DeepSeek opens access to TTS, ASR, OCR and other APIs
要闻
DeepSeek now provides conversational features including web search and file input through three APIs—Messages, Responses and Chat Completions—and offers an open-source demo chat client that developers can start by entering an API key.
DeepSeek now provides conversational features in the Messages API, Responses API and Chat Completions API; its documentation specifies deepseek-flash but gives no launch date. According to the company, these features cover web search, deep thinking search, file input, URL reading and basic risk controls for model content safety, matching those in DeepSeek Chat on the web and in the App. Searches and URL reading run on the server, while the client only needs to display the returned content blocks and send them back unchanged in the next turn. The accompanying open-source demo client can be started by entering an API key and can import history exported from DeepSeek Chat to continue conversations. The company says API keys, conversations and uploaded images are stored only in the browser, and the client is for demonstration purposes only. The documentation also explains that max_tokens, called max_output_tokens in the Responses API, limits the output of each inference step within a turn rather than the turn’s total output.
OpenAI has introduced ChatGPT Wallet in partnership with American Express, Mastercard and Visa, enabling its dot Agent to shop on users’ behalf. Initial access is limited to Pro subscribers aged 18 or older in the United States, with no transaction commissions charged initially.
OpenAI announced that ChatGPT Wallet is available from its launch day to Pro subscribers aged 18 or older in the United States, supporting purchases made by dot Agent on users’ behalf. According to the company, the feature works with American Express, Mastercard and Visa, and will initially charge no transaction commissions. Users enter their card details once, but must review the items and total price before each purchase and authorize the transaction using Face ID or fingerprint verification. The company says the payment network generates a dedicated payment token containing a restricted-use card number, security code and expiration date, so neither the Agent nor the merchant can access the user’s actual card details. It also says payments made with existing credit cards retain their rewards points and benefit protections.
Google Antigravity head Varun Mohan announced that the platform is resetting Gemini usage quotas for all users to mark a week of concentrated releases, and encouraged users to build projects with swarms.
Google Antigravity is resetting Gemini usage quotas for all users, an action announced by its head, Varun Mohan. According to Mohan, the reset marks a week of concentrated releases; he also encouraged users to build projects with swarms. The report comes from 橘鸦 AI 早报, which provided this source link: https://x.com/_mohansolo/status/2121380918355234957 .
腾讯 unveils Hy Realtime, a real-time speech model in the 混元 family
模型发布
Tencent has released its real-time voice model Hy Realtime in Yuanbao voice calls and Tencent customer service, adding Agentic tool calling and singing. It can use conversation context to decide whether to search the web, while Tencent Cloud API access is still pending.
Tencent has released the Hy Realtime real-time voice model and deployed it in Yuanbao voice calls and Tencent customer service. The source does not specify a release date, and Tencent Cloud API access is scheduled to open soon. The new Agentic tool-calling capability uses conversation context to decide whether a web search is needed, handling search decisions, tool calls and answer generation together. Another addition is the ability to select and sing songs based on combined requests specifying singers, genres, languages and other criteria, jointly developed by the Hunyuan team and Tencent Music’s TME Lingyin AI team. According to the official announcement, the model upgrades speech understanding, instruction following and human-like expression, and outperforms the previous generation in audio quality, pronunciation, prosody and emotional expression.
DeepSeek Harness adds an experimental conversation import plugin
开发生态
DeepSeek has added an experimental conversation import plugin to DeepSeek Harness, allowing conversations exported from its official website to be moved into a workspace and continued. The plugin is disabled by default, restricts imports to the owner, and can write up to 1000 conversations per import.
DeepSeek added an experimental conversation import plugin in DeepSeek Harness version dsh-v0.2.2-rc.3. The plugin must be enabled in the experimental plugins section of the plugins page, and import permissions are restricted to the owner. After a user uploads the ZIP file produced by exporting all conversations from the official website, recent conversations are written to the selected workspace. Their original titles, timestamps, thinking, and web search cards are retained, and users can continue any of the conversations. Each import can write up to 1000 conversations. The uploaded archive must not exceed 512 MiB, and each entry must not exceed 256 MiB after decompression.
