ByteDance released the consumer version of the Doubao mobile assistant on September 14, 2026, adding local memory, cross-app Agent capabilities, and automated commands. It will ship on the nubia NaviX Ultra launching on September 16, 2026, with a simultaneous rollout to existing nubia M153 users.
ByteDance formally launched the consumer version of the Doubao mobile assistant on September 14, 2026. The version will be included on the nubia NaviX Ultra, which launches on September 16, 2026, and will begin rolling out to existing nubia M153 users on the same date. The earlier technical preview was released on December 1, 2025. According to the company, the consumer version upgrades the AI button, voice wake-up, and screen-based question answering. The AI button now supports fingerprint authentication, allowing the assistant to be invoked without a second unlock after the user’s identity is verified. Voice wake-up has been optimized for far-field use, indoor and outdoor noise, and rapid speech. On-screen information can now serve both as context for answers and as direct input for task execution. Examples presented by the company include using a photo of a new home to find cabinets no wider than 1.2 meters and priced at no more than RMB 1,000, adding a photographed class schedule to a calendar, and removing a white bottle from a photo through a voice instruction.
The new version also connects personal Context and local memory with Agent-based task execution. According to the company, the assistant can search local photos, text messages, and notes. Its recording feature integrates Feishu Minutes and supports speech transcription, real-time translation, speaker identification, and automatic summaries of key points. Users can also save the current screen as searchable local memory through voice input, a three-finger upward swipe, or by pressing the AI and volume buttons together. The Agent expands support for third-party App ecosystems and first-party interfaces callable by the model. It can work across Apps to add health-check precautions to notes, book a ride with an additional stop, send a vehicle plate number to a colleague on Feishu, or control a bedroom air conditioner. The system also includes task queuing and priority insertion, shortcut commands, and automated commands triggered by a time, location, or event with user authorization. ByteDance said its privacy and security approach follows the principles of reasonable necessity and user control, with tiered restrictions on the Agent’s operating scope. The company also acknowledged that GUI Agent capabilities, the cross-app ecosystem, and task reliability still require improvement.
Apple unveils next-generation Apple Intelligence, revamping Siri AI
要闻
Apple introduced Siri AI, powered by the next generation of Apple Intelligence, with its 2027 software releases. The beta is rolling out first in English, with French, Japanese, Korean, Portuguese, and Spanish support to follow.
Apple released the next generation of Apple Intelligence with its 2027 software releases and began rolling out the new Siri AI it powers in beta in English. According to Apple, Siri AI can combine personal context, onscreen content, and information from the web to answer questions and retrieve needed information from messages, emails, photos, and other content. It can also perform systemwide app actions, such as drafting an email from scratch or editing and sharing a set of photos. Users can ask follow-up questions or review past conversations in a dedicated Siri app, while conversation history privately syncs through iCloud across iPhone, Mac, iPad, Apple Watch, and Apple Vision Pro.
On iPhone, users can invoke Siri AI with the side button or by swiping down from the Dynamic Island, while iPad and Mac provide access through Spotlight. Systemwide menus also support questions about images, files, or text shown onscreen. Apple Vision Pro provides a 3D interface that can be placed in the user’s space. Apple also says AFM Core Advanced will provide more expressive voices and dictation that automatically handles capitalization, punctuation, and formatting on supported devices. Visual Intelligence covers iPhone, iPad, Mac, and Apple Vision Pro; the iPhone Camera app adds a Siri mode, while iPad integrates the capability into the screenshot workflow. Photos, Safari, and Image Playground also add editing, intelligent browsing, and photorealistic image-generation features, respectively.
ChatGPT desktop voice feature gets a roughly 60% price cut, with usage limits increased to 2.4 times
要闻
ChatGPT has cut the price of voice features in Codex and Work on its desktop app by approximately 60%, with the change already in effect. Users can now access 2.4 times the previous amount of ChatGPT Voice usage.
ChatGPT has completed a pricing adjustment for voice features in its desktop app, reducing voice prices in Codex and Work by approximately 60%. The new pricing is already in effect. The adjustment is limited to the Codex and Work voice features in the desktop app. Following the change, users can access 2.4 times the previous amount of ChatGPT Voice usage. According to its explanation, this provides more time for orchestrating tasks by voice and more usable tokens for practical work.
ChatGPT gift cards are now available in the United States. Recipients can add the card value to their account wallet and use it for eligible subscriptions, renewals, and usage credits, while redemption is limited to the US and accounts billed in US dollars.
ChatGPT gift cards have officially launched in the United States, with redemption currently available only in the US and limited to accounts billed in US dollars. Recipients can visit ChatGPT’s official redemption page to add the card value to their account wallet, then use the balance on the ChatGPT web app for eligible subscriptions, renewals, and usage credit purchases. The cards can be used to give ChatGPT to friends and family and will enter the retail channels where consumers commonly shop for gifts before the holiday season.
