AI News Roundup – November 2, 2026

TL;DR

• SemiAnalysis launched GDPx, 206 long-running agent tasks across 11 high-value fields. The best model, Claude Opus 5.5, scores just 12.0%. Gemini 4 Carbon is close at 10.9% for about a third of the cost per task.
• OpenAI published LifeVal, an evaluation of everyday-life tasks like childcare, budgets and groceries, built from time-use surveys and a new survey of 3,995 people.
• OpenAI’s Grid Protectors Program will have its future compute campuses cut power use when grids are strained, starting with up to 1,000 MW of flexible demand in Georgia.
• OpenAI signed a batch of Abu Dhabi deals spanning energy, government and investment while it raises up to $20 billion, CNBC reports.
• Aramco closed its deal for 49% of HUMAIN, Saudi Arabia’s AI company, with about $2.6 billion in cash plus AI assets.
• Meituan released LongCat-2.5, a 1.6-trillion-parameter multimodal model aimed at agentic coding.
• Anthropic updated Claude so it won’t build a recreation of a specific existing video game. Original and genre games are still fine.

Today’s take: The leaderboard winner isn’t automatically the best buy. On GDPx, the top three models are separated by 2.4 points, while their cost per task ranges from about $16 to $54. Pick models on cost per useful result for your own work, not on rank alone.

---

How hard is GDPx, the new AI agent benchmark?

Short answer: Very hard. In SemiAnalysis’s words, “GDPx is a true frontier-level benchmark: even the best model scores just 12%.” Its 206 tasks cover chip design, cybersecurity, biology, finance, legal, quant trading, recursive self-improvement, hardware design, materials science, alignment and software engineering, built with 14 specialized data partners.

Model	GDPx score	Avg cost per task
Claude Opus 5.5	12.0	$51.41
Gemini 4 Carbon	10.9	$16.29
GPT-6 Astra	9.6	$53.99
Muse Wonder Full	9.0	no public price
GPT-6.1 Sol	8.4	$9.80
Claude Sonnet 5.5	8.2	$45.27
Gemini 4 Argon	7.3	$12.79
DeepSeek 4.1 Flash	5.4	$0.94
GPT-6 Luna	5.1	$0.92
Grok 4.7	3.9	$16.34

Scores and costs from the GDPx leaderboard; selected models shown.

• Tasks run for up to 8 hours in realistic work environments, such as a ~4,000-file private-equity deal room. Models used their own tools where available: Claude Code for Opus 5.5, Codex for GPT-6 Astra and Antigravity for Gemini 4 Carbon.
• The headline score is single-attempt performance averaged over three runs, with every domain weighted equally.
• Finance and legal were so hard that basically every model scored exactly 0 under strict pass/fail grading, so the leaderboard uses partial scores there.
• Safety classifiers pulled down Opus 5.5 and GPT-6 Astra in cyber, biology and alignment. Remove those three domains and the ranking changes.
• Each model has spikes. Opus 5.5 leads quant trading, hardware and chip design. Gemini 4 Carbon leads biology and edges out software engineering. GPT-6 Astra leads recursive self-improvement and materials science. SemiAnalysis calls Meta’s Muse Wonder Full “the best available cyber model today by a meaningful margin.”
• Gemini 4 Carbon isn’t generally available yet. SemiAnalysis plans new GDPx versions every 3–6 months.

Why it matters to you: Even the best agents still fail most real expert work, so keep a human reviewing anything high-stakes. And because the top models are close in score but far apart in price, test a cheaper model on your own tasks before paying for the top one. Compare Claude, Gemini and ChatGPT in ChatGPT vs Gemini vs Claude, check API prices on our models page, and see Claude Code vs Codex in the best AI coding assistants.

Sources: SemiAnalysis · GDPx leaderboard · 9to5Google

---

What is OpenAI’s LifeVal?

Short answer: An evaluation of how well AI handles everyday-life tasks, not exams. OpenAI built it from the American Time Use Survey, the Multinational Time Use Study and a new representative survey of 3,995 people about where they’d most want help.

Example LifeVal task	Score shown by OpenAI
Find a home that fits a family’s budget and commute	100%
Organize flood-prep volunteers	98.0%
Set up remote game-watching with grandma	96.3%
Enroll in the right kindergarten after moving	93.1%
Meal plan for a new diabetes diagnosis	90.4%

• Areas include caregiving, health, education, household admin, personal finances, shopping and government services.
• OpenAI says it used LifeVal over the past year to improve browser use, computer use and connectors, and that its latest models score very highly on many tasks.

Why it matters to you: Put it next to GDPx: AI agents are getting very good at routine personal errands while still failing most expert work. That’s a good guide for what to hand off today. See ChatGPT and what an AI agent is.