Antigravity has introduced Dynamic Workflows, which lets an Agent generate and run Python scripts to coordinate multiple Agents for parallel processing and staged execution. Tasks can be launched with /run-workflow in both the App and CLI.
Antigravity has introduced Dynamic Workflows to coordinate multiple Agents through scripts in its App and CLI; the official documentation does not specify a launch date. According to the documentation, when a user enters /run-workflow followed by a task description, or requests a workflow in natural language, the Agent reads the built-in run-workflow skill and generates a Python script. The script is presented for approval before execution, although permission presets that already allow execution, such as Turbo, can bypass this step. Users can also reject execution and request changes. The documentation lists per-file reviews, per-package audits, and writing a failing test before fixing code as applicable tasks. The workflow runs in the background, allowing users to continue the same conversation while inspecting phases, logs, and each Agent’s status. Each child Agent has its own conversation and cannot see the script or other Agents, so the script supplies the necessary information through its prompt. Once execution finishes, the main Agent reads the logs and output and reports the results. Stopping the parent conversation also stops all workflow Agents that are still running.
Antigravity makes Orchestrator available to Pro users
开发生态
Antigravity now offers Orchestrator (Argon) to paid Google AI Pro users and trial users with eligible premium devices, allowing 2 Gemini 4 models to divide planning and execution within the same task.
Antigravity has made Orchestrator (Argon) available to paid Google AI Pro users and trial users with eligible premium devices; the source does not specify a launch date. This option assigns planning and reviews to Gemini 4 Argon and implementation to Gemini 4 Hydrogen. Google AI Ultra includes both Orchestrator (Argon) and Orchestrator (Carbon), which uses Gemini 4 Carbon for planning. According to the official explanation, the models automatically hand control back and forth within the same task rather than classifying a request upfront and assigning the entire request to one model. The company says this collaboration reduces completion time and token usage compared with using the planner model alone.
Orchestrator maintains a single Agent in the conversation and switches between planning and execution within each turn; it is not a mechanism for running multiple Agents in parallel. It has a fixed configuration, so users cannot select a thinking level, and the CLI options --effort and /effort do not apply. Users can select Orchestrator through the model selector beneath the prompt box or through /model in the CLI; selections made through the CLI interactive picker persist across sessions. To run every step on one model or specify a thinking level, users can directly select Gemini 4 Carbon, Gemini 4 Argon, or Gemini 4 Hydrogen, subject to plan availability.
Google AI Studio now lets users claim monthly Cloud credits included with their subscriptions
开发生态
Google has integrated Google Developer Program credit redemption into AI Studio, allowing AI Pro and Ultra subscribers to claim monthly Cloud credits and use them for the Gemini API without switching websites.
Google has launched in-workspace redemption and management of monthly Google Cloud credits included with subscriptions in AI Studio. Google AI Pro and Google AI Ultra subscribers can configure billing, link a billing account, and check their credit balance directly in the workspace. According to Google, the credits are provided through the Google Developer Program and can be used for the Gemini API to test prompts, build AI applications, and deploy them from AI Studio, with monthly allocations varying by subscription tier. Access to developer benefits requires an active Developer Program profile; users without one will have a profile created automatically when they redeem credits through the new workflow. The Program is free to join and brings together tools, learning paths, and benefits across platforms including AI Studio, Google Cloud, Firebase, and Android.
Anthropic is rolling out Projects in stages, introducing it to Claude Code and upgrading existing project features in chat and Cowork. The company says users can join a waitlist for a chance to gain access earlier.
Anthropic is rolling out Projects in stages, with the update covering Claude Code and existing projects in chat and Cowork. The company describes Projects in Claude Code as a new feature and the changes in the other two environments as upgrades. According to the company, users can submit requests as they arise, and Claude will pick up each one, coordinate the different work threads, and report progress along the way. Access requests are currently being accepted through a waitlist, which the company says offers an opportunity to gain access earlier.
