OpenAI has stopped accepting new subscriptions to the $200 ChatGPT Pro 20x plan, according to Codex lead Tibo. Existing accounts are unaffected, while all other plans and the API remain available.
OpenAI has paused new subscriptions to ChatGPT Pro 20x, with Codex lead Tibo announcing the change to the $200 plan. According to his explanation, subscriptions of this type place the greatest strain on the system, and halting new sign-ups is intended to preserve the experience of existing users and continued access to Astra while maintaining broader access through the smallest possible adjustment. All other subscription plans and the API remain available, existing accounts are unaffected, and the team is working to expand capacity as quickly as possible. Another user found that the Pro purchase page on the ChatGPT website is already displaying a notice that subscriptions have been paused.
深度求索 releases and open-sources DeepSeek V4.1 Flash
要闻
DeepSeek has released and open-sourced DeepSeek V4.1 Flash, a native multimodal MoE model with 552B parameters and a 1-million-token context window. Its API is live under reduced peak and off-peak pricing.
DeepSeek has formally released DeepSeek V4.1 Flash and open-sourced it on Hugging Face under the MIT license, with the API and new pricing taking effect at launch. The model uses a Causal Encoder Decoder architecture, activating 8B parameters per token during prefill and 16B during decoding. It optimizes KV Cache through mechanisms including CSA2 and compresses the global KV Cache to 890 bytes per token. According to the company, this reduces HBM and SSD requirements to 1/4 and 1/8 of the previous generation, respectively. The model was pretrained from scratch on a 45T multimodal corpus with subsequent context expansion, and its reasoning intensity can be continuously controlled at integer levels from 1 to 100.
The company said results from multiple tests showed that V4.1 Flash outperformed V4 Pro in performance, cost, speed, and total completion time, and that V4 Pro will be phased out. From 12:00 Beijing time on September 14, 2026, until V4.1 Pro launches, requests sent to V4 Pro will be routed to V4.1 Flash and billed at the V4.1 Flash rate. V4 Flash and V4 Flash Vision Exp have already been discontinued, and their original model names are temporarily being routed to V4.1 Flash. The API continues to use peak and off-peak pricing, with off-peak rates set at 1/2 of peak rates.
DeepSeek Harness released dsh-v0.1.5-rc.1, the first release candidate in the 0.1.5 series. It compiles the main user- and developer-facing changes since dsh-v0.1.2-rc.1 and includes a full Changelog.
DeepSeek Harness has released dsh-v0.1.5-rc.1, the first release candidate in the 0.1.5 series. According to the official release notes, it compiles the main user- and developer-facing changes after dsh-v0.1.2-rc.1 through dsh-v0.1.5-rc.1. The release notes explicitly identify it as a candidate rather than the final 0.1.5 release and provide the Full Changelog for dsh-v0.1.2-rc.1...dsh-v0.1.5-rc.1. The source does not list individual features, fixes, or compatibility changes.
DeepSeek App and web version merge and upgrade three modes
要闻
DeepSeek’s V4.1 Flash is now integrated into the DeepSeek App and web version, where the Fast, Expert, and Vision modes have been merged and upgraded. The product update also adds a minor mode and one-tap logout.
DeepSeek has released the V4.1 Flash model and updated the DeepSeek App and web version, with the related changes now in effect. The update replaces the separate Fast, Expert, and Vision modes by merging and upgrading the three and connecting them to V4.1 Flash. On the feature side, both the App and web version now include a minor mode and one-tap logout. The source does not disclose the model’s parameters, the exact date of the update, or the detailed rules for the new features.
DeepSeek has released three code repositories. One of them, DeepSelect, provides a TopK kernel for DSA and the sampler, with official tests reporting a 2–20x speedup over torch.topk.
DeepSeek has open-sourced three code repositories, including DeepSelect, a TopK kernel implementation for DeepSeek Sparse Attention (DSA) and the sampler. DSA is used in DeepSeek V3.2, DeepSeek V4, and DeepSeek V4.1. According to the official documentation, performance testing in tests/test.py shows a 2–20x speedup over torch.topk on the same input. The test uses effective memory bandwidth because TopK performs no floating-point operations. The repository covers only some workloads, and the fastest approach depends on input dtype, batch_size, vocab_size, and topk. The listed configurations include bfloat16 with topk = 512, and float32 with vocab_size = 129280 and topk = 512. The row stride of input tensor x must be aligned to deep_select.get_stride_requirement()[0] bytes, with padding required for unaligned input. NaN checking is always enabled; with the default abort_when_nan_found=True, the kernel calls trap() and aborts. Rows with row length <= topk are not checked for NaN.
