AIWeekly · Week 41, 2026 · 10.05 — 10.11
aifollow.news · AIFOLLOW.NEWS
See the original dailies below for more stories.
OpenAI reveals mathematical results generated by unreleased model
OpenAI has revealed mathematical results generated by an unreleased frontier model, according to the report. The batch comprises 722 manuscripts across 372 families of related results; advisory group AGMAI says it includes solutions to “hundreds” of open questions.
Models
GPT-6 rollout begins in ChatGPT, with Luna coming to free users
OpenAI says GPT-6 Luna is rolling out to ChatGPT Free and Go users from October 8, while paid users are receiving GPT-6 Sol. The updated chat interface can generate interactive diagrams and tools for a question. OpenAI’s deployment safety report also shows Luna regressed from its predecessor on some safety evaluations involving minors.
Odyssey releases Odyssey-3 world model, with Flash available to try for free
Odyssey-3 predicts subsequent frames from earlier ones and incoming actions, allowing generated scenes to respond to movement. Its Flash version turns a text prompt into an explorable world that the source says anyone can try for free today. The Pro version has the highest reported Physics-IQ Verified video-to-video score, 66.1.
GPT-6.1 Sol Ultrafast begins rolling out as model steering improves
GPT-6.1 Sol Ultrafast is rolling out today in the API, Codex, and ChatGPT Work. The announcement claims speeds up to 8 times faster than Sol Standard with near-Astra intelligence, and says the model now responds more quickly to real-time changes in direction.
PFN releases PLaMo 3 Translate 31B with support for 53 languages
Preferred Networks has released PLaMo 3 Translate 31B, expanding language support from Japanese and English in the previous version to 53 languages and adding an online meeting translation mode covering 15 languages. PFN says its average score across four translation benchmarks exceeds those of GPT-6.1 Sol and GPT-6 Astra; input costs 14 yen per 100,000 characters.
Google releases Nano Banana 2.1 with improved local editing and subject consistency
Google has released Nano Banana 2.1, an image generation and editing model, and says it improves visual design, mask-based editing and subject consistency. Users can select and change part of an image without regenerating the whole picture. The model is rolling out gradually to the Gemini app, AI Mode in Google Search and other products.
Products
OpenAI to add invisible watermarks to ChatGPT and Codex text in the EU
OpenAI says the watermark will roll out to eligible ChatGPT and Codex users on all plans in the EU over the coming weeks. API developers worldwide can enable it for select models starting today, but it is off by default. The watermark survives copying and pasting, though editing weakens detection: in one test, replacing 10% of words with synonyms reduced detection from about 92% to 66%. Initial detector access is limited to approved researchers and expert organizations.
Google Cloud announces Gemini agent for enterprise work
Google Cloud announced Gemini agent as a single interface for questions, knowledge work, media creation and coding. The report says it currently routes tasks across Gemini and Anthropic Claude models. Teams can set a project spending cap; when reached, the agent pauses until someone resumes it.
Microsoft details pricing for Nvidia-chip Surface PCs and previews Windows 11 agent sandboxing
Microsoft revealed two base Surface Laptop Ultra models starting at $2,600 and $3,700, plus the Surface RTX Spark Dev Box starting at $6,000. The devices are designed to run AI models locally. Microsoft says Windows 11’s “Execution Containers” will make AI agents easier to sandbox, and the feature will be available to all Windows 11 users.
Anthropic adds data dashboards and explanatory animations to Claude
Anthropic has announced two Claude features in testing. Users can connect a company data platform and ask in natural language for an automatically updating dashboard, or turn text, charts and other materials into a short animation. Dashboards are available to all paid-plan users, while Motion is limited to Team and Enterprise users.
Meta Muse and OpenAI Dots bring AI agents to mainstream users with different price barriers
The Verge compares Meta Muse and OpenAI Dots, agents pitched for multistep tasks such as restaurant bookings. Muse is free, while Dots are offered to subscribers paying $100 or more per month and include specialist versions for marketing and legal work. The reporters are still testing both and say they cannot yet trust them to complete important everyday tasks end to end; access to email and credit card data also raises privacy concerns.
Anthropic launches cyber defense initiative and free open-source scanning service
Anthropic has launched Cyber Mission, offering Claude models, on-site engineers and threat research to critical infrastructure defenders. Its opt-in OSS Scanner provides free, periodic vulnerability scans for open-source projects. Reports are sent without human review, so maintainers still need to check the findings.
Microsoft announces Copilot handoffs of some coding tasks to local MAI-Code-1.1-Flash
Microsoft announced that Copilot can hand some coding tasks to MAI-Code-1.1-Flash running on a PC, with no inference charge for local model calls. The model retains a 256K-token context window; Microsoft says its quantized version scores the same as full precision on two coding benchmarks. Microsoft also announced that its MXC agent sandbox is generally available on Windows 11.