Cognition makes free SWE-2 access a permanent benefit for Devin subscribers
开发生态
Cognition has updated Devin Teams, removing Desktop seat charges and making free SWE-2 a permanent benefit of Teams, Pro, and Max. Each user can run 2 concurrent Free sessions, while administrators gain SSO and cost controls.
Cognition has announced an upgrade to Devin Teams that lets everyone on a team use Devin Desktop without paying Desktop seat fees. The update introduces Free mode, powered by its own SWE-2, with 2 concurrent free sessions per user and the option to pay for more. Free access to SWE-2 in Devin Desktop and CLI will be a permanent benefit of Teams, Pro, and Max. According to the company, SWE-2 is its most capable coding model to date, performs close to frontier models at a fraction of their cost, and starts work faster by focusing on the parts of a codebase relevant to the task.
The administrative features cover authentication and usage limits. Teams can connect Okta, Microsoft Entra ID, or other SAML or OIDC identity providers and disable email login after connecting SSO. Administrators can set user usage limits across up to 3 tiers, with requests for additional usage either requiring administrator approval or being approved automatically within a specified maximum. Devin Coach provides feedback before engineers submit a prompt, such as suggesting a lighter mode for a small task. Teams customers can also request a standard Business Associate Agreement (BAA) through Settings in the product without an enterprise contract; the source states that a BAA must be signed before a vendor accesses protected health information. Eligible teams can apply to the startup program for up to $65,000 in credits and other support.
Command Code announced an increase in the GLM-5.3 allowance on its GOAT service from $20 in credits to $70 in credits, stating that usage on the $10 monthly plan would rise to 7 times its previous level.
Command Code announced a higher usage allowance for GLM-5.3 on its GOAT service, setting the allowance at $70 in credits, up from $20 in credits. According to the company, the $10 monthly plan now provides 7 times the previous usage. The announcement also covered context capacity and allowances across tiers: the company said GLM-5.3 supports a 1M token context on the service, while the Pro and Max tiers provide additional allowance.
Vercel Token Pass offers $150 in credits for $30 a month
开发生态
Vercel has increased the monthly tokens credit included in Token Pass from $80 to $150, while keeping the price at $30 per team per month. The credit covers selected open-weight models served through AI Gateway.
Vercel has changed the tokens credit included in each Token Pass billing period from $80 to $150, keeping the price at $30 per team per month; the source does not specify an effective date. According to the company, the credit applies to selected open-weight models served through AI Gateway, including Kimi K3, Kimi K3.1 Pro, GLM 5.3, GLM 5.3 Flash, and DeepSeek V4.1 Flash. The credit resets each billing period, and unused credit does not roll over.
What happens after the credit runs out depends on the credentials used for requests. According to the company, requests using standard AI Gateway keys or OIDC tokens fall back to the team's AI Gateway balance, while requests using Token Pass-scoped keys stop without incurring overage charges. Token Pass is an Early Access add-on for Pro teams, with limited spots and a waitlist once they fill. Teams can subscribe in the AI Gateway dashboard and connect a coding agent by running vercel token-pass in the Vercel CLI; agents that support AI Gateway can also use a Token Pass API key.
llama.cpp releases v1.0.0, its first stable major version
开发生态
llama.cpp has released its first stable major version, v1.0.0, and specified that future API and ABI incompatibilities in llama.cpp and ggml will be marked by major-version changes, giving downstream applications that depend on their library interfaces explicit version boundaries.
llama.cpp has released library version v1.0.0, with future incompatible API and ABI changes in llama.cpp and ggml to be reflected in major-version increments. According to the team, many downstream applications have come to depend on its library interfaces in the 3.5 years since the project began, making both API flexibility and stability necessary. llama.cpp currently supports more than 150 model architecture variants and all mainstream speculative decoding methods. Its underlying machine learning engine, ggml, supports hardware through dedicated backends, each of which can be compiled as a dynamic library and optionally loaded at runtime. On development practices, the team says it is incorporating LLMs into its development cycle while continuing to prioritize code quality, control over the architecture, and an understanding of implementation details. The team is also using HF Buckets and HF Jobs to expand CI capacity and, with hardware support from NVIDIA, AMD, Qualcomm, Arm, IBM, and Intel, is adding self-hosted runners across all major hardware platforms.