Cognition released the SWE-2 coding model. The company says it scored 50.0% on FrontierCode 1.1 Main, came within 1 point of Fable 5.1, and cost 64% less.
Cognition released SWE-2 and made it available in Devin Desktop and CLI at launch, while also beginning its rollout to Devin Web and Fusion. According to the company, SWE-2 achieved higher scores at lower cost than SWE-1.7 and Grok 4.6 on both FrontierCode 1.1 Main and DeepSWE 1.1. On FrontierCode 1.1 Main, SWE-2 medium scored above SWE-1.7 while using 58% fewer turns and costing 81% less on average. Its first substantive edit occurred at a median of step 18, compared with step 48 for SWE-1.7.
SWE-2 was post-trained from the 2.8T-parameter Kimi K3 using the training infrastructure and methodology developed for SWE-1.7. Cognition says it scaled RL to the multi-trillion-parameter range for the first time and added 5–6 points across multiple benchmarks. The added RL algorithm covers all reasoning-effort levels in one training run. Its reward function is R=S−λ_eC, where S represents rollout success, C combines inference cost in USD with rollout time, and λ_e is set to match the slope of the base model’s Pareto curve at the corresponding effort level.
杭州 partners with 智谱 to launch a citywide Coding initiative
要闻
Hangzhou and Zhipu have jointly launched the “Hangzhou Citywide Coding Plan.” The official documentation lists 7 FAQs covering V1/V2-to-V3 upgrades, plan repurchases, quarterly and annual Max purchases, review times, discount rates, and application materials.
Hangzhou and Zhipu have jointly launched the “Hangzhou Citywide Coding Plan,” but the official page does not state an announcement date or effective date. The available documentation lists 7 questions in FAQ format, covering whether users can receive a reduction when upgrading an older V1/V2 plan to V3, whether existing plan holders qualify for a reduction when buying a new version, whether users who purchased a quarterly Max plan through the campaign can also buy an annual Max plan, and how long reviews take after an application is submitted. It also asks about discount rates for individual plans and separately addresses the review materials required from people working in Hangzhou and from undergraduate, master’s, and doctoral students in the city; the provided page content does not show the answers, specific discount rates, or review time frames.
高德 launches ABot-Earth 0.7, capable of generating kilometer-scale 3D cities in 10 minutes
模型发布
Gaode’s next-generation 3D-native urban world model, ABot-Earth 0.7, has been deployed across multiple AI services. According to the company, it can generate a kilometer-scale 3D urban scene in about 10 minutes on a single consumer GPU, improving generation efficiency by 1,000 times over conventional methods.
Gaode officially launched its next-generation 3D-native urban world model, ABot-Earth 0.7, on September 11, 2026, and deployed it in AI services including 飞行街景 2.0, 导航 Live, and 避雷指南. According to Gaode, the model can turn either a single satellite image or a text description into a three-dimensional urban scene. On a single consumer GPU, it takes about 10 minutes to generate a kilometer-scale 3D urban scene, with generation efficiency improving by 1,000 times over conventional methods, and it supports continuous generation spanning planets, cities, and street-view landmarks.
Cohere releases open-source translation model North Small Translate
模型发布
Cohere and Cohere Labs released North-Small-Translate-1.0, an open-weight translation model covering 50 languages, and reported an average score of 83.6 across all languages in the WMT benchmark.
Cohere and Cohere Labs released the North-Small-Translate-1.0 machine translation model and made its weights available on Hugging Face. The model uses a sparse Mixture-of-Experts architecture with 25 billion active parameters and 218 billion total parameters. It supports 50 languages and has a context capacity of 16K input plus 16K output. According to official data, it averaged 83.6 across all languages in the WMT benchmark, exceeding DeepL, Google Translate, GLM 5.2, and Mistral Large 3. The model is available in BF16, FP8, and NVFP4 W4A16 versions under the non-commercial CC BY-NC 4.0 license and is also subject to the Cohere Labs usage policy. Commercial use requires contacting Cohere’s sales team.
