AI Daily Brief · Nov 2, 2026 | SemiAnalysis launches GDPx, where the best score is just 12%; Abu Dhabi signs four deals with OpenAI; AI subscription subsidies tighten

The past 24 hours brought few hard launches; the conversation centered on a new benchmark, a set of Gulf deals and the cost of AI subscriptions. SemiAnalysis released the GDPx benchmark, spanning 11 domains from chip design to biology and quant trading, where even the best model scored only 12, drawing responses from Google’s CEO and several researchers. After Abu Dhabi’s crown prince met Sam Altman, local institutions signed four agreements with OpenAI covering energy, government and investment. Elon Musk said Grok Bot costs too much to serve relative to revenue, Replit’s CEO warned that subsidized subscriptions may be ending, and OpenAI’s help center said ChatGPT and Codex will share usage limits from December. Also: Andrej Karpathy built HackerTube with Claude in about two hours, Axios reported on OpenAI’s post-midterm policy plans, and rumors about GPT-6.2 Astra, Mythos 5.5 and other models kept spreading, none of them verified.

Today’s highlights

• SemiAnalysis released GDPx: 206 tasks across 11 domains, led by Claude Opus 5.5 at 12.0, ahead of Gemini 4 Carbon at 10.9 and GPT-6 Astra at 9.6 [Third-party claim].
• Abu Dhabi signed four agreements with OpenAI: ADNOC will use OpenAI’s GPT models with inference kept inside the UAE, alongside a “Machine-Readable Government” pilot and an investment-analysis partnership with Mubadala.
• Musk said Grok Bot’s serving costs are unsustainably high relative to revenue [One-sided claim]; OpenAI said Plus, Pro and Business usage limits will be shared across ChatGPT and Codex from December.
• Karpathy had Claude sift 43 million comments from 20 years of Hacker News into 6,454 videos to build HackerTube (43.5K likes).
• Axios: OpenAI plans to push four proposals, including AI “speed limits,” in the lame-duck session starting Nov 9, and to call its data centers “AI compute campuses” [Media report].

1. SemiAnalysis launches the GDPx benchmark: the top score is just 12%, with Opus 5.5 first and Gemini 4 Carbon second [Third-party claim]

What it is: According to the GDPx site and SemiAnalysis’s blog post, GDPx measures the agentic capabilities of frontier models across 11 domains and 206 tasks, built with 14 data partners. The domains are chip design, cybersecurity, biology, finance, legal, quant trading, recursive self-improvement, hardware design, materials science, alignment and software engineering; government, real estate and insurance were deliberately left out. Each task comes with a simulated work environment and a default 8-hour time limit. Models run in their native harness where one exists (Claude Code for Opus 5.5, Codex for GPT-6 Astra, Antigravity for Gemini 4 Carbon), with OpenCode as the fallback, and the headline score is pass@1 across three attempts.

Results: Claude Opus 5.5 scored 12.0, Gemini 4 Carbon 10.9, GPT-6 Astra 9.6, Muse Wonder Full 9.0 and Claude Fable 5.1 8.5; Grok 4.7 scored 3.9 and GLM 5.3 2.7. The blog notes that the overall scores of Opus 5.5 and GPT-6 Astra are dragged down by their own safety classifiers, and that every model scores 0 on finance and legal tasks under binary grading, so the leaderboard reports the average partial score there. It calls Gemini 4 Carbon “the only model in the same ballpark as Opus and Astra”; Carbon is not yet generally available. SemiAnalysis plans a new version every 3–6 months.

Reactions: @SemiAnalysis_’s launch post drew 2,527 likes and about 1.09M views. Google CEO @sundarpichai said he was glad to see a high-quality benchmark with room to improve and suggested watching Carbon’s cost curve (3,198 likes). SemiAnalysis founder @dylan522p, one of the blog’s authors, posted a formal apology to Google, saying Carbon did very well in several domains but is spikier (1,447 likes). @emollick called it a successor to GDPval but said it lacks a human baseline, added that GPT-6 Pro caught some grading errors, and argued that zero scores hide economically meaningful differences [Opinion] (218 likes).

