Global AI Model Releases, Official Previews & Lab Products

Alibaba Qwen GAs Qwen Intelligence: AI-phone Agent solution shipping on Honor Magic9

Alibaba officially released Qwen Intelligence at the 2026 Yunqi Conference: a modular full-stack AI-phone Agent solution built on the Qwen model family, providing three phone-optimized Agents (Mobile Planner, Mobile-Use, Mobile Creative) and four open evaluation benchmarks. Honor Magic9 (Sep 28) is confirmed as the first device shipping with Qwen Intelligence; Honor Robot Phone is also supported. Alibaba positions it as a non-hardware, open-partnership product for terminal manufacturers.

Meta Muse GA: #1 US App Store in 13 days (2.5M downloads, 642K US DAU); stock +11% on AI-agent re-rating

Meta's AI agent 'Muse' (launched Sept 8; available US & Canada on iOS + Android) confirmed GA as a Meta-owned general AI client product. Piper Sandler (CNBC) verifies strong App Store user volume and ratings, triggering investor re-rating of Meta's AI capex; Meta stock +10% on the day, +28% Sept YTD. Quantified post-launch data (Sensor Tower / Apptopia): 730 K downloads in first 5 days; 2.5 M cumulative by day 13 (iOS ≈ 1.5 M, Android ≈ 1.1 M); US mobile DAU 642 K vs ChatGPT same-period 231 K. Muse holds #1 US iOS free chart, ahead of ChatGPT, Claude, and Grok. Meta positions Muse as 'the world's first personal AI assistant for everyone': an agentic product that autonomously executes web tasks (research, form-filling, online shopping, restaurant booking, pet-care) and integrates email, calendar, Spotify, Instagram, and OpenTable; it stores user preferences and gates sensitive operations (privacy, payments) behind explicit authorization. Stock re-rated +11 % on the reporting Monday.

xAI Officially Launches Grok 4.7 (~2.1T Parameters), API Pricing at $2/$6 per Million Tokens, with Simultaneous Release on Cursor / Grok Build / Cloud Platforms

On September 22, xAI released its next-generation foundation model, Grok 4.7. The official announcement states that the model has approximately 2.1 trillion parameters, an increase of roughly 40% compared to the 1.5 trillion in version 4.6. It supports text and image inputs, features a 500K context window, and offers four inference intensity levels: low, medium, high, and xhigh. API pricing is set at $2 per million tokens for input and $6 for output. A high-speed version is also available, which doubles the output speed and doubles the price. The model is accessible via Cursor, Grok Build, third-party coding tools, model routing services, and cloud platforms, with the Grok API also opened simultaneously. In CursorBench 4.0, focusing on long-duration programming tasks, the model's price and performance rank among the leaders in its category. The company also claims that its safety protection system is at the highest existing level, achieving 62.4% on the LatchBio biosafety benchmark.

Zhipu Officially Releases GLM-5.3-FlashX (Up to 200 tokens/s): After Predecessor 'Ox Alpha' Launched Anonymously Amid Surging Demand, Built on 100,000 Domestic Chips for Inference Computing Power

Zhipu officially released the foundation model GLM-5.3-FlashX on September 18, with a maximum generation speed of 200 tokens/s. The previous version, GLM-5.3-Flash, was initially launched anonymously as 'Ox Alpha' and playfully dubbed the 'Ox Arrival' model by developers. Following rapid growth in demand, Zhipu increased investment in infrastructure and inference optimization on a computing base provided by 100,000 domestic chips, releasing the new FlashX version.

OpenAI releases Astra for Law: A legal vertical SKU based on GPT-6 Astra, integrating CourtListener case law library and 73 plugins, currently in invite-only beta

OpenAI has officially launched Astra for Law, a vertical product SKU for law firms and legaltech companies, built on GPT-6 Astra. It provides dedicated tools, settings, and context with zero data retention. It includes a legal search index covering 2.3 billion web resources and CourtListener case databases, allowing retrieval of 99.9% of public US case law. Key capabilities include finding similar cases, distinguishing dicta, mining adverse precedents, drafting pleadings, and identifying contract risks. In legal benchmarks, it achieves a 54% accuracy rate (versus 38.7% for standard GPT-6 Astra), a 24% increase in relevant precedent retrieval, and up to a 54% improvement in key section localization. The plugin ecosystem consists of 73 plugins (26 partner-built, 47 community). Users can invoke them directly in ChatGPT. The product name ties to GPT-6 Astra reasoning capabilities and auto-upgrades with model updates. Currently in invite-only beta for selected firms, with API access planned later.