GPT-Live-1 arrives on the OpenAI API with simultaneous listening and speaking and delegated reasoning
模型发布
OpenAI has added GPT-Live-1 to its API, enabling developers to use one model to build voice applications that listen, speak, and respond to interruptions while delegating deep reasoning and tool calls to backend models such as GPT-6 Astra.
OpenAI has made GPT-Live-1 available in its API for developers to build natural voice applications and business workflows. The model was previously introduced in ChatGPT and uses a single model to process voice input and output simultaneously. According to OpenAI, it can respond immediately when users interrupt, add details, or change direction during a conversation, while delegating deep reasoning and tool calls to backend text models such as GPT-6 Astra or to third-party models. Developers can also use a system prompt to adjust tone, speaking speed, and conversation style. OpenAI says GPT-Live-1 improves its handling of background noise and silence, natively provides ASR transcripts, and supports the deployment of full-duplex voice Agents in telephone settings.
OpenAI launches public beta of Agents API for hosting cloud-based Agents
开发生态
OpenAI has opened the Agents API public beta to all developers, allowing them to create cloud-based Agents with a single API call. The service carries no separate usage fee, with charges limited to the tokens and tools consumed by each Agent.
OpenAI has now launched the Agents API public beta for all developers, exposing the same harness and infrastructure that power Codex through an API for building and running cloud-based Agents. According to the company, developers can specify a task, model, tools, and runtime environment in a single call, while OpenAI runs the agent loop on its own infrastructure and handles model calls, tool use, context management, and orchestration of long-running sessions. Runtime options include an OpenAI-hosted sandbox, the developer’s own infrastructure, or sandboxes provided by partners such as Cloudflare, DigitalOcean, and Vercel. The API is powered by the open-source Codex harness and has no additional usage fee; charges come from the tokens and tools used by the Agent.
Codex has released a Unity plugin on the ChatGPT plugin page, where users can now access it. According to its staff, this is an official Unity Codex plugin, with skills provided by Unity engineers covering project setup, editor control, rendering, multiplayer, and other development workflows.
Codex has released a Unity plugin on the ChatGPT plugin page, and users can now obtain it there. According to Codex staff, this is an official Unity Codex plugin, with its skills provided by Unity’s own engineers. Its capabilities cover project setup, package management, real-time editor control, UI, sprites, tilemaps, and text, as well as shaders, rendering and post-processing, physics and AI navigation, audio, multiplayer and voice chat, localization, and Web performance.
Cursor launches Projects to coordinate Agent delegation and assign large tasks to subagents
开发生态
Cursor has launched Projects in beta, with a coordinator agent directing thousands of subagents and retaining context across months of work. The company says new users merge 30% more PRs, while users who primarily use Projects merge six times as many.
Cursor has launched Projects and is now rolling it out to all users in beta, with creation available from the left-hand navigation. Projects uses a coordinator agent to manage large development efforts that span multiple chats. According to the company, users specify goals through chat, while the coordinator does not write code itself and instead assigns research, implementation, and testing to thousands of subagents, allowing it to remain responsive to new instructions. Each Project runs by default on its own cloud computer, continues operating when the user closes their laptop, and can start a local agent when testing is required on the user’s machine. A shared set of files synchronizes across cloud and local environments to accumulate research, codebase knowledge, and work preferences. Subscriptions can also monitor a Slack channel, run on a schedule, or follow PRs, then fix CI and act when PRs open or merge based on the signals detected.
The main Projects workflows cover features, migrations, and continuous maintenance. For a feature, Agents first research the system and record their findings in shared context, after which the coordinator creates a plan and assigns implementation and testing in parallel. It can start an Agent on the user’s computer when the feature is ready for a local trial, then monitor logs and handle bug reports after release. For migrations, the coordinator applies an agreed approach incrementally across the codebase; Cursor has used this process to adopt new frameworks and replace styling systems across hundreds of PRs. One internal design-system Project scans every new PR, extracts components that belong in the design system, and adds a lint rule after detecting the same mistake twice. The company says this Project is on track to handle 20 to 100 PRs per day. It also says internal statistics after several months of use show that new users merge 30% more PRs, while users who primarily use Projects merge six times as many.