Claude Duet invites everyone on its US waitlist to the research preview
产品应用
Claude Duet has sent research preview invitations to all US Pro, Max, and Ultra subscribers on its waitlist at the cutoff for this round. The preview includes 3 usage-limit resets for Pro subscribers and 1 each for Max and Ultra subscribers.
Claude Duet product lead karan ahuja announced on X that research preview invitations had been sent to all US Pro, Max, and Ultra subscribers on the waitlist at the cutoff for this round; the source provides no specific announcement or cutoff date. According to ahuja, the preview gives Pro users 3 usage-limit resets and Max and Ultra users 1 each to allow room for testing. The waitlist remains open for registration. Regarding availability outside the US, ahuja said expansion to more countries is planned soon and user feedback is still being collected, but did not announce specific countries or launch dates.
Google integrates Wallet financial insights into Gemini
产品应用
Google announced the rollout of a Google Wallet connection for all Gemini app users in the U.S., allowing responses to draw on their spending habits. Users can also link credit card and bank accounts through Plaid for additional financial insights.
Google announced that it is beginning to roll out Google Wallet connectivity to all Gemini app users in the U.S. across web, Android and iOS; the source does not specify a calendar date for the launch. Users can choose to connect Wallet under Connected Apps in Personal Intelligence settings and use @Google Wallet in conversations to ask about personal finance, budgeting and related topics. According to Google, Gemini can then draw on spending habits to generate responses. Users can also link credit card and bank accounts to Wallet through Plaid for additional financial insights.
Google says Gemini has read-only access to account information and cannot move, manage or invest users’ money. Users decide whether to connect and can turn the connection on or off at any time. Google also introduced Memory Summary, which brings saved preferences, interests and ongoing goals together for users to review, edit or remove. The feature is available on web and mobile to all Gemini app users where Personal Intelligence is available, under Profile > Personal Intelligence > Memory Summary.
Meta’s personal AI agent Muse now offers outbound calling to paid subscribers in the United States, placing calls and waiting for connections for tasks involving insurance, cable television, motor vehicle offices and restaurant reservations. The company says the feature is intended to give users back some free time.
According to Juya AI Daily’s report dated November 14, 2026, Meta’s personal AI agent Muse has introduced outbound calling, with access currently limited to paid subscribers in the United States. Users can delegate calls involving insurance, cable television, motor vehicle office matters and restaurant reservations to Muse, which places the calls and handles the time spent waiting to connect. Meta Chief AI Officer Alexandr Wang said many users had previously requested the feature. The company described the purpose of the launch as helping users reclaim some free time.
Anthropic unveils Anvil, an automated red-teaming system
技术与洞察
Anthropic is using its automated red-teaming system Anvil to test new safeguards before deployment. According to the company, it matched or surpassed the best human jailbreak results on all 8 questions in its biological safety tests, taking 24–48 hours.
Anthropic has developed Anvil, an internal automated red-teaming system, and is using it to check the robustness of new safeguards before deployment; the source does not specify a publication date. The system searches for jailbreaks using frontier internal versions of Claude with safeguards disabled. According to the company, across the 8 questions in its standing biological safety bug bounty, Anvil matched or exceeded the best human results against every classifier tested, finding those jailbreaks in 24–48 hours rather than weeks to months. In cybersecurity tests, it induced a safeguarded model to write complete, working exploits for known historical vulnerabilities in the JavaScript engines of Chrome and Firefox, achieving arbitrary code execution. Against anti-distillation classifiers, Anvil produced its first substantial reasoning leak after 435 requests and 1.2 hours, while experts needed 102,200 requests and 27.9 hours to find a similar jailbreak. The company says vulnerabilities found by Anvil are used for pre-deployment fixes, building static evaluations, and conducting adaptive evaluations.
