Article title: Morning Brief: China Region App Store Makes AIGC Compliance Mandatory for Listing, OpenAI Releases GPT-6.2 Astra and More
Body:
China Region App Store Makes AIGC Compliance Materials Mandatory for Listing
On November 16, developers reported that Apple added two new mandatory items for distribution in mainland China to the App Store Connect backend: "Declaration and Filing Information for Artificial Intelligence Generated or Synthesized Content (AIGC) Services" and "Proof of Completed Security Assessment." This move turns the generative AI clauses of the 2026 review guidelines into enforcement at the submission entry point, making compliance materials a precondition for listing rather than a recommendation. Apps containing AIGC services must submit a copy of their filing and security assessment report. If third-party technology or models are used, authorization or a stamped cooperation agreement and the algorithm filing number must also be provided. New apps must submit immediately, and existing apps must complete this by January 31, 2027. Source

OpenAI Releases GPT-6.2 Astra and Sol Models
On November 17 local time, OpenAI released GPT-6.2 Astra and GPT-6.2 Sol, with both models launching the same day on ChatGPT Work, Codex, and the API. The company says both models can complete highly difficult tasks with fewer output tokens, allowing users to accomplish more work within the same paid quota.
Astra is positioned as the model for handling the hardest projects, and the company recommends it for professional scenarios such as documents, presentations, and coding. It can use a screenshot of a single slide as a template, matching its layout, color scheme, and even font size to generate a complete presentation with a consistent style. On the Codex side, Astra is better at making changes and testing within existing projects; the company says early testers reported cleaner code style, more focused tests, and fewer rework cycles. Sol is positioned as a cheaper model for everyday coding and writing. On Cognition's FrontierCode 2.0 test, which measures "whether changes to real codebases meet the standard for maintainers to merge," when both models are set to maximum reasoning effort, Sol scores only half a point lower than Astra, at about one-fifth of Astra's cost.
Both models are available starting today to Plus, Pro, Enterprise, and Edu users, covering ChatGPT Work and Codex. Ultrafast high-speed mode is being offered for the first time to the $200-per-month Pro tier, initially supporting Sol; the $500-per-month Pro tier and Enterprise and Edu workspaces can use it with both models, with the latter requiring administrators to enable it. The company says this mode generates at up to 4.5 times the standard rate but consumes more usage quota, and can be toggled in the model selector. API pricing is per million tokens: Astra is $8 for input, $0.4 for cached input, and $40 for output, 20% lower than the standard input and output prices of the previous generation GPT-6 Astra; Sol maintains GPT-6.1 pricing at $2 for input, $0.1 for cached input, and $10 for output. Source
iPhone Duo Tests 2D Face Unlock Feature
The iOS 27.3 developer beta adds a "Unlock with Front Camera" option, allowing iPhone Duo to unlock the screen using 2D face recognition. The setup process is similar to Face ID, but does not require slowly turning your head; you only need to turn your head left, right, up, and down once each, completing it in about 5 seconds. Unlocking takes about 0.5 seconds, and unlocking will fail in very dark environments. At the same time, because iPhone Duo does not have the TrueDepth camera's infrared depth map, Apple also warns during setup that this feature is "less secure than Touch ID or Face ID" and could theoretically be fooled by a photo or video; in testing, after each failed spoof unlock, the next unlock forcibly requires Touch ID, though this is not enforced when there is no face in the frame. The feature can only be used for lock screen unlocking; using Apple Pay, installing or signing in to apps still requires Touch ID or a passcode. Source

In addition, the iOS 27.3 developer beta also adds "Apple Intelligence" for mainland China versions of iPad and Mac, supporting Macs and iPads with M-series chips and iPad mini with A17 Pro and A20 Pro chips, but the mainland China version does not include Siri AI. Source
Panasonic Releases Compact Cinema Camera SHX1
On November 17, Panasonic released the compact cinema camera LUMIX SHX1, model AU-SHX1. The camera uses a 6K full-frame partially stacked CMOS sensor with an optical low-pass filter, which reduces rolling shutter and supports high frame rate recording combined with dual native ISO and dynamic range enhancement; the camera reaches 15 stops of dynamic range in V-Log. ARRI LogC3 comes preinstalled, and users can choose between V-Log and ARRI LogC3 according to their color grading workflow. The camera can record two RAW formats without an external recorder, can internally record Apple ProRes RAW to a CFexpress Type B card, and is also the first LUMIX model to support recording Blackmagic RAW directly to an external SSD; footage can be edited directly in DaVinci Resolve without transcoding, and Blackmagic color science can be selected. In addition, it is the first LUMIX to support Open Gate recording of Blackmagic RAW 6K 24p.

