Codex users note which models they reach for by default, and move routine chat to Instant, which will not count. For Business seats, decide in advance who may buy credits and set a monthly ceiling, or a few agentic coding sessions could become an unplanned line item.
Aramco Closes Its $2.6 Billion Deal for 49% of Saudi AI Company HUMAIN
The Gulf’s AI buildout got a second big signal over the weekend. Aramco’s announcement that PIF and Aramco have closed the HUMAIN transaction, dated November 1, says Aramco, through a wholly owned subsidiary, now holds 49% of HUMAIN, with the Public Investment Fund keeping 51%. Aramco is transferring key AI assets and contributing about SAR 9.9 billion ($2.6 billion) in cash in 2026, plus AI, compute and digital assets. HUMAIN, launched in May 2025, is building data centers, cloud infrastructure, models including the Arabic large language model ALLaM, and applied AI solutions. It plans to deploy 3 gigawatts of AI data center capacity by 2030 and aims to become one of the leading exporters of AI globally.
Aramco frames the deal as a way to turn its existing digital and AI work into revenue and to run its own operations on HUMAIN’s infrastructure. Chief executive Amin Nasser said the company has achieved $11.3 billion in what it calls Technology Realized Value since 2023, with a significant portion coming from AI solutions. The closing lands on the same weekend that ADNOC signed with OpenAI, underlining how the region’s national energy companies are positioning themselves as both large AI customers and owners of AI infrastructure.
“The HUMAIN deal is very important to Aramco, building on the company’s long track record of innovation and technology.” — Amin H. Nasser, Aramco President and CEO
Oil producers now own a growing share of the world’s planned AI compute, backed by cheap energy and state balance sheets. For buyers, that means more regional capacity and possibly lower prices for inference outside the United States. It also raises questions about data residency, export controls and contract terms. If you are offered Gulf-hosted capacity, have legal review where data sits and which laws govern access.
South Korea’s Chip Exports Rise 259% as President Lee Proposes 21.3 Trillion Won for AI, Chips and Data Centers
The AI hardware boom is showing up in national accounts. South Korea’s trade ministry release on October exports says total exports rose 93.7% year over year to $115.28 billion, and semiconductor exports rose 259.4% to $56.44 billion, above $50 billion for a second straight month. The ministry attributes the strength to high unit prices and volumes as global big tech companies keep investing to lead in AI. Computer exports rose 496.6% to $5.85 billion on server infrastructure spending and demand for server SSDs, and electrical equipment exports hit an all-time monthly high of $1.99 billion on demand for transformers, wires and cables for AI data centers. Cumulative exports for January to October reached $930 billion, and the ministry expects annual exports to pass $1 trillion in November.
A day later, the government said it would push more money into the same industries. Yonhap’s report on President Lee Jae Myung’s budget speech says Lee told the National Assembly on Monday that the 2027 budget of 820.9 trillion won, up 12.8%, would boldly invest in AI and future growth engines, including 21.3 trillion won for a semiconductor manufacturing hub, data centers, physical AI and AI promotion under its “three megaprojects.” Yonhap puts a separate allocation for a sovereign AI model with the world’s highest performance at 4.6 trillion won, while CNBC’s report on South Korea’s sovereign AI spending plan gives 4.7 trillion won and says the 21.3 trillion won figure is nearly double the 2026 level. CNBC adds plans to train 2,000 specialists in semiconductor design and the regional chip industry.
“With exports maintaining solid growth and exceeding $110 billion again in October, annual exports are now set to reach $1 trillion in November.” — Trade Minister Kim Jung-kwan
Memory is now nearly half of what South Korea sells abroad, and Seoul is reinvesting the windfall into more AI capacity. That keeps memory suppliers firmly in pricing power. Companies planning server builds or large laptop refreshes should assume DRAM and SSD costs stay elevated through 2027, lock in supply agreements early, and treat any memory price dip as a buying window rather than a trend.
Meta Reportedly Built a Humanoid Prototype Called Max With China’s Fourier in Three Months
Meta’s robotics push has a body, according to a weekend report. The Next Web’s summary of The Information’s report on Meta’s humanoid robot says Meta built a prototype humanoid with Shanghai-based Fourier in about three months. Known internally as Max, the robot weighs about 48 kilograms, stands about 1.5 meters tall, runs on NVIDIA’s Jetson Thor chip and has a screen that mimics facial expressions, built-in speakers and five-fingered hands for picking up small objects. Li-Chen Miller, Meta’s vice president of product for robotics, reportedly spent almost two months at Fourier’s factory, and Meta showed Max to some staff at an internal robotics summit in late August. Meta told The Information the work is for internal research and experimentation and that it has no plans to sell the units.
Work on a second version began in September. Engadget’s report on Meta’s household-chore robot prototype says Max 2 is due to get a redesigned torso and hands, a projector and additional sensors to help it handle stairs, and that there is no launch date for a mass-market version. The Next Web reports that Meta sees humanoids as a physical form for Muse, its AI agent, with the hope that they can eventually do household tasks. Neither outlet independently verified the report, and Meta has not published anything about Max itself.
Three months from requirements to a walking prototype shows how far off-the-shelf humanoid platforms and edge AI chips have come, and how much of that supply chain runs through China. For companies exploring robotics, the practical path is the same one Meta took: partner with an existing hardware maker, customize, and focus your own effort on the software and the tasks. Budget early for supplier and compliance risk.
What We Left Out
Posts on X claimed that a hidden reference to a “Claude Mythos 5.5” appeared briefly on an Anthropic web page, and separate posts claimed a “GPT-6.2 Astra” name surfaced in OpenAI software. Neither company has announced such a model. When we checked on Monday morning, Anthropic’s Mythos page referred only to Mythos 5.1, and we found no primary source for either name, so we treat both as unverified. Canada’s planned equity investment of up to C$1.6 billion in Cohere, which circulated widely this weekend, was announced on October 29 and falls outside today’s window.
Model-name leaks travel fast and rarely change anything you should do this week. Our rule is simple: plan around shipped models with published pricing, limits and documentation, not around rumors. If your roadmap depends on a capability you have only seen in a leaked string, write down the fallback plan using models you can buy today, and revisit it when a real announcement arrives.
Need help navigating AI for your business?
Our team turns these developments into actionable strategy.
Contact SEN-X →