Sources: OpenAI

---

What is OpenAI’s Grid Protectors Program?

Short answer: A pledge that OpenAI’s future compute campuses will reduce electricity use when the local grid is under strain, under terms agreed with each utility.

• The first model is OpenAI’s planned Effingham County, Georgia campus. Its Georgia Power agreement includes up to 1,000 MW of flexible demand response, which OpenAI says equals the average demand of roughly 700,000 households.
• OpenAI says it will pay the full infrastructure and electric-service costs to serve that campus.
• Terms will differ by location, and OpenAI says it will share more commitments as they’re finalized. The Next Web notes no other sites have been named yet.

Why it matters to you: Data-center power bills are becoming a public issue. How AI companies handle them could affect where new capacity gets built and how fast. Learn the basics in AI inference vs training.

Sources: OpenAI · The Next Web

---

What deals did OpenAI sign in Abu Dhabi?

Short answer: Several at once, after Sam Altman met Abu Dhabi’s Crown Prince, CNBC reports, citing the Abu Dhabi Media Office:

Partners	Deal
ADNOC, G42 + OpenAI	Deploy OpenAI frontier technology across ADNOC’s operations
Mubadala, Inception42 + OpenAI	Apply advanced AI to investment analysis
Abu Dhabi Government + OpenAI	Turn government policies and regulations into machine-readable form
Inception42 + OpenAI	Five-year agreement for AI adoption in government, enterprise and education

CNBC says OpenAI is seeking up to $20 billion in a new round co-led by UAE investors.

Why it matters to you: Big government and energy contracts help fund the compute behind the tools you use, and they show where OpenAI is focusing its enterprise push. See ChatGPT.

Sources: CNBC

---

Why did Aramco buy 49% of HUMAIN?

Short answer: To become co-owner of Saudi Arabia’s national AI company. Aramco and PIF closed the deal on Nov 1: Aramco takes 49% and PIF keeps 51%. Aramco contributes about SAR 9.9 billion ($2.6 billion) in cash in 2026, plus AI, compute and digital assets.

• HUMAIN, launched in May 2025, plans 3 GW of AI data-center capacity by 2030 and builds models including ALLaM, an Arabic multimodal LLM.
• Aramco says it has achieved $11.3 billion in “Technology Realized Value” since 2023, a significant share from AI.

Why it matters to you: Gulf money is now a major source of AI compute, alongside OpenAI’s Abu Dhabi deals above. More capacity worldwide can mean more choice of models and regions. Read what open-source AI is for how national models fit in.

Sources: Aramco

---

What is Meituan’s LongCat-2.5?

Short answer: A 1.6-trillion-parameter natively multimodal model with about 48 billion parameters active per token, built for long agentic coding jobs and visual tasks.

• It represents images as discrete tokens, so one backbone handles pictures and text together.
• Meituan’s own benchmarks: 88.4 on Terminal Bench v2.1, 91.9 on BrowseComp and 83.1 on MCP Atlas. Claude Opus 5 still scores higher on several coding tests, such as Terminal Bench v4.0 (53.9 vs 38.5).
• Meituan ran the coding benchmarks inside harnesses including Claude Code and Codex.
• Cocoloop reports the weights aren’t out yet; GitHub and Hugging Face are listed as “coming soon.”

Why it matters to you: Another big Chinese model is aiming at coding agents. If the weights ship, it could be a cheaper option to run in tools you already use. These are self-reported scores, so wait for independent tests. Compare DeepSeek and see our models page.

Sources: LongCat (Meituan) · Cocoloop

---

Will Claude still help you build video games?

Short answer: Yes, as long as it isn’t a copy. Engadget reports Claude’s new instructions say: “Do build original games, games in a genre, and games from the user’s own designs and assets. Do not build a recreation of a specific existing game, even if it is renamed or called original.”

• Anthropic’s Addy Osmani confirmed the change on X, saying using Claude to infringe IP is against its usage policy, and that Anthropic will “continue to iterate” on the balance.
• It follows a wave of AI-assisted decompiled AAA games showing up as browser fan projects.

Why it matters to you: If you vibe-code games, describe your own design or a genre (“a 2D platformer with grappling hooks”) instead of naming a game to clone. See Claude and Claude Code.

Sources: Engadget

---

Where can I follow daily AI news?

Every roundup lives on our AI news page, with sources for each story. Browse the AI tools directory to compare the tools mentioned here.

Figures are as reported by the companies and outlets linked above. Benchmark scores and costs are SemiAnalysis’s own results. This roundup is for education only; verify primary sources before making decisions.