DeepSeek rolls out answer read-aloud to all users on the web and App versions
产品应用
DeepSeek has rolled out answer read-aloud (TTS) to all users on both the web and App versions. The feature offers four voice options: Shell, White Wave, Starfish, and Undercurrent.
As of September 15, 2026, DeepSeek had completed the full rollout of its answer read-aloud (TTS) feature on both the web and App versions, covering all users of both clients. The feature converts DeepSeek’s text responses into spoken audio, with four selectable voices: Shell, White Wave, Starfish, and Undercurrent. The rollout applies to both the web and App versions rather than a single client; all users on both platforms are covered and have access to answer read-aloud and the four listed voice options.
Anthropic has launched Claude for Financial Advisors, combining research, meeting preparation, and documentation workflows. Its plugin supports nine partners, while regulated activities such as investment recommendations still require human review and approval.
Anthropic has launched Claude for Financial Advisors to help financial advisors conduct research, prepare for client meetings, and produce documentation through connectors and workflow skills. According to the company, the plugin can be installed in a single step, with guided setup allowing advisors to choose connections to BlackRock, Charles Schwab, Addepar, Envestnet, iCapital, Orion, Wealthbox, Wealth.com, and Zocks. The new connectors cover CRMs, custodians, portfolio reporting, financial planning software, estate planning tools, and meeting capture, and can be used alongside existing Claude connectors including Microsoft 365, Salesforce, DocuSign, Box, FactSet, S&P Global, and Morningstar. BlackRock will also launch its own plugin for Advisor Center, while S&P Global and LSEG already offer their own plugins.
According to Anthropic, the skills included in the plugin can gather information, generate summaries, and draft client communications. Firms can adopt them directly from the Claude for financial advisors skill repository or adapt them to their own operations. Claude can prepare briefs, draft analyses, and stage CRM updates, but investment recommendations, client communications, compliance determinations, and other regulated activities require human review and approval. The compliance skill screens client-facing language against the SEC Marketing Rule, flags potential issues, and helps document review activities and AI-use policies. Kitces research cited by Anthropic found that a typical advisory practice spends only one-sixth of its time in client meetings. EBRI’s 2026 Retirement Confidence Survey found that more than four in 10 American workers do not know whom to approach for reliable financial or retirement planning advice.
Bolt.new launches Forge and a $9-per-month Lite subscription plan
开发生态
Bolt.new has introduced Lite, a Forge-powered plan priced at $9 per month, with access rolling out from September 14 of a year not specified in the source. Individual Pro users can receive up to 50× Forge usage for free until October 14, for which the source also gives no year.
Bolt.new’s Lite plan will begin rolling out access on September 14 of a year not specified in the source, and Forge, its underlying engine, is also scheduled to start its rollout that day. Lite costs $9 per month, and users who join the waitlist will receive an access code by email when their slot opens. According to the company, annual billing can save up to 28%. Bolt.new defines Forge as a new open-source model mode and says it will add the best-performing options to Forge as new frontier open-source models emerge. It also states that individual Pro users can receive up to 50× Forge usage for free, without waiting or using an access code, until October 14 of a year likewise not specified, on top of existing Pro features. Teams and Enterprise are not included in this research preview.
Cline launches the native desktop app Cline Desktop
开发生态
Cline has released Cline Desktop v0.0.27, a native coding agent that loads workspaces without opening VS Code and supports Apple Silicon Macs, Intel Macs, and Windows.
Cline released the standalone native app Cline Desktop v0.0.27 for macOS on Apple Silicon and Intel, as well as Windows, with all downloads available from GitHub. According to its documentation, users can point the app at a workspace and let the Agent read, edit, and run code while the interface displays the full workspace name and git branch, without requiring a VS Code window. The app uses the same open-source Cline Agent found in the IDE and CLI, supports user-provided API keys, and connects natively to Anthropic, OpenAI, Gemini, open-weight models, or a local endpoint. Users can switch models during a project or sign in with a Cline account.
According to Cline, follow-up instructions can be queued while a task is running, then edited, reordered, sent immediately, or deleted. Editing an earlier message causes the app to fork the session, restore the workspace to the corresponding checkpoint, and rerun from the revised prompt. The sidebar brings together sessions from the desktop app, CLI, extension, and scheduled runs, with options to filter them by source or mark them as favorites. Cline says new builds will ship frequently, the current version may still have rough edges, and Beta feedback will shape subsequent features.
