Boss lens: the vendor itself says keep Opus for daily work and save Fable for harder jobs—Pro’s included week is capped at half of weekly usage, and Fable burns quota faster, so do not budget as if it were permanently bundled (inference).
AI NEWS
Today’s AI News
Chinese and English are the same day. This page lists every item, each with the original link, the author, and the engagement.
Latest day
November 10, 2026
11-09 07:50 – 11-10 07:50 (UTC+8)
Top 3
Boss lens: an agent that sits in the same chats as the team can take over cross-group collection and follow-ups—but rollout still depends on permissions, who steps in when a task fails, and whether the trail can be audited (inference).
Boss lens: cheaper per-edit pricing plus stabler multi-round changes matter if your team revises posters or product shots over and over—the 100-round and usability claims are the vendor’s, so trial it on your own assets first (inference).
All
Anthropic launches Claude Fable 5.2: a step above Opus 5.5 on its evals, with API prices cut about 20% to $8/$40 per million tokens
Boss lens: the vendor itself says keep Opus for daily work and save Fable for harder jobs—Pro’s included week is capped at half of weekly usage, and Fable burns quota faster, so do not budget as if it were permanently bundled (inference).
CollapseAnthropic launches Claude Fable 5.2: a step above Opus 5.5 on its evals, with API prices cut about 20% to $8/$40 per million tokens
Anthropic on Nov 9 released Claude Fable 5.2, its top-tier model, and Claude Mythos 5.2 for organizations in its verification programs. They share one underlying model; Fable adds cybersecurity and biology safeguards so it can be offered to all paying users. Anthropic said Fable 5.2 is a step above Opus 5.5 on nearly every capability it evaluated, generates outputs up to about 30% faster than the 5.1 generation, and costs $8 per million input tokens and $40 per million output tokens, down from $10 and $50. An optional fast mode speeds generation by up to about 2.5 times at $16/$80. The company still recommends Opus 5.5 as the daily driver for most coding and knowledge work, with Fable reserved for harder, more open-ended tasks; on FrontierCode 2.0 it said Fable 5.2 at High effort scored about 45.3%, roughly level with Opus 5.5 at Max (about 45.5%), at about half the cost per task. Through Nov 16 at 11:59 p.m. PT, Pro plans and standard Team seats can use Fable against weekly plan limits (up to 50%) and receive one Fable reset; Max, Ultra and premium Team seats receive one to three resets by tenure, expiring March 8, 2027. Benchmarks and demos in the post—including about 1,000 agents formalizing Fermat’s Last Theorem in 30 hours, and a Python service that started about 26 times faster after a move to Rust—are Anthropic’s own figures.
Alibaba’s Qwen Office 2.0 ships as a Team Agent that can join work chats; Intime’s voucher team used it across 60+ store groups
Boss lens: an agent that sits in the same chats as the team can take over cross-group collection and follow-ups—but rollout still depends on permissions, who steps in when a task fails, and whether the trail can be audited (inference).
CollapseAlibaba’s Qwen Office 2.0 ships as a Team Agent that can join work chats; Intime’s voucher team used it across 60+ store groups
IT Home reported on Nov 9 that DingTalk said Qwen Office 2.0 is generally available, shifting the product from a personal assistant toward a Team Agent for groups and organizations. The headline feature, Collaboration, lets people and a digital worker share a workspace linked to multiple DingTalk chats, documents, AI spreadsheets and meeting notes; once a task is assigned, the agent can summarize progress, capture decisions and follow to-dos. An official example: Intime Department Store’s voucher team previously had five operators collecting submissions from more than 60 store chats and filling a spreadsheet by hand; they now attach those chats and one AI spreadsheet to the same workspace so the agent can aggregate updates. The release also adds event-triggered Auto Tasks (a named person’s message, being @-mentioned in a specified group, or an approval), human-AI co-editing, a materials library, and sync across phone, computer and web. The company said tasks can keep running in the cloud after a computer is shut down.
ByteDance releases Seedream 5.5: up to 100 edit rounds on one image; about 0.28 yuan per 2K square edit, less than half of the prior Pro
Boss lens: cheaper per-edit pricing plus stabler multi-round changes matter if your team revises posters or product shots over and over—the 100-round and usability claims are the vendor’s, so trial it on your own assets first (inference).