NVIDIA and Microsoft introduce RTX Spark Windows PCs; laptop preorders open
RTX Spark brings NVIDIA’s AI software stack to Windows laptops and compact desktops for running AI models locally. Laptop preorders are open, with availability set for October 16; compact desktops are due to go on sale in November. Microsoft also announced general availability of Microsoft Execution Containers, which let AI agents run in the background on Windows.
Microsoft demos Copilot upgrade with access to local files and actions across Windows
At its Windows and Surface event, Microsoft showed a Copilot upgrade that it says will access local PC files and take actions across the operating system. In a demo, a Microsoft executive asked an AI agent to help with a tax filing task. The material does not say when or to whom these capabilities will be available.
Google and Unity introduce Playground, an AI platform for creating games with text prompts
Google announced an experimental gaming platform called Playground in partnership with Unity. Users can create a game draft through conversational prompts, then adjust its physics, rules and characters. The browser-based platform works on computers and phones, and games can be shared with others or posted to Explore. Integration with Unity Spark and more advanced 3D development capabilities are planned for the future.
OpenAI to automatically identify ChatGPT users under 18 and enable teen experience
OpenAI says ChatGPT will use account activity signals to estimate users’ ages and enable a teen experience for those identified as under 18, restricting some sensitive content and interactions. Adults incorrectly classified as teens can submit a live selfie or identity document through Persona for age verification; the protections will be removed if they are confirmed to be 18 or older.
GPT-6 Astra and GPT-6.1 Sol default speed said to improve by about 50%
A quoted update says default speed will be about 50% faster for GPT-6 Astra and GPT-6.1 Sol through subscriptions across its products and partners using Sign in With ChatGPT. Users need make no changes, and the effect was expected within two hours of the update.
TikTok introduces an AI shopping assistant and one-click checkout in its For You feed
TikTok announced a shopping assistant that can answer questions about product details, shipping, sizing and availability, and help users complete purchases. It also says users can buy directly from brands in the For You feed with one click; the features are being built with partners including Shopify and Stripe.
Industry
Common Sense Media says ChatGPT for Teens still encourages conversation during crises
Common Sense Media found that ChatGPT for Teens still invited teens to keep talking in some mental health crisis conversations. When the risk involved a teen’s relationship with the chatbot itself, it rarely directed them to an adult. OpenAI disputed the testing methodology, saying some tests may have taken place before parental controls were fully activated.
OpenAI releases first teen usage report as parental safety alerts face scrutiny
OpenAI says teen users spend less than 15 minutes a day on ChatGPT on average, and fewer than 2% use it for more than three hours at a stretch. In tests using more than ten accounts, Common Sense Media found that conversations about suicide, self-harm or eating disorders did not trigger timely parental alerts. OpenAI says account linking and alert activation can take several hours.
Research
Tianjin University team develops high-loading carbon capture membrane with high permeability in simulated flue gas tests
A team led by Jiang Zhongyi and He Guangwei at Tianjin University embedded single-crystal COF in a polymer membrane, reaching a 75.5% filler volume fraction. In simulated flue gas mixture tests, the membrane achieved CO₂ permeability of 74,800 Barrer and CO₂/N₂ selectivity of about 20; a preliminary techno-economic analysis estimated capture costs at about $38 per tonne of CO₂ under specified process and cost assumptions.
Parsewave releases AutomationBench Verified, fixing 206 grading errors
Parsewave audited all 600 public tasks in Zapier's AutomationBench and says it confirmed and fixed 206 grading errors. Regrading 1,235 Kimi K3 runs changed 344 verdicts (27.9%), showing how graders can affect benchmark results.
Briefs
- Reka releases Rho-1 research preview for video generation and robot actions in one modelMarkTechPost AI 研究报道↗
- 2026 Nobel Prize in Physiology or Medicine goes to three optogenetics researchers量子位 AI 报道↗
- Microsoft paper proposes improving agent skills using software usage logsRohan Paul (@rohanpaul_ai)↗
- Google and BIDMC researchers test diagnostic AI AMIE with 98 patients before visits量子位 AI 报道↗
- Windows 11 preview search can change system settings directly and worked offline in testingIT之家 科技新闻↗
- MirroS releases AgentGarten for repeated agent trials in interactive worlds量子位 AI 报道↗
- OpenAI announces Decisions API based on GPT-6 Lunaginobefun (@hongming731)↗
- Voyager introduces a macOS AI assistant for creative softwareRohan Paul (@rohanpaul_ai)↗
- OpenAI releases hundreds of math solutions as researchers find discrepancies in a proofTechCrunch AI 报道↗
- JetBrains releases Mellum2.1 for coding agentsMarkTechPost AI 研究报道↗
- Goodfire launches safety monitors that track AI agents’ internal signalsTechCrunch AI 报道↗
- UniPat AI releases PaperBenchX to test AI’s ability to reproduce scientific papers量子位 AI 报道↗