ElevenLabs has released the ElevenLabs MCP connector. The company says one installation covers both ElevenAgents and ElevenCreative, with the latter using more than 50 models to generate voiceovers, music, images, and videos.
ElevenLabs has released the ElevenLabs MCP connector. According to the company, the same MCP also supports ElevenLabs Agents, allowing one installation to provide access to both ElevenAgents and ElevenCreative. ElevenAgents manages voice and conversational Agents within existing tools, including reviewing recent performance, creating new Agents, updating configurations, and estimating LLM costs before changes go live. Through the same connector, ElevenCreative provides access to more than 50 models and supports natural-language generation of voiceovers, music, images, and videos.
上海人工智能实验室 releases the open-source Agentic model Atria Dawn Preview
模型发布
Shanghai AI Laboratory has open-sourced the Agentic model Atria Dawn Preview. According to the lab, it uses a 744B-parameter MoE architecture based on GLM-5.2, supports a 256K context window, and is available online through an API.
Shanghai AI Laboratory has released Atria Dawn Preview, a preview of its new-generation Agentic model, with the model weights now open and online access available through an API. The model uses GLM-5.2, which has a 744B-parameter MoE architecture, as its foundation. According to the lab, it supports a 256K context window and targets research and engineering scenarios that require continuous environment understanding, tool use, and multi-step task completion, covering four areas: discovery, creation, delivery, and cybersecurity. The weights are available under the MIT license on Hugging Face and ModelScope in Atria-Dawn-Preview and FP8-quantized versions.
Anthropic and Accenture release a practical guide to taking enterprise AI from pilot to production
技术与洞察
Anthropic and Accenture jointly released a guide with seven considerations for moving enterprise AI from pilot to production. The cited research found that only 23% of executives reported sustained enterprise-wide impact, while 42% of organizations had no single owner overseeing AI costs and outcomes.
Anthropic and Accenture jointly released a practical guide for CIOs and technical leaders that sets out seven considerations for moving enterprise AI from pilot to production, along with the related decisions leadership must make. Accenture’s Pulse of Change report, published in July 2026, found that only 23% of C-suite leaders said their AI initiatives had achieved sustained, enterprise-wide impact. The guide states that pilots are often run by handpicked teams selected for enthusiasm and capability, within defined scopes and timelines, while their budgets are commonly protected, making them unrepresentative of normal production conditions. Accenture’s Tokenomics research, published in September 2026, found that 42% of organizations relied on shared IT and finance accountability without a single owner responsible for AI costs and outcomes. According to the companies, the guide draws on observations from enterprise deployments built on Claude and Accenture’s implementation experience across industries and geographies.
Anthropic details a service for analyzing the impact of scaling tests
技术与洞察
Anthropic said its CI job volume grew 25-fold in six months, while three fixes lasted only 70 days, 29 days, and less than one day. The company ultimately rebuilt its test selection service to scale horizontally with stateless workers.
Anthropic said it rebuilt its test impact analysis service after CI job volume grew 25-fold in six months. It moved much of the in-memory processing previously handled by a single process to an in-memory data store, allowing listener workers to scale horizontally in a stateless form. According to the company, engineers now ship eight times as much code per quarter on average as during 2021–2025, with Claude authoring 80% of it. The number of tests in the codebase grew tenfold, while the number of additional engineers remained limited.
According to Anthropic, the original v0 service selected tests for each change based on past performance and package relevance. Because a single writer recorded each test’s history in order, the service could not be sharded horizontally. When multiple CI jobs ran every second, the listener fell behind the PR queue, and 20 minutes of lag could prevent tens of thousands of test updates from reaching the selector. The team successively doubled the number of cores running the service and split packages into shards with independent workers; the three fixes lasted 70 days, 29 days, and less than one day. After the redesign, any listener worker could append a result to the in-memory store’s journal and continue processing without retaining local state.
Anthropic explains how healthcare organizations use Claude Tag to build human-AI teams in Slack
技术与洞察
Insight Health and Tennr are trialing Claude Tag in Slack channels that do not touch PHI; 97% of the former’s critical alerts closed without engineer intervention, while the latter completed more than 15 tickets in about one month.
Anthropic said Insight Health has operated Claude Tag for three months in engineering and support Slack channels that contain no PHI, while Tennr made it the primary maintainer of its internal recruiting portal in July 2026. Claude Tag is currently in beta and is not covered by Anthropic’s Business Associate Agreement. According to the company, some healthcare organizations use it only in channels and with connectors that do not touch PHI, while administrators can restrict the channels and systems it can access. In Insight Health’s critical-alert channel, Claude Tag reads the codebase and Linear, while a separate Agent called Zeus queries production data through a BAA-covered Claude API and masks PHI before the information reaches Slack. The two investigate alerts, create and deduplicate tickets, open draft PRs, and monitor tests before an engineer reviews and merges the changes. Since Claude Tag went live, 97% of alerts have closed without engineer intervention.