CollapseByteDance releases Seedream 5.5: up to 100 edit rounds on one image; about 0.28 yuan per 2K square edit, less than half of the prior Pro
IT Home reported on Nov 9 that ByteDance’s Seed team released Seedream 5.5, an image model focused on stabler multi-round editing, native transparent assets and more natural texture. The company said pixel-level preservation outside the edited region allows up to 100 consecutive edits on the same work, with less color drift and quality fade, and that the model can emit RGBA images with an alpha channel so designers need less cutout work. It is rolling out on Jimeng, Doubao, Doubao Work and Volcano Engine, with a Pro version planned later. The API still bills on input image plus output image and does not charge for text tokens; a 2K square edit costs about 0.28 yuan, less than half of Seedream 5.0 Pro. ByteDance’s own tests said usability more than doubled versus the previous version in film/TV and office scenarios, and that severe text-rendering failures fell to 0.4% and mild ones to 1.7% on a set covering Chinese, English and dense small type. Those quality figures are the vendor’s.
Meta launches Muse Ember, its smallest model, at $0.10/$0.50 per million tokens; it says it matches Haiku’s best score at about one-third the cost
Boss lens: a cheaper small model for the repetitive steps in an agent stack can cut volume costs—pricing is official, but the quality claims are vendor and benchmark numbers, so re-test on your own workflow (inference).
CollapseMeta launches Muse Ember, its smallest model, at $0.10/$0.50 per million tokens; it says it matches Haiku’s best score at about one-third the cost
Meta on Nov 9 introduced Muse Ember, its smallest and cheapest model, built for the repetitive work agents do all day: tool calls, coding subtasks, routing, extraction and long-document parsing. List pricing is $0.10 per million input tokens, $0.004 per million cached input tokens and $0.50 per million output tokens, with a 1-million-token context window and up to 512,000 output tokens; prompts longer than 200,000 tokens rise to $0.30 input and $1.50 output. Citing Artificial Analysis, Meta said Ember delivers more intelligence per dollar than Claude Haiku 5.5 or GPT-6 Luna at any reasoning setting, matches Haiku’s best score at about one-third the cost, and sits on the efficiency frontier of that leaderboard. It is available today as muse-ember-1.0 on Muse Code and the Meta Model API, and with Muse Wonder and Muse Spark 1.3 on Google Cloud and Oracle Cloud. The comparisons come from Meta’s post and the Artificial Analysis results it cites.
OpenAI: the same dot can now connect to multiple local computers, and you choose which machine runs each task
Boss lens: a cloud assistant that can pick which local machine to use lets you separate testing from everyday work—check the plan limits and machine cap before you rearrange the team’s setup (inference).
CollapseOpenAI: the same dot can now connect to multiple local computers, and you choose which machine runs each task
OpenAI Developers posted on Nov 9 that users can now connect multiple local computers to the same dot and choose which machine a given task uses. The example was building and testing an app on a dedicated machine while you keep coding on a laptop; a follow-up said an iOS change can be tried in Xcode on a second Mac with screenshots sent back, or the same bug reproduced on a machine with an older SDK. Dot is the always-on cloud assistant OpenAI introduced at DevDay in late September, with its own cloud computer and browser, so it can keep working after you close your laptop. This is an official feature update; the posts did not say which subscription tiers get it or how many computers can be linked at once.
ChatGPT Enterprise is now on AWS Marketplace, so buyers can purchase it on their existing AWS bill and apply eligible cloud commitments
Boss lens: buying ChatGPT Enterprise on the AWS bill can remove a separate contract if you already have cloud commitments—this is procurement, not a data-residency guarantee, and pricing is still a private quote (inference).