OpenAI disbands its standalone robotics team, moving staff to its main research team
行业动态
OpenAI has dissolved its stand-alone robotics team, moving researchers into its main research team and applied robotics and hardware staff into Applied. The company attributes the change to general-purpose models such as GPT-6 Astra being directly applicable to robotics, while retaining a 2027 target for the first public demonstration of its own humanoid robot.
According to Fortune’s report dated November 13, 2026, OpenAI has confirmed the dissolution of its stand-alone robotics team without disclosing the exact date of the reorganization. Robotics researchers have moved to the main research team under Chief Research Officer Mark Chen, while the applied robotics and hardware teams have joined the Applied organization under President Greg Brockman. The company says general-purpose frontier models, including GPT-6 Astra, have demonstrated capabilities that can be applied directly to robotics, prompting it to integrate the relevant researchers into general-model teams. It says robotics remains a core part of its long-term vision.
The reorganization has not changed the company’s disclosed plan to develop its own humanoid robot, with CEO Sam Altman still targeting a first public demonstration in 2027. Experiments involving general-purpose models controlling robots have also produced relevant test data. In an Nvidia study, researchers supplied Astra with camera images, robot information and a control interface. An Agent powered by the model achieved a 74.8% success rate on a simulated manipulation benchmark, compared with 51.0% reported by the best specialized robotics models on that leaderboard. The comparison involved simulated robots in a digital environment, not physical robots.
Anthropic outlined an IPO of Class A common stock on Nasdaq under ANT, with pricing planned for November 19, Japan time; no year was specified for the schedule. The company also disclosed that its net loss includes US$21.9 billion in non-cash charges.
Anthropic announced plans for an IPO of Class A common stock, with trading expected to begin on Nasdaq under ANT on November 19, U.S. time, without specifying the year for the schedule. According to the company, the roadshow is scheduled to start on November 11, Japan time, and the final offering price is expected to be set on November 19, Japan time, based on investor demand and market conditions, with pricing and listing arrangements subject to regulatory approval. IPO participants need a brokerage account and are not guaranteed to receive all or any of the shares they request; purchases after trading begins will be at the prevailing market price. In its financial disclosures, the company said its rising valuation resulted in US$21.9 billion in non-cash charges included in net loss, related to changes in the fair value of convertible promissory notes, derivative liabilities and SAFEs. It also described a calculation that divides the ratio of revenue to compute and infrastructure costs for the third quarter of 2026 by the same ratio for the first quarter of 2025, but the supplied source does not provide the result.
微软 apologizes for widespread Microsoft 365 service outage
行业动态
Multiple Microsoft 365 services were disrupted for nearly a day and a half, also affecting Anthropic’s Claude Microsoft 365 connector. Microsoft apologized and said it had addressed the root issue and was restoring service in stages for the remaining customers and regions.
Microsoft 365 experienced a service outage lasting nearly a day and a half, affecting Exchange Online, Outlook, SharePoint and some Teams services, and Microsoft issued an apology in a post. The disruption also affected the Microsoft 365 connector functionality in Anthropic’s Claude. Microsoft attributed the outage to a problem in an earlier network capacity upgrade that surfaced after subsequent network changes, saying it was unrelated to cyberattacks, AI or agents. According to the company, the scale of demand required a phased restoration, with stability checked at each stage. Microsoft said it had addressed the root issue and was continuing to restore service for the remaining customers and regions, with a full incident report to be published after its review is complete.
Sierra releases Poppy 1.0, an open standard for personal agents
行业动态
Sierra has released Poppy 1.0, an open standard for personal Agents incorporating feedback from five weeks of partner testing, which companies can implement independently. All Shopify stores now automatically publish a Poppy file, requiring no additional implementation by merchants.
Sierra has released Personal Agent Protocol, or Poppy, 1.0. Companies can implement the protocol independently to let personal Agents interact with them through business Agents, websites, and APIs. According to the company, this version incorporates lessons from five weeks of design partners testing drafts against real websites, APIs, and Agents. The protocol is built largely on existing standards including MCP, OpenAPI, and UCP, and uses OAuth authentication to let customers control personal Agents’ permissions and manage their access over time. The official implementation kit includes setup instructions and a validator for checking implementations.
