X AI Daily¶
Translation
English translation of the Chinese original.
1. AI Models, Products, and Applications¶
OpenAI uses an unreleased "cyber model" to run the whole path from vulnerability discovery to patching: 948 Linux kernel privilege-escalation vulnerabilities, a $1.2 million API bill¶
What happened: OpenAI's official account said its security research team built an autonomous pipeline for vulnerability discovery, exploitation, and patching; in a single run, it used a not-yet-released cyber model to validate 948 local privilege-escalation vulnerabilities in the Linux kernel and generate patches, at a total API cost of $1.2 million. Converting with the figures the post itself reports, that is about $1,266 per vulnerability.
Why it matters: This is one of the few public cases where the entire chain of "find vulnerabilities — validate — write patches" is handed to a model to run in bulk, with a bill that can be priced. With 414K views and 3,891 likes, it is the most widely spread non-product-launch official post in the window. It sends two signals at once: security research is turning into batch work billed by the token; and OpenAI has an undisclosed cyber-specific model in hand.
Primary X post: @OpenAI announcement post (11-05 07:13, 414K views, 3,891 likes).
Verification: The publisher is OpenAI's official account; "a single run", "948", and "$1.2 million" are all the company's own claims, with no third-party review attached. Whether the vulnerabilities really exist, whether the patches were merged upstream, and the name and availability scope of the "upcoming cyber model" are all unverified.
Assessment: What is worth recording is the cost magnitude and the shape of the process, not the number 948. The official wording is "validated and generated patches", which is not the same as upstream adoption; and the results of an unreleased model on its own pipeline cannot be directly turned into a security product available to others.
Kimi K3.1 Pro releases its weights: one of the largest entries in the open-weight camp¶
What happened: Kimi's official account announced at 00:00 on November 5 that it was releasing the model weights of Kimi K3.1 Pro, with a caption of just one line: "continuing to advance open intelligence". Hours later, a license named "Kimi K3.1 Pro" appeared on Hugging Face; the relayed specs are about 2.8 trillion total parameters, 104 billion activated parameters, and a 1 million token context, officially described as "an open-weight, natively multimodal agentic model, the strongest 3T-class model to date". The same day, Vercel's AI Gateway listed two entries, kimi-k3.1-pro and a high-speed version, marked as zero data retention.
Why it matters: K3.1 Pro was first released as an API at the end of October; what is being given this time is downloadable weights, which is different in nature. The official post has 413K views and 6,229 likes, the most widely spread post on the open-weight topic that day; for teams that need private deployment or cost-sensitive agent workloads, this is one of the largest open-weight models currently available.
Primary X post: @Kimi_Moonshot announcement post (11-05 00:00, 413K views, 6,229 likes). Supplementary: weights already on Hugging Face (7,826 views, 132 likes), parameters and context (2,705 views, 46 likes), listed on AI Gateway (2,673 views, 42 likes).
Verification: Cross-checked against the Hugging Face model page: confirmed the official wording "open-weight, natively multimodal agentic model, the strongest 3T-class model to date" and the 1 million token context; the LICENSE text is a custom MIT-style license: use, modification, redistribution, and commercial use are permitted, but running a "model-as-a-service" business with more than $20 million in revenue over 12 months requires a separate agreement, and commercial products with more than 100 million monthly active users or more than $20 million in monthly revenue must prominently display "Kimi K3.1-Pro", with internal use exempt. The 2.8T total parameters and 104B activated parameters follow the relayed figures and Neural Brew's report; this time they were not checked item by item on the model card.
Assessment: The license clears self-hosting and the vast majority of commercial use; the threshold only catches cloud providers reselling the API at large scale. What to watch next is how fast the community can actually get weights of this size running.
Follow-up: Tencent Hunyuan releases the full Hy4 version, adding vision and computer operation; weights still not released¶
What happened: At 14:43 on November 4, Tencent Hunyuan's official account released the full Hy4 version, highlighting "newly added vision capabilities" and "firmly among the leading open models", listing scenarios such as finance, front-end, gaming, office, and research, and saying constraints were put in place for cybersecurity and biosecurity. The status recorded in the earlier 11-04 daily was "only a teaser; parameters and vision capabilities seen only in a leaked document", and this official post ended the teaser phase. Relayers add: Hy4 is a multimodal upgrade of the preview, can see the screen and operate a computer, scores above DeepSeek V4.1 1017; the weights are currently unavailable.
Why it matters: The official post has 481K views and 2,690 likes, and the same day there was also a shorter "Introducing Hy4" post with 43K views and 1,078 likes. On the same day as Kimi, one side released weights while the other released only an API, forming a contrast in how China's leading models use the word "open".
Primary X posts: @TencentHunyuan announcement post (11-04 14:43, 481K views, 2,690 likes), shorter launch post (11-04 15:50, 43K views, 1,078 likes). Outside observation: @TeksEdge (11-05 00:06, 14K views, 156 likes).
Verification: The official post is first-hand information from the party involved; "leading open model" is the vendor's own assessment. Pricing and benchmark details are covered by the Chinese-language relays in Section 3 and are all unverified. The point that the weights are not released is consistent with outside observation.
