X AI Daily¶
Translation
English translation of the Chinese original.
1. AI Models, Products and Applications¶
Anthropic releases Claude Fable 5.2 and Mythos 5.2: about 30% faster, 20% cheaper¶
On November 9, Anthropic released Claude Fable 5.2 and Claude Mythos 5.2, which are the same underlying model. Fable 5.2 comes with cybersecurity and biosecurity safeguards and is open to all users; Mythos 5.2 is available only to organizations that pass review under the Life Sciences and Cyber Verification programs. The company says both improve on coding, reasoning and vision, with output up to 30% faster than 5.1, at 8 USD per million input tokens and 40 USD per million output tokens; there is also an optional fast mode, up to 2.5x faster, priced at 16 USD / 80 USD. Pre-release test results cited by the company include: designing three working electronic devices from scratch, solving two open problems in black hole physics, and formalizing a proof of Fermat's Last Theorem within 30 hours.
Why it matters: this is the most widely spread AI release in this window. CNBC reported that it is Anthropic's first frontier model release since Dario Amodei called for slowing AI development in September.
- Core posts: @claudeai release post, about 7.1 million views, 69K likes, 6,468 reposts; @AnthropicAI launch post, about 488K views, 11K likes; @claudeai pre-release test cases post, about 653K views, 13K likes.
- Verified: Anthropic's official announcement confirms the same underlying model, up to 30% faster speed, 8/40 USD pricing, fast mode at 2.5x and 16/80 USD pricing, and the test results above; Claude developer docs confirm a 1M context, 128K maximum output, and list four potentially breaking changes; the 20% price cut comes from CNBC's report.
- Take: the Fermat's Last Theorem formalization, the black hole problems and so on are the vendor's own account of pre-release testing, not independent reproduction. The developer docs still recommend starting most tasks with Opus, and switching to Fable 5.2 only if evals fall short.
Follow-up: Meta releases its smallest model Muse Ember: 0.10 USD per million input tokens, aimed at agent grunt work¶
On November 9, Meta released Muse Ember (muse-ember-1.0), the smallest and cheapest model in the Muse family, intended for tool calling, coding subtasks, request routing, information extraction and long-document parsing. It supports text, image, audio and video input, with a 1M token context and 512K maximum output. Pricing is 0.10 USD per million input tokens and 0.50 USD per million output tokens, with inputs above 200K billed at long-context rates. Meta says it is the first model trained with Muse Wonder Full. It launched on the Meta Model API and Muse Code, and @AIatMeta also says it is already available on Google Cloud.
Why it matters: this follows up on Alexandr Wang's weekend teaser of a "release this week." Third-party evaluator Artificial Analysis puts it level with Haiku 5.5 at about one-third the cost per task; cheap execution models paired with large models for planning are becoming the standard division of labor for agents.
- Core posts: @AIatMeta release post, about 612K views, 2,374 likes; @alexandr_wang, about 148K views, 1,819 likes, saying it is close to Muse Spark 1.3 (max) in capability at 17x lower price (on the AA basis); @ArtificialAnlys evaluation post, about 95K views, 767 likes, giving it an AA Intelligence Index of 36, level with Claude Haiku 5.5 (max), at about one-third the cost per task. @OpenRouter has also made it available.
- Verified: Meta's official blog confirms the positioning, 1M context and 512K output, 0.10/0.50 USD pricing, training with Muse Wonder Full, and "about 0.14 USD per AA task"; Meta Model API docs confirm the input modalities and model ID.
- Take: the comparisons with Haiku and Luna use Meta's chosen Artificial Analysis basis; real-world agent performance still depends on user feedback.
OpenAI unveils Control Room: cutting off research model output within 35 seconds¶
On November 9, OpenAI unveiled Control Room, a system that centrally registers and tracks internal research inference jobs, covering agent sampling in training and evaluation, data processing and more. Working with its misalignment monitoring systems, it can stop a job from getting further model output within at most 35 seconds of a stop decision. The cutoff sits at the inference layer, so even if the job's code is still running it cannot get output; if Control Room itself goes down or is tampered with, all inference requests are denied by default (fail closed).