CollapseChatGPT Enterprise is now on AWS Marketplace, so buyers can purchase it on their existing AWS bill and apply eligible cloud commitments
OpenAI said on Nov 9 that ChatGPT Enterprise is available through AWS Marketplace. Organizations can discover, buy and manage the suite using their existing AWS purchasing and billing relationship and apply eligible spend toward AWS commitments. The listing covers ChatGPT, ChatGPT Work, Codex and dots; OpenAI said features ship at the same time as purchases made directly from the company, and that the suite complements OpenAI models already on Amazon Bedrock—builders can use those models while business teams use ChatGPT Enterprise day to day. Eligible customers can also get the Healthcare and Financial Services editions through the same channel. The post does not publish a public price list and asks buyers to confirm eligibility with an OpenAI or AWS account team. This is a change in procurement and billing, not a promise that all data will stay on AWS.
Apple launches about $50 Intelligence Usage Packs: Logic Pro can add a one-time pack after the monthly cap, with unused usage rolling over
Boss lens: Apple is now selling one-time top-ups after a monthly cap—about $50, rollover for 12 months, no family sharing—so if staff rely on these creator features, check whether the included monthly limit already covers them (inference).
CollapseApple launches about $50 Intelligence Usage Packs: Logic Pro can add a one-time pack after the monthly cap, with unused usage rolling over
9to5Mac reported on Nov 9 that Apple has started selling Intelligence Usage Packs inside Creator Studio apps. In Logic Pro, Intelligence Features > Show Usage > Buy More Usage currently offers a single $49.99 one-time pack. Apple’s support text says the packs are one-time purchases, do not support Family Sharing, and can be used only by the Apple Account that was signed into iCloud at purchase; extra usage is consumed only after the monthly limit is reached, unused amounts roll over month to month, and they expire 12 months after purchase. Logic Pro is the first app to sell the packs; coverage said Final Cut Pro, Pixelmator Pro and the iWork apps Pages, Keynote and Numbers are expected to follow later this month. This is Apple’s first extra charge for some cloud-backed Apple Intelligence features, which until now shipped with the device at no separate usage fee.
Baidu discloses ERNIE Evolve, an automated-research system that it says raised ERNIE 5.5 Flash deterministic throughput by about 33.7%
Boss lens: an automated-research loop that actually landed production inference gains is worth a look if you run your own serving—the speedups are the vendor’s, measured on its own stack, so do not pencil in the same percentages (inference).
CollapseBaidu discloses ERNIE Evolve, an automated-research system that it says raised ERNIE 5.5 Flash deterministic throughput by about 33.7%
Baidu’s ERNIE team published a technical post on Nov 9 describing ERNIE Evolve, an automated-research system that proposes directions, edits code, runs isolated experiments and keeps only changes that pass independent evaluation and review. On public tasks it reported a NanoChat AutoResearch validation BPB of 0.8763, a NanoGPT Speedrun Track 1 time of 38.9 seconds and 2,590 steps on Track 3, beating the verified bests listed in the post on those two public tracks. The team then applied it to inference-efficiency work on ERNIE 5.5 Flash, which is still in development: on Kunlunxin P800 low-latency serving, about 168 hours and four search directions raised generation speed to 623.6% of baseline; a roughly three-day search over three directions raised deterministic-inference throughput by 33.7%, cut average time per token by 26% and first-token latency by 12%; extreme decoding speed rose another 14.9%. Baidu said the changes passed its own load tests and correctness checks. The figures are the company’s.
Meituan LongCat says LongCat-2.5 is now open-sourced: about 1.6 trillion total parameters, 48 billion active per token, under MIT for commercial use
Boss lens: open weights plus an MIT license mean a long-horizon agent model can be evaluated in-house rather than only as a hosted demo—fit still depends on your chips and tasks (inference).
CollapseMeituan LongCat says LongCat-2.5 is now open-sourced: about 1.6 trillion total parameters, 48 billion active per token, under MIT for commercial use
IT Home reported on Nov 9 that Meituan LongCat has open-sourced LongCat-2.5, the native multimodal model it first announced on Nov 2. The release uses the MIT license, allowing commercial inference and further training without a separate grant. The model has about 1.6 trillion total parameters, activates about 48 billion per token, and supports a 1-million-token context window, aimed at long-horizon agents, large-codebase development and visual coding. The report said the team jointly optimized the architecture and deployment for domestic chips with tight memory and bandwidth, and ships a W4A4 scheme that quantizes expert weights and activations to 4 bits and runs native INT4 matrix multiplies. Weights are on Hugging Face as meituan-longcat/LongCat-2.5. This is the open-weight follow-through of the Nov 2 announcement, not a brand-new model debut.