Claude Code desktop app now lets users pop out any panel into a separate window
开发生态
Anthropic’s ClaudeDevs account says the Claude Code desktop app can now pop any panel into a separate window, move diff or terminal panels to a second screen, and run multiple sessions side by side or stacked.
Anthropic’s developer account ClaudeDevs says the Claude Code desktop app now includes a separate-window feature that allows any panel to be detached from the main interface. Users can drag a diff or terminal panel to a second screen for separate viewing while Claude continues working in the main window. A detached panel can be docked back in its original position at any time to restore the previous layout. According to the account, the desktop app also supports running multiple sessions simultaneously and displaying them side by side or stacked.
TraeCode launches limited-time model discounts in September, with the Seed series 90% off
开发生态
TraeCode has launched September model discounts that are applied automatically by user status, model, and time slot, with no claim process required. Seed-Evolving and Seed-2.1-Pro cost 10% of the regular price for all users until 23:59 on September 24, 2026.
TraeCode launched limited-time model discounts in September 2026, applying the relevant price directly according to user status, model type, and usage period without requiring manual configuration. Seed-Evolving and Seed-2.1-Pro are available to all users at 10% of the regular price for two weeks, ending at 23:59 on September 24, 2026. According to the company, the former focuses on Coding and multi-step Agent tasks, while the latter is intended for long-context work and complex engineering. Seed-2.1-Turbo and Seed-Code cost 25% of the regular price for members and 50% for free users. The company says the two models are intended for frequent tasks such as daily coding and quick fixes. The production versions of DeepSeek V4-Pro and V4-Flash have added a nighttime discount, giving all users a price of 50% from 22:00 each day to 08:00 the next day. Paid users of V4-Flash receive the 50% price at all times.
First preview of Gemini API documentation launches directly in AI Studio
开发生态
Google AI Studio has integrated the first preview of the Gemini API documentation into its service. Developers can access it at ai.studio/docs and obtain Agent-ready content through URL extensions, Docs MCP, and Agent Skills.
Google AI Studio officially announced that the first preview of the Gemini API documentation is now available in AI Studio at ai.studio/docs. Logan Kilpatrick described it as a fully integrated documentation experience for both humans and Agents and said he had waited two and a half years for it. According to his explanation, this release is only the first step toward a more extensive redesign of the experience. Philipp Schmid also explained that adding .md to the end of any documentation URL provides a markdown version. The company additionally offers Docs MCP, which can be added with npx add-mcp, and Agent Skills, which can be added with npx skills add google-gemini/gemini-skills.
Nebius launches free AI Builder Program with over $400 in credits and discounts
开发生态
Nebius has launched the free AI Builder Program for anyone to join. The page title advertises more than $400 in credits and discounts, while the FAQ itemizes five tool credits totaling $250 and a Nebius Academy certification available for $1.
Nebius has launched the free-to-join AI Builder Program. The page title says the program provides more than $400 in credits and discounts, while the FAQ specifically lists $25 for Nebius Token Factory, $100 for LangChain’s LangSmith, $50 for Toloka, $50 for Tandem, and $25 for Tavily, totaling $250 across five tool credits. It also provides access to a Nebius Academy certification for $1. According to the company, the program is primarily intended for learning, testing, and experimentation, with separate options available for larger-scale and production workloads. Partner offers are subject to eligibility, availability, and each partner’s terms.
The program’s resources include open models, frameworks, dev tools, cookbooks, blueprints, and runnable examples covering inference, real-time web search, observability, evals, coding agents, RAG, and fine-tuning. Members also receive access to technical Office Hours, Discord, and community resources. The enrollment process consists of registering for free, describing the project being built, verifying an email address, and claiming credits for the required tools. According to Nebius, as a project grows, members can be connected with the Nebius Startup Program, distribution channels operated by Nebius and its partners, enterprise workshops, and hackathons.
Arena offers GPT-Image-2.5 Sunburst for free for a limited time
开发生态
Arena is making OpenAI’s GPT-Image-2.5 Sunburst image model free in Direct Mode for 72 hours, allowing users to select it by name until 8:00 a.m. Pacific Time on September 13, 2026.