Why it matters: it is rare for a frontier lab to publish engineering details of its internal "kill switch," and financial news accounts relayed it heavily, spreading it beyond AI circles.
- Core posts: @OpenAI release post, about 264K views, 2,144 likes; a representative relay, @DeItaone news alert, about 83K views, 200 likes.
- Verified: OpenAI's official article confirms that the 35-second ceiling comes from the caching layer for inference authorization, that the goal is to bring it to zero, and the design for stopping output when the service is down; that fail closed already covers all frontier model research jobs, with remaining research jobs rolling out within days, comes from FourWeekMBA's account of the original and was not seen directly in the excerpt on the official page.
- Take: it governs OpenAI's internal research infrastructure, not an external product; the 35 seconds is counted from the "stop decision," so the time for humans to review an alert is not included.
NVIDIA releases GR00T 2 at CoRL: an open "world-action model"¶
At CoRL, the robot learning conference in Austin, NVIDIA released GR00T 2, an open "world-action model" built on Cosmos that predicts both future frames and robot actions, usable for parallel grippers, dexterous hands and humanoid robots. The company says it was trained with human first-person video, 100,000 hours of which are available for download as Human-1M-100K; on 50 tasks with no targeted demonstrations it completed at least 40, with a 71% average success rate with one demonstration per task and 95% with five. Isaac Lab 3.0 was released at the same event.
- Core posts: @NVIDIARobotics release post, about 30K views, 379 likes.
- Verified: NVIDIA's official blog confirms 40/50, 71%, 95% and the 100,000 hours of open data; the GR00T 2 research page gives a different account: 40% zero-shot success, reaching 95% after fine-tuning for 30 minutes with five demonstrations.
- Take: the results are NVIDIA's own testing, and The Next Web notes the paper acknowledges the test tasks were selected from ones the model already learned well; Foxconn, Siemens and others are still in testing.
Hot on X, facts unconfirmed: new Gemini 4 "Krypton" checkpoint¶
Several testers on X say a new checkpoint, Gemini 4 Krypton, has appeared inside Google, and have posted samples such as generated 3D scenes, arguing it is stronger than Gemini 4 Carbon.
- Representative posts: @bidyut_io, about 17K views, 255 likes; @Lentils80 comparison video, about 15K views, 253 likes; @abhinavflac samples, about 10K views, 193 likes.
- Unverified: no official Google statement was found; the model's identity and capabilities come from the testers' own accounts, and a single sample cannot represent overall performance.
2. AI Compute and Industry¶
No new AI compute and industry items with first-hand sourcing in the window and worth listing separately were found this issue.
3. Chinese AI Ecosystem¶
ByteDance releases Seedream 5.5: 100 consecutive edits without drift, native transparent backgrounds¶
ByteDance's Seed team released the image model Seedream 5.5 on November 9, with three upgrades: pixel-level preservation of unmodified regions during multi-turn editing, eliminating color shifts, with support for 100 consecutive edits to the same image; native generation of RGBA transparent or semi-transparent assets with an alpha channel, removing the need for cutouts; and more realistic lighting, skin texture and rendering of dense small text. The model is rolling out on Jimeng, Doubao, Doubao Work and Volcano Ark, and overseas via Dreamina, CapCut and BytePlus.
- Core posts: @testingcatalog launch post, about 12K views, 251 likes; Chinese discussion @Dicklong1999, 670 views, 13 likes; Turkish hands-on comparison @ozansihay, about 18K views, 100 likes.
- Verified: ByteDance Seed's official blog confirms the three upgrades and 100-round editing; the domestic launch channels are covered in Cocoloop's report, which says the blog does not state the maximum resolution or API pricing.
- Take: 100 rounds is the official upper limit, and the blog has no quality comparison between round 1 and round 100, so real performance needs users' own testing. Some posts claim it beats GPT Image 2.5 and Nano Banana 2.1; that claim was not verified in the excerpt of the official blog.
X discussion and third-party material on image releases such as MiniMax H3-Image disagree on the parameters, so they are not included in this issue.