OpenAI deploys Control Room on research infrastructure: engines check authorization before each generation and can halt output within 35 seconds
Boss lens: labs are putting a kill switch at the inference layer and failing closed if that switch dies—if you let agents run unattended, decide in advance who can stop a job and what happens to the work after it stops (inference).
CollapseOpenAI deploys Control Room on research infrastructure: engines check authorization before each generation and can halt output within 35 seconds
OpenAI on Nov 9 described Control Room, a central system that tracks model workloads across its research infrastructure and stops them quickly when safety or security concerns arise. It covers research inference such as training sampling, evaluations and data processing: each job gets a job_id, and inference engines ask Control Room for permission before producing output; once a monitor or a person marks the job stopped, the next request is denied. The company said the longest gap from a stop decision to the point where further progress is blocked is 35 seconds, caused by an authorization cache built for traffic volume. Termination is implemented in the more centralized inference layer rather than in researchers’ varied workload code, and the system fails closed: if Control Room itself goes down, all inference requests are denied. It is already deployed on frontier-model research workloads and will finish rolling out to remaining research jobs over the following days. This is internal research infrastructure, not a control panel shipped to outside customers.
The Information: inference provider Fireworks told customers including Harvey of a more serious hack; some are considering leavingUnconfirmed
Boss lens: a breach would normally send customers looking for a new inference vendor, but scarce GPUs can trap them—if critical work sits with one provider, ask about backups and how long a switch would take, not just the price (inference).
CollapseThe Information: inference provider Fireworks told customers including Harvey of a more serious hack; some are considering leaving
The Information’s Stephanie Palazzolo posted on Nov 9 that Fireworks, among the most highly valued inference providers, had told some large customers—including legal-AI firm Harvey—that it suffered a more serious hack exposing customer data, and that some customers are considering moving their work elsewhere. The public post and headline-level summary present this as a people-familiar-with-the-matter account of an incident disclosed in roughly the past week. The full article is paywalled; this item uses only the public post and does not list which classes of data were exposed, how many customers were affected, or whether Fireworks has issued a public incident note. Engineer Gergely Orosz later wrote that inference vendors mainly sell scarce GPUs, easier operations and data security; a breach would normally be very damaging, but GPUs are tight enough that customers often have few places to move on short notice.
Epic ships Unreal Engine 5.9 MegaJam Preview, calling it the first Unreal release built primarily with agentic workflows
Boss lens: even an engine vendor now ships a build directed by people and written largely by agents—and it also warns of a one-week sprint, shorter QA and known bugs, so faster delivery still needs review and a rollback plan (inference).
CollapseEpic ships Unreal Engine 5.9 MegaJam Preview, calling it the first Unreal release built primarily with agentic workflows
Epic Games posted on its developer forum (dated Nov 9) that, to make the Epic AI MegaJam on Nov 12–15, it decided about a week earlier to cut a Unreal Engine 5.9 Preview. The studio called it the first Unreal release built primarily with agentic workflows: a handful of people directed fleets of agents to port AI tooling from the UE6 branch and to implement the web as a target platform, with agents writing web-platform code from UE4’s retired HTML5 support while engineers directed, reviewed and tested. Epic said the preview had about a week of development and a shorter QA pass than usual, that some agent-written code is still being cleaned up, and that parts of the web renderer are prototypes to be refactored before the full 5.9 release. Experimental additions include Unreal MCP launching with the editor, an in-editor Unreal Assistant that can connect Claude Code, Codex, Cursor or OpenCode, and running projects in a browser via WebAssembly and WebGPU. The post lists many known issues and asks people to test on a copy of their project rather than converting production work.
Meta expands Instagram’s AI age estimation worldwide, placing suspected teens in Teen Account protections even if they entered an adult birthday
Boss lens: if you market to younger audiences on Instagram, the platform may now reclassify some users into teen protections even when they typed an adult birthday—worth a check by whoever runs those accounts, not a new ad product (inference).