Assessment: Hy4's product form is clear this time: multimodality plus computer operation, aimed at office work and professional software. But "open model" currently means only an open API; the weights and license are still parked at "planning to open source in the near future".
Cognition releases SWE-2 Flash: a small-activation coding model whose weights are shared only with partners on a targeted basis¶
What happened: On November 5 at 03:55, Cognition released SWE-2 Flash, positioned as a fast and cheap coding model. The company says it beats all US open-weight models on major benchmarks, including ones three times its size; to balance transparency with cybersecurity risk, the weights are not published directly but instead go into a program of secure sharing with partners and researchers. Relayed specs: 276B total parameters, 12B activated, about 200 tokens/second, one-fifth the cost of SWE-2, three times faster, and free in Devin Desktop and Devin CLI for Pro/Max/Teams until November 30.
Why it matters: 171K views, 1,516 likes. It demonstrates two things at once: small-activation models can approach giant models on coding and terminal tasks; and the qualifier "US open-weight" shows the scope of the comparison was deliberately narrowed. Combined with weights being open only to vetted partners, this is a public trade-off by a frontier lab between transparency and misuse risk.
Primary X post: @cognition announcement post (11-05 03:55, 171K views, 1,516 likes). Usage-side relays: selectable directly inside Devin (3,808 views, 86 likes), comparison with Opus 5.5 plus the free scope (17K views, 121 likes).
Verification: The official post is from the party involved; the parameters, benchmark scores, and free scope come from relays, the original official blog is not in the sample, and the free scope explicitly excludes Devin cloud sessions (per the relay). Benchmarks are vendor-reported and unverified.
Assessment: What is worth recording is the engineering path of "small activation + strong RL", not the sentence "beats a model three times larger". The benchmark scope is limited to US open-weight models, and comparison with closed frontier models still needs third-party reproduction.
ChatGPT Images 2.6: adds GIF and reference images, two new models enter the API, generation 50% faster¶
What happened: On November 5 at 07:35, OpenAI announced that ChatGPT Images was upgraded to 2.6, adding GIF generation and reference images, and said two new models would enter the API with image generation 50% faster.
Why it matters: 564K views and 9,225 likes, the highest-engagement official post in this window, more than OpenAI's vulnerability research post the same day. The focus of iteration on the image product line lands on "controllable input" (reference images) and "animated output" (GIF), and both point directly at the production pipelines for e-commerce, advertising, and social assets.
Primary X post: @OpenAI announcement post (11-05 07:35, 564K views, 9,225 likes).
Verification: The official post is from the party involved, but it gives only three sentences of description. The names and pricing of the two new API models, the baseline for the 50% speedup, and the limits on GIF length and number of reference images are all unverified pending official documentation.
Assessment: This is a patch-level product update, not a model generation. What matters more for real workflows is reference-image consistency; GIF is more about filling a format gap for social output.
Anthropic uses Claude to solve the first 3D structure of a bacterial vault particle, and writes up Claude's mistakes along with it¶
What happened: On November 5 at 04:11, Anthropic said its biology research team, working with Claude, solved the first three-dimensional structure of the vault particle in bacterial life. Vault is one of the largest protein structures in cells, discovered in 1986 by Leonard Rome and Nancy Kedersha, and its function is still unknown; previously it was known only in complex cells such as those of animals and plants, and this work shows that similar proteins in bacteria and the viruses that infect them can also assemble into vaults. The company says manual analysis of such structures usually takes weeks to months, and this time the first model was completed in 34 hours. The official article also writes up several simple errors Claude made along the way, and how the team corrected them together with machine learning researchers.
Why it matters: 274K views, 2,877 likes. Life sciences is the field where "AI accelerates research" is easiest to turn into a slogan; the value of this post is that it gives a time comparison (weeks versus 34 hours) and does not hide the model's errors.
Primary X post: @AnthropicAI announcement post (11-05 04:11, 274K views, 2,877 likes).
Verification: Cross-checked against the Anthropic research article and the preprint PDF: what was solved is a vault assembled from phage protein L05, with 96 chains (48 per half, versus 78 in humans), a nominal cryo-EM resolution of 3.2 Å, and a volume about 25% larger than the human vault; purification went from about three days to about 90 minutes, the first structural model was completed in 34 hours, and the final map took about a week. The preprint has been submitted to bioRxiv and has not yet been peer reviewed; the particle has so far been expressed only in E. coli, and its function remains unknown.
Assessment: This is a process victory, not a puzzle solved. The function of vault is still unknown, and reading "a model in 34 hours" as "AI independently did structural biology" overestimates the model; the company itself put the errors into the body of the text.
Axiom formalizes the entire Stacks Project in Lean: agents write about 16.5 million lines of machine-checkable code¶
What happened: In the early hours of November 5, several posts relayed that Axiom Math had formalized the entire Stacks Project, a core reference work in algebraic geometry, in Lean. The relayed scale is about 16.5 million lines of code, 63,000 files, and 663,000 declarations, preserving the original project's definition dependencies and chapter structure and linking back to the original text; the process was carried out by AI agents under human supervision. The Stacks Project is said to have been started by Johan de Jong in 2005.