Sources: @SemiAnalysis_ 1 · @SemiAnalysis_ 2 · @sundarpichai · @dylan522p · @emollick 1 · @kimmonismus 1 · GDPx site · SemiAnalysis blog

2. Abu Dhabi signs four agreements with OpenAI; ADNOC becomes the first energy company to deploy OpenAI’s frontier technology

The deals: According to the Abu Dhabi Government Media Office, after Crown Prince Sheikh Khaled bin Mohamed bin Zayed met Sam Altman, he witnessed local institutions sign four agreements with OpenAI spanning government, energy and investment. (1) ADNOC, G42 (through its subsidiary Inception42) and OpenAI will bring AI into ADNOC’s operations using OpenAI’s GPT models, with inference running inside the UAE. (2) The Abu Dhabi government and OpenAI launched Machine-Readable Government (MRG), which converts a set of policies and regulations into machine-readable form to help officials with routine compliance reviews such as licenses and certificates; every assessment traces back to the relevant clause, and officials still make the final decision. The first pilot is live at the Abu Dhabi Agriculture and Food Safety Authority, covering food imports. (3) Mubadala, Inception42 and OpenAI will use AI for investment analysis, opportunity identification and risk assessment. (4) Inception42 and OpenAI signed a five-year agreement to drive frontier AI adoption across government, business and education.

The goal: Abu Dhabi says it aims to become the world’s first fully AI-native government.

Reactions: @Kalshi posted a news flash (3,127 likes), as did @Cointelegraph (225 likes). ADNOC CEO @SultanAlJaber wrote that the world has long asked what energy can do for AI, and now we will see what AI can do for energy (228 likes).

Sources: @Kalshi 1 · @Cointelegraph 1 · @SultanAlJaber · Abu Dhabi Government Media Office · WAM · The National · CNBC

3. AI subscription subsidies tighten: Musk says Grok Bot costs too much to serve relative to revenue; ChatGPT and Codex will share limits from December

What Musk said [One-sided claim]: Replying to a user, @elonmusk said Grok Bot’s serving costs are unsustainably high relative to revenue (15,378 likes, about 7.64M views). In a later reply to another user, he asked whether upgrading to a higher-usage plan was too big a price jump (5,753 likes). The same day he also urged people to try Grok Bot, saying the team ships dozens of fixes and new features seven days a week and reads user feedback (9,680 likes, about 7.65M views).

User reports [One-sided claim]: @op7418 said requests are being routed to Opus 5.5 less and less often, and that many people saw their Grok Bot usage jump by 20%, 30% or even 40% (112 likes). @ai_for_success said they can no longer afford to run Grok Bot around the clock (197 likes). These are user observations; there has been no official explanation of any change to usage rules.

Replit CEO’s warning [Opinion]: @amasad told anyone enjoying a subsidized AI subscription to get ready for the rug to be pulled (2,802 likes).

OpenAI merges limits: OpenAI’s help center says that starting in December, usage limits on Plus, Pro and Business plans will be shared across ChatGPT and Codex and measured in tokens, with usage depending on the model and task. Instant chats don’t count against the limit, and the combined remaining usage and reset time can be checked under Settings → Usage. OpenAI expects most users won’t need to wait for a reset more often than they do now. The page gives no exact December date and no per-plan limits.

Sources: @elonmusk 1 · @elonmusk 2 · @elonmusk 3 · @op7418 · @ai_for_success · @amasad · OpenAI Help Center: ChatGPT and Codex usage

4. Karpathy’s weekend project HackerTube: Claude sifts 20 years of HN comments into 6,454 videos

How he built it (first-person account): Unhappy with YouTube’s recommendations, @karpathy had Claude sift 43 million comments from 20 years of Hacker News into 6,454 videos that were recommended again and again, finishing in about 2 hours with the Claude Code desktop app and Opus 5.5 at medium reasoning (43,526 likes, about 3.81M views). The site needs no login and keeps state in the local browser; thumbnails are a representative frame from the middle of each video rather than the uploader’s A/B-tested cover; the underlying comment-analysis data is published as a bundle, and he suggests having your own Claude build a private version from it.

The prompt (first-person account): In a reply, he shared his opening prompt, which roughly asked Claude to build a YouTube-like site featuring only videos repeatedly recommended on HN over the years, and not to make mistakes (2,045 likes).