CollapseMeta expands Instagram’s AI age estimation worldwide, placing suspected teens in Teen Account protections even if they entered an adult birthday
Meta posted on about.fb.com (dated Nov 9, with a Nov 8 timestamp also visible; TechCrunch and others reported it on Nov 9) that it is expanding Instagram’s AI age-estimation technology worldwide. Accounts it suspects belong to teens will be placed in Teen Account protections even if they registered with an adult birthday, a change Meta said will automatically cover millions more teens with limits on who can contact them, what they see and how long they spend in the app. Parents in the United States will start seeing an in-feed explainer this month; in Australia, Meta has also begun using Apple’s app-store age-range signal as an extra cue. The same technology continues to roll out on Facebook, with a global Threads launch planned later this year. Meta tied the move to its agreement with 52 bipartisan U.S. attorneys general. The post did not publish a false-positive rate or how often appeals succeed.
China’s foreign ministry: Beijing is first to ratify the pact founding a World AI Cooperation Organization; 38 countries became founding members in July
Boss lens: China finishing ratification first is a signal that a new intergovernmental venue may take shape—useful context for cross-border data and model work, but not a new rulebook yet (inference).
CollapseChina’s foreign ministry: Beijing is first to ratify the pact founding a World AI Cooperation Organization; 38 countries became founding members in July
Xinhua reported on Nov 9 that foreign-ministry spokesman Guo Jiakun said China had become the first country to ratify the agreement establishing a World AI Cooperation Organization, calling it a sign of Beijing’s commitment to getting the body set up. The question cited a decision by the 14th National People’s Congress Standing Committee at its 25th session to approve the pact. Guo said President Xi Jinping announced the organization’s creation in July at the 2026 World AI Conference and high-level meeting on global AI governance, and that 38 countries became founding members. As initiator and host, he said, China would push the organization toward actual operation, focus on practical cooperation and capacity-building for the Global South, and welcome more countries to join. The remarks mark completion of a ratification step, not a newly issued set of regulatory rules.
Ethan Mollick: he asked Fable 5.2 to build a hands-off journey through 3D worlds; about 25,000 agents created more than 450 worlds
Boss lens: a single prompt plus tens of thousands of agents can produce a huge volume of work—and, as Mollick later noted, those agents can also grow their own bureaucracy and slow down, so set reviewers and a stop rule before you scale (inference).
CollapseEthan Mollick: he asked Fable 5.2 to build a hands-off journey through 3D worlds; about 25,000 agents created more than 450 worlds
Wharton professor Ethan Mollick posted on Nov 9 that he asked Claude Fable 5.2 to build an endless, hands-off journey through 3D worlds linked by portals. About 25,000 agents created more than 450 worlds in code, inspired by books, art and places. He linked a playable page and the GitHub repository emollick/ai-portals. The write-up is a first-person experiment, not a third-party audit of hours or quality; treat the worlds as whatever the repo and demo actually contain. In a follow-up he added that the agents eventually built a bureaucracy that slowed development, an example of jagged capability showing up at scale.
Guoyu: the parking assistant parkuma actually took one day to finish, because sandbox and cloud-browser pieces were already in his harness platform
Boss lens: a one-day app is often the visible tip of a platform that already solved sandboxing and browsing—if you chase one-off speed without leaving a reusable base, the next build will not be as fast (inference).
CollapseGuoyu: the parking assistant parkuma actually took one day to finish, because sandbox and cloud-browser pieces were already in his harness platform
Guoyu (@turingou) posted on Nov 9 that many people had complimented parkuma’s completeness, but that the vertical parking assistant actually took only one day to finish. Before building it, he wrote, he had already spent a long time on a harness platform, and details such as the sandbox and cloud browser had been solved earlier in the wanman project. This is a follow-on to his Nov 5 note that a parking-fee app took about 15 minutes once the platform worked: today’s point is that the more complete parkuma shipped in a day because the platform had already been thickened. It is a first-person build log, not a third-party timesheet, and the post does not list a full feature set or externally checkable usage figures.