Astra for Law: Frontier intelligence built for your practice. A new offering powered by GPT-6 Astra with tools, settings, and context to support the expertise and judgment of lawyers and legal technology firms.— OpenAI · @OpenAI 于 X · Sep 18, 2026, 4:15 AM

Alibaba Qwen Officially Launches Qwen3.8-Omni-Flash: The First Full-Modality Agentic Foundation Model

Alibaba Qwen officially released Qwen3.8-Omni-Flash, describing it as the lab's first full-modality foundation model centered on agentic capabilities. It natively integrates audio-visual understanding, reasoning, and tool calling, enabling content comprehension, task planning, tool execution, and result delivery within a single model.

🚀 Meet Qwen3.8-Omni-Flash, Qwen's first omni-modal model built around agentic capabilities! Native audio-video understanding, reasoning, and tool use come together in one model: understand the content, plan the task, execute with tools, and deliver the result. Highlights: 🥳 -— Alibaba_Qwen · @Alibaba_Qwen 于 X · Sep 18, 2026, 11:16 AM

Kimi Launches Financial Industry Solution (Vertical SKU Officially GA)

Moonshot Kimi officially released its financial industry solution on September 17, integrating over 10 authoritative data sources, 9 financial skills, and 5 compliance security measures to provide data modeling and report delivery capabilities. Dozens of financial institutions, including top brokerages, commercial banks, venture capital funds, and investment research data firms, are already leveraging Kimi models, Kimi Enterprise, or Kimi Work to drive business innovation or co-build financial solutions.

OpenAI ChatGPT Ads Officially Launches; Shopify Becomes First E-Commerce Partner

OpenAI’s ChatGPT Ads product has officially launched, with Shopify named as the first commerce partner. Merchants can directly install and manage ad campaigns for ChatGPT users within the Shopify admin dashboard and set their own budgets, while product catalogs are pulled directly by OpenAI. This function is OpenAI’s proprietary commercial advertising product, representing the general availability of a vertical SKU-level lab product.

Anthropic Officially Announces the Merger of Claude Cowork and Chat into a Unified Claude, with Rollout to Pro/Max Users Within Weeks

Anthropic has officially confirmed via its official channels that Claude Cowork and chat will be merged into a single Claude product: users can quickly ask questions or deliver reports, Claude will continue processing after the user closes their laptop, actively follow up on unclear points, and final decision authority remains with the user. This merged experience will be rolled out to Pro and Max subscribers within a few weeks.

Claude Cowork and chat are merging into one Claude. Ask a quick question or hand over a report, and Claude takes it from there, even after you close your laptop. If something's unclear, Claude asks—you keep the final say. Rolling out to Pro and Max over the next few weeks.— claudeai · @claudeai 于 X · Sep 17, 2026, 12:20 AM

Anthropic opens beta for 'Salesforce in Claude' vertical product (37 prebuilt sales skills)

Anthropic has officially opened the beta for its 'Salesforce in Claude' offering. This vertical sales SKU connects Salesforce customer, opportunity, and pipeline data directly into Claude conversations. It features 37 prebuilt sales skills supporting tasks like call preparation, deal review, pipeline dashboard generation, and forecast sending, all without leaving the chat interface.

Salesforce in Claude is now available in beta. It brings your accounts, opportunities, and pipeline into Claude, with 37 pre-built sales skills. Prep a call, review a deal, create a pipeline dashboard, or send your forecast without leaving the conversation.— claudeai · @claudeai 于 X · Sep 15, 2026, 11:02 PM

Anthropic ships Claude for Financial Advisors, a vertical SKU integrating BlackRock and Vanguard financial tooling

Anthropic shipped Claude for Financial Advisors, a lab-owned vertical SKU that integrates Claude with financial-analysis and risk-management tooling operated by BlackRock and Vanguard. Executives from both Anthropic and BlackRock said the system accelerates research, administrative, and portfolio-monitoring tasks for financial advisors, marking one of Anthropic's most significant moves into the financial industry.