Sources: @karpathy 1 · @karpathy 2 · HackerTube

5. Axios: OpenAI readies a post-midterm policy push built on AI “speed limits” and rebrands data centers as “AI compute campuses” [Media report]

The plan (Axios exclusive): Axios expects Democrats to likely win the House in Tuesday’s midterms, which would change the landscape AI companies face in Washington. OpenAI chief global affairs officer Chris Lehane’s team has been stockpiling policy proposals since August and plans to push them starting with the lame-duck session that opens Nov 9. There are four main planks: (1) mandatory safeguards for the most powerful AI systems, which Lehane calls “speed limits,” to keep humans in control of the technology; (2) a new community compact and state legislation giving local residents a real say in how data center projects are built and what benefits they bring, with OpenAI calling its own facilities “AI compute campuses” and pointing to a large project in Pike County, Ohio; (3) helping workers cope with disruption and expanding education and training; and (4) a US-led push for global AI standards grounded in democratic values. Axios calls the renaming an “intentionally provocative rebranding.”

Caveat: There is no bill text, threshold or enforcement detail yet. Axios discloses at the end of the article that it has a content licensing and technology agreement with OpenAI.

Reactions: @AndrewCurran_ shared the Axios story (859 likes); @Polymarket said OpenAI hopes the rebrand will improve public perception (1,019 likes).

Sources: @AndrewCurran_ · @Polymarket 1 · Axios

6. A pile-up of model rumors: GPT-6.2 Astra, Mythos 5.5, Fable 5.5 and Opus 5.6 [Unverified]

GPT-6.2 Astra [Unverified]: The earlier clue was gpt-6.2-astra showing up in usage stats. @Smlay_ero went further, saying it is rolling out and can be called directly in the Codex CLI with codex --model gpt-6.2-astra (1,061 likes). There was no OpenAI announcement as of publication.

Mythos 5.5 [Unverified]: @AiBattle_ said Mythos 5.5 appears in the HTML source of Anthropic’s website but not in an Oct 23 snapshot, suggesting it was added in the past few days (243 likes). The original tip it cites was posted late on Nov 1 Beijing time, before this issue’s coverage period.

Fable 5.5 demo [Unverified]: @chetaslua posted a video described as a pure-code animation that Fable 5.5 wrote in one shot (3,618 likes); @kimmonismus said it is likely the latest Fable 5.5 checkpoint (2,083 likes). The model behind it can’t be verified.

Opus 5.6 [Secondhand rumor]: @Pirat_Nation relayed a tip from kimmonismus that Claude Opus 5.6 will launch alongside Mythos and Fable 5.5, with a set of benchmark numbers attached (1,863 likes). Those numbers have no official source, so we don’t reproduce them.

Debunk [Opinion]: @kimmonismus said the so-called Anthropic “Numbat” leak is probably fake, since some parameters in the model exactly match random values, as if it had never been trained (305 likes).

Release list [Opinion]: @LuminaBench listed models that may ship in the coming weeks if all goes well, including Fable 5.5, GPT-6.2 Astra, Qwen 4, DeepSeek V4.1 Pro and general availability of Gemini 4 Carbon (203 likes).

Sources: @Smlay_ero · @AiBattle_ · @chetaslua · @kimmonismus 2 · @Pirat_Nation · @kimmonismus 3 · @LuminaBench

