X AI Daily¶
Translation
English translation of the Chinese original.
1. AI Models, Products, and Applications¶
OpenAI clarifies the three ChatGPT Pro usage tiers and publicly asks whether to add a Pro 50¶
What happened: OpenAI's Tibo (@thsottiaux), who oversees Codex and ChatGPT, posted that because he had seen a lot of confusion, he wanted to set things straight: none of the Pro plans have a 5-hour usage window, and usage relative to Plus is 10x for Pro 100, 20x for Pro 200, and 50x for Pro 500, with Pro 500 additionally including ultrafast. About half an hour later he posted again asking "whether there is any point in a Pro 50."
Why it matters: The two posts drew about 2.12 million and 1.33 million views respectively, making them the most widely spread official posts in the AI circle during this window. They spell out in one go the multipliers that had remained unclear since Pro usage was doubled in late October; the Pro 50 question points directly at the gap between the $20 Plus and the $100 Pro, and users in the Chinese-speaking community (such as @Bryan_Explorer) have said they would cancel Claude and switch over as soon as Pro 50 launches.
Key original X posts (all times below are Shanghai time): - @thsottiaux clarifies usage, 11-08 13:09, 8,225 likes, about 2.12 million views - @thsottiaux asks about Pro 50, 11-08 13:45, 9,886 likes, about 1.33 million views
Verification: OpenAI's Help Center note on usage in Work and Codex states that Pro 100, Pro 200, and Pro 500 currently have no 5-hour limit in Work and Codex, but still have an included total allowance (weekly limits may still apply); the Pro plan description confirms the three tiers cost $100, $200, and $500 per month respectively, and that only Pro 500 includes Ultrafast. The 10x, 20x, and 50x multipliers are not listed on the official help pages; the source at present is Tibo's own posts. Pro 50 is only a question, with no commitment that it will launch.
Assessment: This is a clarification, not a new round of capacity expansion. "No 5-hour limit" does not mean unlimited; heavy users still need to watch the weekly allowance. Whether Pro 50 materializes is the pricing signal worth watching next.
Google DeepMind's Pushmeet Kohli: scientific discovery is entering a "phase transition," with new releases in a few weeks¶
What happened: Pushmeet Kohli, who leads science at Google DeepMind, posted that across DeepMind's various projects they are seeing "a real phase transition in the speed of scientific discovery," and that releases in the coming weeks would let the outside world see part of it; he then added that the focus would be on helping the broader scientific community use these tools responsibly.
Why it matters: The main post drew about 640,000 views and more than 7,000 likes, and X News also carried it as a separate item. DeepMind has a track record of delivering on the AlphaFold line, so this kind of preview carries considerable weight in the community.
Key original X posts: - @pushmeet main post, 11-08 18:15, 7,198 likes, about 640,000 views - @pushmeet follow-up, 11-08 18:15, 831 likes, about 51,000 views
Verification: The posts come from Kohli's own account. They contain no specific project, field, data, or date; "phase transition" is currently only the claim of the person involved and cannot be verified.
Assessment: For now it can only be read as a preview. Once the actual releases land in the coming weeks, we can see how many testable results the "phase transition" can produce.
Garry Tan's open-source harness evaluation: same model, different shell — pass rates on 20 tasks can't be separated, but cost can¶
What happened: YC president Garry Tan published preliminary results from an open-source evaluation. The setup fixes the model (Opus 5.5 or GPT-6.1 Sol) and 20 public tasks, changing only the agent harness (system prompt, tools, memory), and compares pass rate, cost per solved task, and time taken.
Why it matters: Garry Tan's post drew about 54,000 views, and X News carried two related stories. The conclusion differs from the popular claim that "picking the right harness substantially raises success rates": on this batch of tasks pass rates could not be distinguished, and what actually separated the harnesses was cost and speed.
Key original X post: - @garrytan, 11-09 02:09, 345 likes, about 54,000 views
Verification: The results table in the garrytan/g-harness-evals repository shows: on Opus 5.5, pass rates for QM, Claude Code, Hermes Agent, and OpenClaw range from 72.5% to 78.8%, with the 95% intervals overlapping heavily; cost per solved task is about $0.22 for Claude Code and about $0.44 for OpenClaw. On GPT-6.1 Sol, pass rates for Codex, Hermes Agent, and OpenClaw range from 71.3% to 75.0%. The repository itself notes that these are development results on public tasks, not validation on a held-out set, that the cost differences are described only and not tested, and that two OpenClaw cells are incomplete. Conflicts of interest to be aware of: Garry Tan's disclosure file states that YC has invested in Hermes and that QM is a YC project.