Tennr built recruiting.tennr.com with Claude Code and lets recruiters, People team members, hiring managers, and RevOps submit requests in English or as screenshots through a dedicated Slack channel, after which Claude Tag changes the code, deploys it, and reports the result. The team completed more than 15 tickets this way in about one month, including a benefits section generated from a PDF, a Dropbox Sign link, target bonus fields, and self-service administration controls. It also restricted access to a publicly exposed copy API on the day it was flagged, fixed an Ashby import issue that reset equity values on offers already sent and propagated the corrected values, and took responsibility for onboarding documentation and permission management. Enterprise organizations can receive $25,000 in credit by activating Claude Tag and connecting it to GitHub. Team organizations with at least 10 seats can receive $2,500, and both credits expire on October 1, 2026.
Sam Altman calls on frontier labs to jointly develop AI safety standards
行业动态
OpenAI CEO Sam Altman said the company will prepare safety cases before frontier reinforcement learning runs expected to materially increase capabilities, while urging other frontier AI labs to develop shared standards for misalignment, monitoring, and safety.
OpenAI CEO Sam Altman said OpenAI has added a frontier AI safety measure under which any frontier reinforcement learning run expected to materially increase capabilities will require a clear safety case before it begins. This measure is being added to the model pre-release safety processes the company has followed for years. He explained that the Responsible Scaling Policies and Preparedness Frameworks developed years ago were primarily designed for deploying completed models, while shifting the focus toward safety and evaluation during development requires new tools.
On industry coordination and regulation, Altman supported a federal framework that would establish consistent safety requirements for frontier AI, but said action need not wait for antitrust exemptions or legislation; he also welcomed ideas such as independent auditors. He called on other companies to use OpenAI’s practices as a reference and put forward their own proposals, with the aim of forming shared standards covering misalignment, monitoring, and safety. According to his explanation, pacing does not mean stopping: AI progress will continue, but at a slower rate than might otherwise have occurred, while international coordination will require government assistance.
Musk unveiled the Grok roadmap: 4.7 is expected to roughly match Opus 5.0, 4.8 is a 2.5T model trained on a new C++ stack, 4.9 may reach the level of Astra and Fable, and Grok 5 could outperform every other model.
Musk used his personal X account to describe the positioning and training plans for Grok 4.7 through Grok 5, saying Grok 4.8 would finish training and begin RL during the week the post was published. According to Musk, Grok 4.7 should be roughly comparable overall to Opus 5.0 rather than Opus 5.1, with strengths and weaknesses in different areas, while SpaceXAI still needs to fix its multimodal performance. Grok 4.8 is a 2.5T model trained with a new C++ software stack; Grok 4.9 may reach the level of Astra and Fable, while Grok 5 could outperform every other model. On the training process, Musk said the current run was unquestionably better than the previous 2.1T run using Jax, although some errors were not corrected until midway through training. He said data quality had improved over the past few months and that the next 3T run would use improved internal training software and better data, resulting in stronger performance. All of these statements were forward-looking assessments posted on Musk’s personal account.
Anthropic, OpenAI, and Google reportedly discuss forming a joint AI standards body
前瞻与传闻
The Information reported that Anthropic, OpenAI, and Google discussed jointly creating a standards body to test and audit AI technologies. The report has not been officially confirmed by any of the three companies.
According to an exclusive report by The Information, Anthropic, OpenAI, and Google discussed jointly establishing a standards body focused on testing and auditing AI technologies before Anthropic CEO Dario Amodei publicly called for greater coordination among AI developers. The report did not say how far the discussions had progressed, nor did it disclose the proposed organization’s structure or a timetable for its formation. Anthropic, OpenAI, and Google have not officially confirmed the report.
OpenAI reportedly acquires smartphone camera company Glass Imaging for over $300 million
前瞻与传闻
The Wall Street Journal reported that OpenAI has acquired smartphone camera technology company Glass Imaging for more than $300 million. OpenAI has not confirmed the deal and did not immediately respond to a request for comment.
According to The Wall Street Journal, OpenAI has acquired smartphone camera technology company Glass Imaging for more than $300 million, although the report did not disclose the deal’s completion date. The transaction has not been officially confirmed, and OpenAI did not immediately respond to a request for comment. Glass Imaging was founded in 2019 by former Apple engineers Ziv Attar and Tom Bishop. According to the company’s description, the two previously led the team that developed Apple’s Portrait mode. OpenAI is also reportedly developing its own hardware, including smartphones, headphones, and AI companion devices.