LG AI Research opens a free API for K-EXAONE 2.5, saying its agent skills are comparable to DeepSeek V4.1 Flash and GLM-5.3
Boss lens: a free vendor API lowers the cost of a trial—treat the comparison with DeepSeek and Zhipu as the publisher’s claim, and ask how long the free tier lasts before you plan around it (inference).
CollapseLG AI Research opens a free API for K-EXAONE 2.5, saying its agent skills are comparable to DeepSeek V4.1 Flash and GLM-5.3
IT Home reported on Nov 9 that LG AI Research said the same day in Korea it was opening a free API for K-EXAONE 2.5, so developers, companies and research groups can use the model in products and studies at no charge. The release is an additional-training update of the original model launched in August. LG said its results on AutomationBench, SWE-Bench Pro, BrowseComp and Claw-Eval are comparable to DeepSeek-V4.1-Flash and GLM-5.3. IT Home also linked the institute’s own news post. The comparison and the free-access claim come from LG; the coverage did not say how long the free tier will last, whether there is a rate cap, or cite an independent rerun of the benchmarks.
yetone: the local assistant pi can now use Doubao, shown in a settings screenshot
Boss lens: a local assistant that can call Doubao lets files stay on the machine while using a domestic cloud model—confirm which versions, pricing and whether data leaves the device before you standardize on the screenshot (inference).
Collapseyetone: the local assistant pi can now use Doubao, shown in a settings screenshot
yetone, author of the open-source local assistant pi, posted a settings screenshot on Nov 9 with the line that users can now use Doubao inside pi. This is a first-party product note, not an announcement from ByteDance. The post did not list which Doubao models are wired up, how they are billed, or whether conversation content is sent to Doubao’s cloud or used for further training. Pi runs on a user’s own computer; once a cloud model is attached, local file handling and remote inference sit side by side. Treat the details as whatever the product docs and the user’s own settings later show.
Gorden Sun on on-device vision model Mirage: about 51.3 million parameters and 22MB after quantization; the publisher says a full pass takes about 74ms on a Galaxy S26 NPU
Boss lens: a 22MB vision model that can label, read text and outline objects on a phone may keep some inspection photos off the cloud—the 74ms figure is the publisher’s on one flagship device, and 118 bookmarks show demand (inference).
CollapseGorden Sun on on-device vision model Mirage: about 51.3 million parameters and 22MB after quantization; the publisher says a full pass takes about 74ms on a Galaxy S26 NPU
Gorden Sun (@Gorden_Sun) on Nov 9 wrote up Cactus Compute’s on-device vision model Mirage and linked its Hugging Face repo. He said the model has about 51.3 million parameters, shrinks to about 22MB with a specialized 2-bit quantization scheme, and completes a full analysis pass in about 74 milliseconds on the NPU of a Samsung Galaxy S26. He described it as understanding a scene, outlining objects, reading signage or labels, sorting out spatial relations and returning structured decisions for other apps. Cactus’s own page lists the same 51.3 million parameters and about 22MB, plus about 74ms end-to-end on the S26 NPU and about 107ms on CPU. Those latency figures are the publisher’s measurements on a specific phone, not a guarantee for every device. The post drew 118 bookmarks, more than its 98 likes.
Tw93 shares a Grok evening-digest prompt that summarizes accounts he follows at 8 p.m.; the post drew 165 bookmarks, more than 118 likes
Boss lens: a scheduled digest can replace an evening of scrolling—165 bookmarks show demand—but the result depends on whom you follow and whether the summary misses something important (inference).
CollapseTw93 shares a Grok evening-digest prompt that summarizes accounts he follows at 8 p.m.; the post drew 165 bookmarks, more than 118 likes
Tw93 (@HiTw93) posted on Nov 9 about how he keeps up with X without constant scrolling: a recurring Grok prompt that, every evening at 8 p.m., compiles highlights from accounts he follows plus other posts worth reading, using only the official X API and skipping anything already covered the day before. His format opens with three short sentences on the day’s main threads, then one section per person with original links, two to four points and a one-sentence take, plus a closing tools section, aiming for about 10 to 15 sections. He said the daily digest lets him scroll less and stay more focused. This is a personal workflow, not a newly shipped X or Grok product. The post drew 165 bookmarks, more than its 118 likes.