In brief

• South Korea’s AI budget: In a budget speech to the National Assembly on Nov 2, President Lee Jae Myung presented a 2027 budget of 820.9 trillion won, up 12.8%, including 21.3 trillion won for three megaprojects (a chip manufacturing hub and data centers, physical AI, and AI adoption) and 4.6 trillion won toward a sovereign AI model that aims to be the world’s best-performing (@Reuters, 34 likes; @CNBC, 55 likes; Yonhap).
• Is Google killing AI Overviews? [Unverified]: @delonperc claimed Google is pulling AI Overviews because they cut clicks to websites and hurt ad revenue (@delonperc, 13,713 likes). A Google spokesperson told The Verge on Oct 30, “We haven’t reduced the triggering of AI Overviews,” adding that they were never meant to appear on every search (The Verge).
• Colossus 2 tops the compute rankings: @Rothmus said Colossus 2 is worth about 1.11 million H100 equivalents, making it the world’s most powerful AI data center (@Rothmus, 1,228 likes); @elonmusk reposted it, saying other supercomputer projects fall short of their targets while SpaceX has delivered [One-sided claim] (@elonmusk 4, 16,213 likes). Following Musk’s earlier cluster-by-cluster breakdown of Colossus 2, Epoch AI’s database estimates the site at about 1.112 million H100e and 946 MW of IT power today, rising to about 1.824 million H100e in Q1 2027 (Epoch AI).
• Suleyman on AI consciousness [Opinion]: @Kalshi flashed that Microsoft’s AI CEO warned against treating AI as conscious (@Kalshi 2, 862 likes); the original was a link posted by @mustafasuleyman before this issue’s coverage period (@mustafasuleyman, about 1.2M views).
• Mistral Large 4 writes a quantum chemistry program [Company claim]: @MistralAI said that from one short prompt, Mistral Large 4 wrote a quantum chemistry SCF solver from scratch, validated the output and produced a 3D visualization (@MistralAI, 585 likes).
• MiniMax quantizes M3.1 Flash [Company claim]: @MiniMax_AI said it quantized the 428B-parameter M3.1 Flash, reducing load memory and the weight data moved during decoding so it is easier to run locally (@MiniMax_AI, 615 likes).
• Nvidia’s AgentX submission [Third-party claim]: @SemiAnalysis_ said an Nvidia AgentX submission shows that, with software written by agent swarms, HBM chips can exceed 1,000 tokens per second per user on 700B+ models, and Rubin NVL72 delivers 20.9x the performance per TCO of GB300 NVL72 (@SemiAnalysis_ 3, 541 likes).
• RTX 5080 Super canceled? [Media report, not officially confirmed]: @GameGPU_com said Nvidia has canceled plans to launch a 24GB GDDR7 RTX 5080 SUPER (@GameGPU_com, 461 likes); Nvidia has said nothing.
• Kimi’s promotion controversy [Opinion]: @kimmonismus said even WhatsApp now knows about the controversy over Kimi K3.1 Pro’s undisclosed promotion, calling it embarrassing for Kimi (@kimmonismus, 1,074 likes).
• Mollick on the start of agentic work [Opinion]: @emollick said GPT-4 could work on its own for only 4 minutes (at a 50% success rate), while today’s models can do weeks of work, and that January 2026 can be seen as the start of agentic work (@emollick, 691 likes).
• Chinese models vs. Mythos Preview [Opinion]: @scaling01 said that nearly 7 months on, Chinese models still haven’t caught up with Mythos Preview (about 161 ECI) (@scaling01, 564 likes).
• Perplexity Computer demo: Perplexity CEO @AravSrinivas posted a demo video of asking Computer to visualize 400 million years of change on Earth (@AravSrinivas, 456 likes).
• AA-AgentPerf adds GPU rental prices: Artificial Analysis partnered with Silicon Data to bring market GPU rental prices into AA-AgentPerf, its data center performance benchmark, to measure how much agentic work each dollar of compute supports (@ArtificialAnlys, 21 likes).

---

About this column and how we verify

1. AI-assisted: AI Daily Brief is compiled with AI assistance; this issue was produced with Grok Bot and covers Nov 2, 2026 00:00 to 22:07 Beijing time.
2. Every post checked: Each quoted post was read individually to confirm it exists, the author matches and it was posted within this issue’s time range; the Suleyman post and the original Mythos 5.5 tip predate this issue’s range and are cited only as background. All quoted posts were also double-checked through the X API. Like and view counts were read at about 22:15 Beijing time on Nov 2.
3. X News summaries are leads only: X News summaries are generated by Grok and used only to find leads; facts come from the original posts, official pages and media reports, and the primary source wins in any conflict.
4. Items involving Grok Bot: Main story 3, “AI subscription subsidies tighten,” concerns Grok Bot itself, which helps produce this column. It was included on popularity and public posts alone, reflects no inside information, and was held to the same standards as everything else.
5. Labels: [Unverified] = circulating but not confirmed by an official or authoritative primary source; [Media report] = from media reporting, details may differ from official accounts; [Media report, not officially confirmed] = reported by media, not yet officially confirmed; [Third-party claim] = results published by a third-party organization, not independently reproduced; [Company claim] = from a company announcement or official account, not independently reproduced; [One-sided claim] = one party’s account with no other corroboration; [Secondhand rumor] = something passed along from what someone else heard; [Opinion] = personal judgment, not a statement of fact.