OpenAI Officially Launches ChatGPT for Financial Services (Vertical SKU GA, with Morgan Stanley/Evercore as Design Partners)

OpenAI announced via official channels (Greg Brockman's blog) that ChatGPT for Financial Services is now available: a customized ChatGPT Work experience for investment bankers and equity research analysts, featuring integrated financial data sources such as Daloopa, PitchBook, and LSEG News, and powered by GPT-6 Astra reasoning capabilities. It supports investment research, financial modeling, and customized client materials; the service will automatically upgrade with subsequent OpenAI model updates. Morgan Stanley and Evercore participated as design partners in the product's development. OpenAI stated that the tool aims to attract large financial institutions and continue expanding the proportion of enterprise customers.

OpenAI targets work of Wall Street junior bankers with new ChatGPT for Financial Services— CNBC · @CNBC 于 X · Sep 11, 2026, 1:17 AM

OpenAI Launches ChatGPT for Clinicians (GPT-6 Astra Pro) for U.S. Certified Clinical Doctors

OpenAI has announced the launch of its vertical product, ChatGPT for Clinicians. Verified U.S. clinicians can now use GPT-6 Astra Pro for free, with an official registration entry point provided.

OpenAI GPT-Live-1 model API endpoint officially launched

OpenAI confirmed via its official developer channel (@OpenAIDevs) that GPT-Live-1 is now available on the API, enabling developers to integrate the endpoint into their applications to support real-time voice interaction agents and natural conversational capabilities.

DeepSeek officially launches V4.1-Flash base model, replacing V4-Pro and lowering prices

On September 10, DeepSeek officially released DeepSeek-V4.1-Flash, the smallest model in its new architectural series. Featuring a 552B parameter MoE architecture with native multimodal visual understanding capabilities, the model utilizes 8B activated parameters for input and 16B for output, with costs lower than comparable-sized models. The official announcement stated that this model replaces V4-Pro with reduced pricing, surpassing the flagship V4-Pro in multiple benchmark tests.

OpenAI Cuts Official Luna Model Price by ~80%

On September 7, 2026, OpenAI cut the official price of its Luna model by approximately 80%, leading to a tenfold increase in usage.

OpenAI Reveals Execution Details of Internal Model Solving Navier-Stokes Millennium Problem: 88 Hours of Exploration plus Lean Formal Verification; Model Stronger than GPT-6 Astra

On September 9, OpenAI provided execution details after announcing on September 8 that its internal model, stronger than GPT-6 Astra, collaborated with about 10,000 AI agents to solve the Navier-Stokes millennium problem. The team spent 88 hours exploring the core proof strategy, followed by a 17-hour rigorous verification using the Lean formal proof system. The project consumed approximately 130 billion tokens (nearly 300 billion in total) and cost millions of dollars, accompanied by the release of a 165-page paper and formalized code. OpenAI reiterated that the agents had no prior contact with researchers' work and did not utilize specific user data, while thanking Levent Alpöge and Tristan Buckmaster. Greg Brockman and Sam Altman had previously confirmed this as a milestone for AI and mathematics and the "first successful achievement." The Clay Mathematics Institute has not yet formally recognized the proof, and some top mathematicians have raised questions about the logic, attribution, and copyright of prior research. Additionally, the released GPT-6 Astra had previously solved five Erdős math problems at once, including the disjoint set conjecture and the Erdős-Sós conjecture. The official name, version, and API pricing of this internal model remain undisclosed.

DeepSeek Flash Series Official API Pricing Adjustment (Including Breakdown of Rate Cuts)

DeepSeek flash series API pricing adjusted as of September 10, 2026, at 12:00: off-peak input cached tokens 0.02 CNY/M (60% reduction), uncached 1 CNY/M (33% reduction), and output 4 CNY/M (11% reduction); peak hours (Mon-Fri 9:00-12:00 and 14:00-18:00) are double the off-peak rates, with cached 0.04 CNY, uncached 2 CNY, and output 8 CNY per M tokens.

Meta Officially Releases Muse Spark 1.3; Muse Code Arrives on Native Windows; Muse Personal Agent Launches on iPhone and Mac Clients

Meta has officially launched Muse Spark 1.3 via the Meta Model API and Muse Code, achieving benchmark performance comparable to GPT-5.6 Sol. At the time of release, the max reasoning configuration remains in safety testing and is not yet widely available to developers. Meta simultaneously introduced the personal agent Muse, powered by Muse Spark 1.3, which autonomously handles tasks with an interface similar to Grok Bot. Recent updates include Muse Code arriving on native Windows without requiring WSL, supporting PowerShell interaction and running in a default sandbox, along with the personal agent launching on native iPhone and Mac apps, where Mac users can directly download and install it.