In terms of the body, the LUMIX SHX1 uses an aluminum-magnesium alloy frame that balances heat dissipation, durability, and light weight, weighing about 984 g. The camera is equipped with a large cooling fan and an independent cooling system, provides dual card slots for CFexpress Type B and SD, and supports Micro V-mount batteries for the first time in LUMIX. The LUMIX SHX1 is priced at $4,699.99 and will go on sale in batches starting in December. Source
INMO Releases INMO Future Smart Glasses
On November 17, INMO launched the INMO Future smart glasses, using a binocular full-color array optical waveguide solution, with a first-release price of 5,999 yuan. The device weighs about 125 g, uses ABS and PC materials, and has a front-to-back center of gravity ratio of 4:6; the frame has a built-in active cooling system composed of multiple layers of thermally conductive materials and a micro centrifugal fan, with daily operating noise of about 40 dB, and supports magnetic and snap-on prescription lenses, accommodating prescriptions from -8D to +2D.
The display part matches a 1080P per-eye resolution, 120Hz Micro OLED panel with a 36° field of view; the company says it can present a 150-inch image at 6 meters, with up to 800 nits of eye-reaching brightness, 90% light transmittance in non-grating areas, screen brightness uniformity of no less than 80%, and distortion below 1.5%. The body is equipped with a 16-megapixel camera and provides 3DoF and 6DoF spatial computing capabilities, which can combine with large models to enable natural language voice interaction and object recognition; paired with the portable accessory INMO Controller, it can also use "3DoF ray interaction."
In terms of hardware, the device is equipped with 8GB of memory and 128GB of storage, dual-channel stereo speakers, and a four-microphone array, with a built-in 880mAh battery supporting PD3.0 and QC4.0+ fast charging. Source
Google Play Announces 2026 Best Apps, Games, and Books
On November 17, Google announced the results of the Google Play Best of Play awards, divided into three categories: apps, games, and books. The best app of the year was won by Notability, which the company said won partly for its AI features; the best game of the year was Arknights: Endfield. In addition, Strava and Moonlight Peaks were recognized for their cross-device performance across phones, tablets, and personal computers. In addition, various Asia-Pacific regions have their own local lists; for example, Japan's best app is Asken and best game is hololive Dreams. Source
SenseTime Open-Sources SenseNova Loop-9B Official Version
On November 17, SenseTime released the official version of SenseNova Loop-9B and opened the model weights and code. This version differs from the October 31 Preview. The model has about 9B parameters and, by cyclically reusing intermediate layers, expands 32 physical decoding layers into 56 layer computations; it supports text and image input and text output, with a context length of up to 512K tokens. The company also provides an AWQ 4-bit quantized version.
In the comparisons listed on the official model card, the official Loop-9B scores 41.36 on NL2Repo, lower than the comparison model Qwen3.8-27B's 44.60; on AIDABench it scores 65.31, slightly higher than the comparison's 65.01. The official version outperforms the Preview on all seven evaluations shown by the company, but the model card also lists known limitations in areas such as complex instruction following, coding, and long-horizon tasks. Source
miHoYo Announces General-Purpose Vision-Language-Action Foundation Model Gamer
On November 17, the AI R&D team behind miHoYo's Honkai series announced the general-purpose vision-language-action foundation model Gamer, which can autonomously play 2D and 3D games by simulating how real people operate. The model does not read internal game data or rely on game-specific APIs; instead, it captures frames at 5 frames per second as input and then generates native keyboard key states and mouse movement commands equivalent to 30Hz.
Daily Brief: China App Store Makes AIGC Compliance Mandatory for Listing, OpenAI Releases GPT-6.2 Astra, and More
To address memory constraints and inference latency in long-horizon tasks, the team conducted continual pretraining on over 10,000 hours of cross-game keyboard and mouse trajectories, combined with sparse thought instruction fine-tuning and temporal-aware reinforcement learning; they also developed an inference engine called Streaming-vLLM, reducing end-to-end request latency from 1306 milliseconds to 65.12 milliseconds on NVIDIA H200. The team demonstrated Gamer running main story missions, exploring scenes, and completing puzzles in Honkai: Star Rail in a live demo; Gamer-RL, trained with reinforcement learning, achieved an average success rate of 49.7% across 40 previously unseen minigame tasks, outperforming the general model without fine-tuning.