Assessment: The most useful information is that "this task set is too easy to separate anything," which is also how the author reads it. Until it is rerun on harder tasks, it cannot show which harness is stronger — only that cost can differ by a factor of two for equivalent results.
Hot on X, facts unverified: suspected GPT-6.2 Sol gradual rollout (follow-up)¶
What happened: On 11-06 this daily reported speculation about the release timing of GPT-6.2 Astra and Sol; what is new this issue is user-side claims of being "routed." From the evening of 11-08, large numbers of users in both Chinese and English said that when they selected GPT-6.1 Sol they were routed to GPT-6.2 Sol, and posted outputs such as a Three.js 3D scene generated in one shot; a self-test also circulated in the Chinese-speaking community: turn off web access and ask the model whether it knows about Fable 5.
Why it matters: It spread very widely, and some used it to predict Sol launching Monday and Astra on Thursday. But this kind of "gradual rollout" claim appears before every major version release.
Key original X posts: - @kimmonismus, 11-08 20:40, 1,388 likes, about 78,000 views - @daniel_mac8, 11-08 21:33, 418 likes, about 47,000 views (release-date prediction, personal speculation) - @ai_funss, 11-08 22:45, 69 likes, about 30,000 views (Chinese)
Verification: As of RUN_AT, no post confirming a GPT-6.2 Sol gradual rollout or release was found on the accounts of OpenAI, Sam Altman, or Tibo. The "do you know about Fable 5" method can only show that the model's knowledge cutoff is relatively recent; it cannot prove a version number; claims such as "stronger than Opus 5.5" and "about half the price" all come from personal tests or predictions and are unverified.
Assessment: The heat is real; the facts are not yet. Go by OpenAI's official announcements.
Meta's Alexandr Wang previews a "release this week"¶
What happened: Meta's Alexandr Wang posted a short video with text to the effect of "before this week's release, first teach everyone's muse to dance," and earlier the same day had also posted something joking about muse.
Why it matters: The main post drew about 360,000 views. On 11-08 this daily reported that Meta was pushing Muse Wonder's memory capability downmarket; here the person involved is directly previewing another new release this week.
Key original X post: - @alexandr_wang, 11-08 11:00, 3,866 likes, about 360,000 views
Verification: From Wang's own account. The post does not state the product, form, or date of the release; it is a preview by the person involved, and the specifics are unverified.
Assessment: All that can be established is that Meta will have a Muse-related release this week; everything else waits for an official announcement.
2. AI Compute and Industry¶
No new compute, funding, or earnings news directly related to AI met the inclusion criteria in this window.
3. Chinese-Language AI Ecosystem¶
Rumor (follow-up): unreleased Qwen-4-Flash-Early tops a third-party game-development leaderboard¶
What happened: On 11-08 this daily reported speculation about whether Qoduck is Qwen 4; what is new this issue is data from a third-party leaderboard. GameFeel Bench, maintained by independent developer @follistatindev, updated to v1.4 and added two Qwen 4 checkpoints labeled "Early": Qwen-4-Flash-Early scores 49.9 overall, ranking first, above Claude Opus 5.5 (XHigh) at 47.2 and Sonnet 5.5 (Max) at 45.0; Qwen-4-27B-Early scores 44.6. The Chinese-speaking community then widely relayed this as "Qwen 4 crushes Opus 5.5."
Why it matters: The original post drew about 390,000 views, and the single most-viewed Chinese relay drew about 38,000. If the numbers hold, the most notable point is cost: the per-task cost listed on the leaderboard is about $0.40 for Qwen-4-Flash-Early and about $2.60 for Opus 5.5.
Key original X posts: - @follistatindev, 11-08 08:07, 864 likes, about 390,000 views - @NFT_Chen, 11-08 16:18, 324 likes, about 38,000 views (Chinese relay) - @MichaelLee009, 11-08 17:34, 58 likes, about 16,000 views (Chinese relay)
Verification: The GameFeel Benchmark website shows Qwen-4-Flash-Early 49.9, Opus 5.5 (XHigh) 47.2, Sonnet 5.5 (Max) 45.0, and Qwen-4-27B-Early 44.6, consistent with the posts. Three points need to be stated. First, this is a narrowly scoped benchmark maintained by an individual, testing only game development and creativity; the lead comes mainly from the creativity and game-feel sub-scores, and Opus is higher on the bug-fixing item. Second, the cost basis is not uniform: Qwen is billed at the tester's API rates, while Anthropic is converted from subscription pricing. Third, the codename "hawk" appears only in the post and not on the leaderboard page; Alibaba has not confirmed this checkpoint and has not announced a Qwen 4 release date.
Assessment: This is another signal that Qwen 4 is near, but "surpassing Opus 5.5" holds only on one narrow benchmark, and by less than 3 points; it cannot be extrapolated to general capability.