All Shopify stores now automatically publish a Poppy file, with no additional implementation required from merchants; these stores also support Universal Commerce Protocol (UCP) by default. The two protocols serve different roles: UCP handles product discovery, adding items to a cart, and checkout, while Poppy handles Agent identification and permission checks before shopping begins. Shopify co-authored the Poppy commerce extension connecting the two protocols. It supports Agents shopping as guests or signing in to customer accounts, and requires Agents to obtain customer permission before performing certain actions. The extension also connects to UCP Ask, allowing a shopper’s Agent to communicate directly with a store’s Agent.
LangChain raises $300 million at a $3 billion valuation
行业动态
LangChain raised $300 million at a $3 billion valuation in a round led by existing investor CapitalG. The company says it aims to help businesses control their own Agent systems, with more than 7,000 organizations already using its LangSmith platform.
LangChain announced that it raised $300 million at a $3 billion valuation; the announcement did not specify when the financing took place. Existing investor CapitalG led the round, with participation from new investors GIC and DFJ Growth and existing investors IVP and Sapphire Ventures. The company also disclosed participation from Salesforce Ventures, Insight Venture Partners, QuantumLight, and MongoDB Ventures, alongside additional investment from Sequoia, Amplify, and Benchmark. According to the company, its goal is to let businesses choose their models, control their data and infrastructure, and continuously improve Agents while retaining governance capabilities.
According to the company, more than 7,000 organizations have adopted its framework-agnostic LangSmith platform. Its capabilities are organized into Agent Runtime, Agent Improvement, and Agent Governance, covering execution and deployment, testing and monitoring, and centralized management of models, tools, and data. The company says its built-in Agent, LangSmith Engine, can identify errors, write code and prompt fixes, and monitor production traffic for regressions. Its cited use cases include Clay monitoring hundreds of millions of Agent runs per month and Workday running and governing Agents across its organization while reducing deployment time for a new Agent from five months to under a day.
Four US nonprofits sue five AI companies, seeking a pause in development
行业动态
Four U.S. nonprofits have sued five AI companies, asking a court to halt the development or deployment of new models. The plaintiffs claim the companies cannot prevent catastrophic harm from their systems and allege that their operations endanger public health and interfere with the use of property, including computer systems.
Four U.S. nonprofits have filed a lawsuit in San Francisco Superior Court asking a judge to prohibit Anthropic, Google, Meta, OpenAI and SpaceXSI from developing or deploying new AI models; the source does not give the specific filing date. The plaintiffs are the Irreplaceable Education Fund, Public Citizen, Heat Initiative and the Alliance for Secure AI. The complaint alleges that the companies have violated California’s unfair competition law, citing threats to public health and interference with the use of property, including computer systems. The plaintiffs also say they have had to divert resources from other planned work to address the threat of harm posed by the companies’ advancing AI models.
The lawsuit draws on documented cyberattacks and other harmful actions by AI systems, as well as the companies’ own statements about technological risks. The plaintiffs argue that efforts to build AI with broadly human-level intelligence and an increasing ability to perform tasks and learn autonomously could result in a permanent loss of human control over the technology and increase existential risks to humanity. These are the plaintiffs’ allegations and risk assessments, not court findings described in the report. According to the report, the defendant companies did not immediately respond to requests for comment.
SpaceXAI has cut most of the development staff for its video and image generation tool Grok Imagine, affecting about 60 people and leaving only a small group to maintain basic operations, according to The Information. The tool remains available for normal use.
SpaceXAI has cut most of Grok Imagine’s development staff, according to The Information, citing people familiar with the matter; the report did not specify the date of the layoffs. About 60 team members working on the video and image generation AI tool were affected, with only a small group remaining to handle basic operations, while the product remains accessible and usable as normal. The Information interpreted the staffing change as an indication that the company is shifting its focus toward AI coding tools and agent products.