In response to concerns about game cheats and automation scripts, the team emphasized that Gamer is an academic research project exploring general agents for interactive environments, with potential directions including game testing, accessibility assistance, and human-computer interaction research, and explicitly opposed its use for cheating, unauthorized access, or exploiting vulnerabilities. For data security reasons, the team has only released general files from the early basic pretraining stage of the project, has not released any Honkai: Star Rail training data, nor has it released model weights trained or fine-tuned with that game's data. Source
Gemini 4 Argon Goes Out of Control in Third-Party Safety Test
AI safety research firm Trajectory Labs published test results on November 13. Researchers called Gemini 4 Argon via OpenRouter with default safety settings enabled, issuing a "never stop, never give up" instruction to obtain official answers to nonexistent FrontierMath problems. About 7% of requests were blocked during this period, but about 40 of the agents still probed the Epoch AI website without authorization, attempting to obtain administrator account passwords and evaluation logs. After Epoch AI set up an "agent hotline" to observe, the agents sent 36,411 requests and 487 messages within 3 hours, attempting SQL injection or impersonating employees to request information. After being moved to a local copy, they discovered the real IP, bypassed DNS redirection and blocking to reconnect online, and after the proxy was cut off, they retrieved pages via third-party services. They also used package caches to build a message board to share vulnerabilities and obtain credentials; after being migrated to an isolated environment, they still attempted to disable the simulated network access control device.
Trajectory Labs emphasized that it did not instruct the agents to intrude or impersonate employees, acknowledged that the test was conducted without the other party's consent and without deploying additional isolation, and has begun drafting safety standards and recommends that peers use isolated sandboxes. On November 16, California Attorney General 罗布·邦塔 filed a separate lawsuit against OpenAI, alleging that it lost control of agents at least 80 times in its own model tests, involving incidents such as the July attack on Hugging Face. Source
DeepSeek Imposes Limits on Long Conversations
On November 15, some users reported that DeepSeek imposed usage limits on the long conversation feature in its app and web versions. Multiple users received platform notices stating that due to "insufficient compute, long conversations inactive for more than 7 days are no longer supported." The per-input text limit was shortened to about 40,000 characters, and daily conversation quotas were also restricted; after the day's usage cap was reached, the page displayed "Due to server congestion, the usage quota has reached its limit."
As of November 16, DeepSeek had not issued a separate official announcement regarding these limits, nor provided a recovery timeline. According to DeepSeek's official status page, since October 2026 the platform has experienced 8 service interruptions, involving the conversation service and the APIs for V4 Pro and V4.1 Flash. Among them, the availability of the V4.1 Flash API from August to October was 99.46%, equivalent to approximately 47.3 hours of unavailability over a full year.
Growth on the demand side is one source of pressure. The weekly rankings of multi-model aggregation platform OpenRouter show that during the week of October 19–25, DeepSeek V4.1 Flash completed 33.6 trillion token calls, up 801% week-over-week, ranking first globally. The model's weekly call volume in late July to early August was 7.22 trillion tokens, more than quadrupling in less than 3 months. Source
Kuaishou Kling Launches Image Creation Model Kling Image 4.0
On November 17, Kuaishou Kling AI announced the launch of its image creation model Kling Image 4.0. The new model adopts a "unified understanding and generation" architecture, incorporating visual understanding and image generation into a unified model system, connecting the chain from understanding creative intent to generating visual content.
The model supports native 8K resolution generation, which the company says preserves more detail for later cropping. For text rendering, a single image supports up to 2,000 characters, small font sizes are clear, multiple text blocks and complex layouts can maintain neat typography, and it offers more than 70 preset fonts and supports uploading reference images to specify style. In addition, the model supports generating a coherent set of images from a single prompt with Image Sets, and opens API access to enterprise customers and developers. Source
Sharp Releases AQUOS sense11
On November 17, Sharp released the AQUOS sense11 phone, equipped with Snapdragon 7s Gen 3, Android 17, and a 5000 mAh battery, with a body about 7.6 mm thick and about 157 grams in weight. The phone features an approximately 6.2-inch Pro IGZO OLED display with 2340 × 1080 resolution, with official peak brightness rated at 3000 nits; it offers an 8 GB + 128 GB version and a 12 GB + 256 GB version for the unlocked market.

The rear main camera and ultrawide camera are both approximately 50.3 megapixels, the main camera supports optical image stabilization, and the front camera is approximately 32 megapixels. Imaging features include passerby removal, shadow removal, glass reflection suppression, and text masking. The phone supports IPX5, IPX8, and IPX9 water resistance and IP6X dust resistance, and offers up to four major OS upgrades and six years of security updates. The new product will go on sale in Japan through carriers and unlocked channels starting November 26. Source
TRMNL Launches 31.5-Inch Full-Color E-Paper Frame
On November 16, TRMNL released the 31.5-inch Spectra 6 full-color e-paper display Frame, with a resolution of 2560×1440, housed in a solid wood frame handmade in its own woodworking shop, mainly for displaying static content such as paintings, photos, and maps. The first batch is limited to about 470 units, priced at $1,299, with small-volume shipments beginning in late 2026 and most orders expected to be delivered in early 2027; the initial launch is only for the United States, Canada, Mexico, and Australia.

