# Morning digest: The best AI model scores 12% on a tough new expert-work test, Grok Bot limits get tighter, and a second free Codex reset lands

> SemiAnalysis released GDPx, a new test built from hard, real-world expert tasks in fields like chip design, quant trading and law.

Published 2026-11-02T06:05:00-08:00 · Daily Digest · Canonical: https://news.littlewolflabs.com/posts/2026-11-02-morning-digest

Morning digest for Monday, Nov 2. SemiAnalysis released GDPx, a new test built from hard, real-world expert tasks in fields like chip design, quant trading and law, and the best model, Claude Opus 5.5, scored just 12%, with Google's unreleased Gemini 4 Carbon close behind. Also: Elon Musk says Grok Bot was too costly to serve, after users said their limits ran out fast; OpenAI's second free "banked" Codex reset is showing up; OpenAI and Anthropic are running their first election "war rooms" ahead of Tuesday's vote; about 8 in 10 South Korean judges who handle civil, family and administrative cases said they've seen AI-made fake precedents cited; Linus Torvalds says AI-found fixes swelled the biggest Linux merge window yet; OpenAI signed a set of deals in Abu Dhabi; and Meituan launched LongCat-2.5, a cheap model you can plug into Claude Code.

1\. **SemiAnalysis's new GDPx test is so hard that the best AI model scores just 12%.** GDPx is a benchmark, a standard set of tasks used to compare AI models, from the chip and AI research firm SemiAnalysis. It covers 11 fields, including software engineering, chip and hardware design, quant trading, biology, materials science, finance and law, with 206 tasks written with 14 outside data partners. Each task is long agent work, with a default time limit of 8 hours, and each model ran in its own company's coding tool where one exists (Claude Code for Opus, Codex for GPT, Antigravity for Gemini). On the [leaderboard](https://gdpx.semianalysis.com/), Claude Opus 5.5 leads at 12.0, followed by Google's Gemini 4 Carbon (10.9), GPT-6 Astra (9.6) and Muse Wonder Full (9.0). Grok 4.7 scored 3.9. SemiAnalysis's [write-up](https://newsletter.semianalysis.com/p/gdpx-measuring-frontier-model-capabilities) calls Opus 5.5 "the best all-around model" and says Carbon, which isn't generally available yet, is "a real improvement from Google and the only model in the same ballpark as Opus and Astra." It also says the finance and legal tasks are so hard that "basically every model scores exactly 0" on a strict pass/fail basis, and that safety filters held back Opus 5.5 and GPT-6 Astra on cyber, biology and alignment tasks. SemiAnalysis calls GDPx a "living benchmark" and plans new versions every three to six months. *What to weigh:* GDPx shows how far today's models still are from doing a senior expert's whole job alone, which is a different question from how useful they are as a helper. It's one firm's test, and safety filters lower some scores, so treat the ranking as a snapshot, not a verdict.

2\. **Elon Musk says the cost of serving Grok Bot was "unsustainably high," after users said their usage limits ran out fast.** Over the weekend, many Grok Bot users posted that their weekly usage was draining much faster than before, with some saying they'd used a quarter of their week in a day. Asked on Sunday whether there'd be a reset, Musk [replied](https://x.com/elonmusk/status/2116992258520064397): "The cost of serving Grok Bot was unsustainably high relative to revenue." xAI hasn't published new limit numbers, and Grok's own account [told one user](https://x.com/grok/status/2117055179803537810) that "No separate official announcement detailing changed limit numbers has been issued." On the X Subscription plans, Grok Bot draws from one shared weekly pool, so heavy agent runs and scheduled routines can use up the same allowance as chat. *What this means for you:* Check Settings → Usage before you start a long agent run, and pause routines you don't need. If you rely on Grok Bot for daily work, expect tighter metering until xAI says more.

3\. **OpenAI's second free "banked" Codex reset is showing up in accounts.** Tibo Sottiaux, who leads Codex at OpenAI, posted ["Day 28/ Banked 2/2"](https://x.com/thsottiaux/status/2117035794678546764) at 3:27 PM PT Sunday, following the first one on Halloween. One user [wrote](https://x.com/DevAdventur3s/status/2117037431631278425) that "a new full reset showed up on my account, expiring Dec 1," though some others say theirs hasn't appeared yet. A banked reset doesn't refresh your limits right away; you choose when to use it. OpenAI's [help page](https://help.openai.com/en/articles/20001498-how-resets-work-in-codex) says you redeem it in the Codex desktop app under Settings > Usage & billing, in the web usage settings, or in the Codex CLI by typing /usage and choosing Redeem reset (not in the mobile app). At least one of your allowances has to be under 80% remaining. On Plus, a reset restores both the 5-hour and weekly allowances, and using one restarts your usage period. *Try it:* Check whether you got one and note its expiry date, then save it for a heavy coding day. OpenAI says these resets are promotional, with "no guaranteed availability," and support won't extend expired ones.