Why it matters: This is a direct rebuttal of the intuition that "AI-generated code at scale must be garbage", and it is also an engineering sample of agents handling extremely long, strongly dependent corpora. The spread was fragmented: individual posts ranged from 13K to 25K views and 23 to 158 likes, roughly 80K views in total.
Primary X posts: @romanftp's breakdown (16K views, 138 likes), @kimmonismus (15K views, 158 likes), @StockMKTNewz (25K views, 26 likes), @wallstengine (17K views, 23 likes).
Verification: Axiom's original official post and research blog are not in this sample, and the three figures of 16.5 million lines, 63,000 files, and 663,000 declarations appear only in relays and X news summaries, and are unverified. The boundary is the same as last time: Lean checks only formal derivations, and whether the definitions and statements are faithful to the original text still needs human review.
Assessment: If the numbers hold, this is the largest publicly known "machine-checkable mathematics library", and its significance lies in a reusable foundation rather than a single result. But until the official repository, Lean version, and compilation records are in hand, it can only count as a claim from sources.
Rumor: Astra 6.2 cuts prices 60% to match Opus 5.5, while the community also bets on a release this week¶
What happened: On November 4 at 03:20, Bindu Reddy posted that Astra 6.2 was cutting its price by 60% to align with Opus 5.5's price, and said that in blind tests users preferred using Astra for apps; the same post then says "Astra 6.2 may be the same price", wording that carries its own uncertainty. Around the release timing, other posts said that "Le GPT Fat" and "ChatBLT" posted by Tibo are teasers for a new flagship, that a new model ID has appeared in the API list, and that some people saw output labeled 6.2 Astra.
Why it matters: 88K views, 1,459 likes, the highest engagement in this window among non-official accounts. If the price cut is real, the unit-cost structure of agent workloads changes immediately, and it presses directly on the cost-performance narrative of Opus 5.5.
Primary X posts: @bindureddy (11-04 03:20, 88K views, 1,459 likes), guess about release timing (11-05 01:48, 11K views, 163 likes).
Verification: OpenAI has made no official announcement, and there is no first-hand source for the size of the price cut, the blind-test conclusion, or the release date; within the same post, "60% price cut" and "may be the same price" contradict each other. Rumor, unverified.
Assessment: This can only be read as a signal of price expectations, not written up as a pricing fact. The real evidence would be OpenAI's pricing page and API changelog.
2. AI Compute and Industry¶
No new AI compute and industry news was found in this issue that was worth separate coverage and could be traced to a first-hand source.
3. Chinese-Language AI Ecosystem¶
Chinese-language hands-on tests fill in two things the official charts left unclear about Hy4: $0.55 per task, and 19.8 on Terminal Bench Science¶
What happened: The official launch is covered in Section 1; the Chinese-language community added two pieces of information. The first is cost and ranking: Hy4 entered the Agent Arena overall leaderboard at No. 14 (the preview version was No. 25), with a median cost per task of $0.55, versus more than $5 for Opus 5.5; the author also admits the results are a notch worse (Opus at +20%, Hy4 at +4.7%). The second is an explanation of the weak spot: the 19.8 on Terminal Bench Science does not mean "40 points worse on scientific common sense" — those are 70 terminal tasks drawn from scientists' real research workflows, requiring an agent to run analysis, simulation, proof, and data processing inside a container; Hy4 scores 19.8, GPT-6.1 Sol 60.9, Opus 5.5 56.2, Kimi K3.1 Pro 41.7. Another post gives the pricing: 6 yuan input, 18 yuan output per million tokens, 0.3 yuan for cache hits, and notes that the weights and system card have not yet been released.
Why it matters: What the Chinese-language hands-on circle did this time was "translate the official charts", turning rankings into unit costs and a low score into leaderboard context, which is far more useful than relaying "crushes DeepSeek". Engagement was low (427, 416, 355, and 189 views, with 0 to 10 likes), but the information density was higher than most English posts of the same period.
Primary X posts: @piazzzzzz's cost comparison (427 views, 10 likes), @zg0820's explanation of Terminal Bench Science (189 views, 0 likes), @LiLensss's pricing and business hands-on (416 views, 1 like), @kqyno7's scores and free period (355 views, 9 likes).
Verification: All of the above are Chinese-language relays of Tencent's official report; the link to the official report could not be opened directly within this sample, so the specific scores and pricing are unverified; the Agent Arena ranking and per-task cost are third-party leaderboard figures and need to be checked back against the leaderboard.
Assessment: Hy4's real positioning is "cheap, good enough, and clearly lacking on research terminal tasks". Whether the $0.55 versus $5 comparison holds depends on the task distribution, and treating it as a general-purpose replacement will lead people into pitfalls.
Coverage note: This was a catch-up run; the three categorized searches ("large models", "labs and companies", "AI moves at tech companies") and the 24-hour slice of top events were not executed, so lower-circulation releases within the window may be missing.