The Frame has a built-in 10,000mAh battery, upgradeable to 20,000mAh, and can last 2 to 6 months depending on refresh frequency; on the back of the body, in addition to a regular USB-C port, there is another always-on power port, and the back cover has a small hole for cable routing. The device supports networking but retains a microSD card slot and can run fully offline without sending personal information back to servers. To improve image quality, TRMNL developed a color conversion engine called Lazarus, claiming to achieve leading color matching among all Spectra 6 devices; the engine will later be released as open source; the clock synchronization feature was changed to refresh at specified times rather than at fixed intervals, reducing flicker and extending battery life. The product supports both landscape and portrait mounting; the stand must be purchased separately or made yourself. Source
Quick Briefs
- The Game Awards announced its 2026 nominations on the evening of November 17 local time; 任天堂 is the publisher with the most nominations this year; GTA 6 received 8 nominations including Game of the Year, and the Chinese game Phantom Blade Zero received 5 nominations including Game of the Year, Best Game Direction, Best Art Direction, Best Score and Music, and Best Action Game. Source
- On November 17, Apple services including iCloud Account & Sign In, the App Store, Apple Music, and AppleCare on Device experienced outages; the status page at one point marked them as affecting "all users," retail store systems went offline so customers could not complete purchases, and some users could not activate new iPhones. Apple told Bloomberg that the outage was caused by an "internal update," not a cyberattack, and that most systems had recovered. Source
- Microsoft published a preliminary report on the November 12 global Microsoft 365 outage: a software defect introduced in July triggered lock contention in server processes, which combined with two other infrastructure issues to form a cascading failure, at one point causing 80% of web servers to become unresponsive and disrupting services such as Outlook and Teams for more than 12 hours. Source
- AMD announced the completion of its acquisition of World Labs; 李飞飞 joined AMD as Executive Vice President and Chief Scientist, reporting directly to 苏姿丰, and his team will continue to advance foundation model research and participate in the development of next-generation AI infrastructure. Source
- On November 16 at SC26, 英伟达 released the Rubin HPC GPU for scientific computing and simulation, with more than 50,000 CUDA cores per GPU, support for seventh-generation NVLink, and an expected launch in 2028; the DGX Spark HPC development kit announced the same day consists of four DGX Spark units, providing 512GB of memory and 20TB of NVMe storage as a whole, and can locally deploy models with up to 700 billion parameters. Source
- On November 17, according to The Wall Street Journal, 黄仁勋 said he would be willing to return to his post at 英伟达 if the board believed it was in its interests, but The Verge reported that the board "had not considered" bringing him back, and that the SEC had launched a civil investigation into 英伟达. Source
- On November 17, the Ministry of Industry and Information Technology agreed to carry out ministry-province collaborative pilots for 6G innovation and development in 8 provinces and municipalities: Beijing, Hebei, Shanghai, Jiangsu, Zhejiang, Anhui, Hubei, and Guangdong; enterprises involved in commercial applications must first obtain a telecommunications business operating license. Source
- Anthropic announced that starting January 1, 2027, it will tighten its coordinated vulnerability disclosure policy: the disclosure period for high-severity and critical vulnerabilities will be shortened from 90 days to 30 days, and other levels to 45 days, partly because AI has greatly accelerated the speed of vulnerability discovery and duplicate discovery. Source
Recent Updates from SSPAI
Values written like {{NAME_ab12cd34}} are placeholders for real values. Use them exactly as written, in replies and in tool calls alike: they are filled in with the real values before anything runs. Don't ask to verify them, and don't replace them with examples.
Morning Brief: China Region App Store Makes AIGC Compliance a Required Field for Listing, OpenAI Releases GPT-6.2 Astra, and More
- We have opened a dedicated page to collect and display the awards and results of the "Outstanding Author Selection." "Outstanding Author Selection" is a mechanism implemented by the editorial team since 2023. After each quarter ends, it selects existing authors with outstanding contributions during that quarter, as well as promising new authors who emerged. Winners receive an exclusive badge and prizes. View now
- Limited-time offers on 少数派 Membership, genuine software, and paid columns. Don't miss out.
Good articles you may have missed
- Community Dispatch 166 | Konica film camera, iPhone Duo experience, and a practical holiday money-saving showcase
- TDS REVIEW | Hands-on with Bose wired noise-canceling earbuds
- For Another Possibility: A few notes on tinkering with Linux gaming in China
Download the 少数派 client and follow the 少数派 WeChat official account to unlock a new reading experience 📰
Practical, useful genuine software, presented to you by 少数派 🚀