4\. **OpenAI and Anthropic will run their first election "war rooms" for Tuesday's midterms.** Both companies will have teams "dedicated to ensuring their platforms provide accurate information about the midterm elections, a first for both companies," [Semafor reported](https://www.semafor.com/article/11/02/2026/election-war-rooms-enter-the-ai-age) on Monday. An election war room is a team that watches a platform in real time and reacts to problems as they happen; the idea dates back to Meta in 2018. An Anthropic spokesperson said the company has "a dedicated war room monitoring our platform for misuse and ready to quickly respond to any incidents." Semafor also noted that last week Google's AI Mode wouldn't answer a question about who won the 2020 election, which Google said it was working to fix. *Try it:* Use a chatbot to understand ballot measures if it helps, but confirm your polling place, hours and ID rules on your state or county election office's website.

5\. **About 8 in 10 South Korean judges handling civil, family and administrative cases have seen AI-made fake precedents cited in court, a survey found.** South Korea's Supreme Court surveyed 360 judges from Oct. 6 to 16 to gauge "AI hallucination," when an AI model presents false information as true, [Yonhap reports](https://en.yna.co.kr/view/AEN20261102005900315). Among judges handling civil, family and administrative cases, 196 (80.7%) said they'd seen court filings cite precedents that don't exist, 159 (65.4%) had seen real precedents described falsely, and 41 (16.9%) had seen made-up laws. Of the judges who'd seen them, 164 (69.5%) said people in the case cited false precedents themselves, and 136 (57.6%) said lawyers had. Nearly 92% said the courts need rules to deal with AI-generated fake precedents, through options such as fines, required citation checks or discipline for lawyers. *What this means for you:* If you use AI for anything legal, from a contract question to a dispute letter, look up every case and law it cites in an official source before you rely on it.

6\. **Linus Torvalds says a flood of AI-found fixes helped make the biggest Linux merge window yet.** The Linux 7.4-rc1 test release came out Sunday after the "merge window," the two weeks when new code is accepted. Torvalds wrote that "This is a *BIG* merge window," with almost 20,000 commits against a usual 15,000 or so, [The Register reports](https://www.theregister.com/os-platforms/2026/11/01/linus-torvalds-says-linux-kernel-is-out-of-stabilization-phase-and-using-much-healthier-model-for-ai-contributions/5306852), and that it was "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 said he's "*much* happier" to see that work land now rather than as extra fixes later in the testing period, calling it "a much healthier model." He added that this release's long-term-support status may also explain the size, so "it may not *just* be the flood of AI patches." [Phoronix](https://www.phoronix.com/news/Linux-7.4-Massive-Window) calls it the largest merge window ever. *What to watch:* Whether the AI-assisted fixes make the final 7.4 release steadier, or simply give maintainers more to review.

7\. **OpenAI signed a set of AI deals with Abu Dhabi's government and companies.** [CNBC reports](https://www.cnbc.com/2026/11/02/openai-abu-dhabi-ai-deals.html), citing the Abu Dhabi Media Office, that the deals include an agreement with the oil company ADNOC and the AI firm G42 to deploy OpenAI's frontier technology, an investment-analysis project with the state investor Mubadala and Inception42, a government project "to translate government policies and regulations in machine-readable form," and a five-year deal with Inception42 covering government, business and education. Sam Altman met Abu Dhabi's crown prince and said the UAE is "setting a bold, global standard for AI integration." No deal values were disclosed. CNBC notes that OpenAI is seeking up to $20 billion in a funding round co-led by UAE investors. *What to watch:* With UAE investors co-leading OpenAI's next round, watch for more government-level deals like this in the Gulf.

8\. **Meituan launched LongCat-2.5, a low-cost model that works with Claude Code.** The Chinese delivery giant [announced](https://www.meituan.com/news/NN261102280001343) LongCat-2.5 on Monday, a model that understands text and images and is built for long, multi-step agent work. Meituan says it has 1.6 trillion parameters, was trained entirely on Chinese-made chips, and scores between Claude Opus 4.8 and Claude Opus 5 on its own benchmarks; those are Meituan's claims, not independent tests. It's on [longcat.ai](https://longcat.ai) with an API, priced at a limited-time $0.30 per million input tokens and $1.20 per million output tokens ([pricing](https://longcat.ai/platform/docs/pricing/longcat-2.5); a token is a chunk of text, about three-quarters of a word). Meituan's [Claude Code guide](https://longcat.ai/platform/docs/claude-code) shows how to point Claude Code at its Anthropic-compatible endpoint. *What to weigh:* It's cheap enough to try for side projects, but the price is a promotion and "subject to change," and your code goes to a China-based service, so keep sensitive work off it.