Arena has launched 72 hours of free access to OpenAI’s GPT-Image-2.5 Sunburst image model in Direct Mode, with the offer ending at 8:00 a.m. Pacific Time on September 13, 2026, or 11:00 p.m. Beijing Time on September 13, 2026. During the offer, users can select the model directly by name on the platform at no charge. The access window was still open when the announcement was published. Arena said the model had just ranked first on each of the Text-to-Image, Image Edit, and Multi-Image Edit leaderboards. After the free period ends, the model will remain available anonymously in Battle and Agent Mode.
Google’s Gemini desktop app for Windows is now available to download globally for Windows 10 and 11 users. It can invoke AI over the current screen with Alt + Space to handle tasks, access information from Google apps, and generate images and videos.
Google has launched the Gemini desktop app globally for Windows 10 and 11, and users can download and install it from gemini.google/desktop. Pressing Alt + Space opens the app over the active workspace, where it can check information in documents, suggest presentation titles, write summaries, organize files, and help with homework or projects without requiring users to leave the current screen. The app also includes existing Gemini capabilities, such as assigning multi-step tasks to Gemini Spark 1, a 24/7 personal AI agent, or retrieving information from Gmail and Google Drive to create project summaries. Users can generate custom images with Nano Banana and produce videos with Gemini Omni 1. According to Google, the app runs in the background without slowing down the PC, and more native desktop capabilities will be added over time.
OpenAI has launched Data agent in ChatGPT Work. Users can query company data in natural language, generate answers and interactive dashboards, and perform actions; the integrations listed by OpenAI include seven data sources and three BI tools.
OpenAI has launched Data agent in ChatGPT Work. After installing the Data plugin from the plugin directory and connecting an account, users can chat with @Data to query company data in natural language, generate answers and interactive dashboards, or perform actions. According to OpenAI, Data agent can connect to approved data sources such as Amazon Redshift, Datadog, Google BigQuery, ClickHouse, Databricks, MongoDB, and Snowflake, import files from Google Drive and SharePoint, and build and manipulate dashboards in Tableau, Power BI, and Sigma. The company says queries retain the existing permissions of the connected account, including table-, row-, and column-level restrictions.
ChatGPT launches stock watchlists for Plus and Pro users in the US
产品应用
ChatGPT says the Finances stock watchlist is rolling out to Plus and Pro users in the United States. After users add stocks, Chat will continuously track related developments and proactively send updates; availability is currently confirmed only for those subscribers.
According to a post reshared by ChatGPT’s official account, ChatGPT is rolling out the Finances stock watchlist in phases, with current availability limited to Plus and Pro subscribers in the United States. Users can specify the stocks they want to follow in Chat. According to the company, Chat will then continue monitoring related developments and proactively send updates. Stocks can be added by giving Chat an instruction, and the official example is “@Finances add Nvidia, Microsoft and Berkshire to my watchlist.” The message confirms no availability beyond the region and subscription tiers stated above, and the feature remains in the rollout phase.
ChatGPT Library adds native integration with Dropbox, Box, and Sharepoint
产品应用
Dropbox, Box, and Sharepoint are now natively integrated with ChatGPT Library across ChatGPT and ChatGPT Work. According to ChatGPT’s official account, the feature will roll out in stages to paid users during the week of the announcement.
ChatGPT’s official account confirmed that ChatGPT Library’s built-in connections to Dropbox, Box, and Sharepoint now cover both ChatGPT and ChatGPT Work, with access being opened gradually to paid users during the week of the announcement. According to Box, OpenAI and Box are bringing enterprise content directly into ChatGPT, with Box providing access to enterprise context in a headless manner. The content remains subject to Box’s existing permissions and continues to use its governance and protection mechanisms.
OpenAI has launched ChatGPT for Financial Services for eligible financial institutions, integrating paid financial data from 4 providers; Morgan Stanley and Evercore participated in its design.
OpenAI announced the launch of ChatGPT for Financial Services, a customized version based on ChatGPT Work that eligible financial institutions can currently obtain by contacting OpenAI. The product uses GPT-6 Astra and includes paid financial data from Daloopa, PitchBook, LSEG News, and Crunchbase, covering earnings call transcripts, financial statements, and company fundamentals, with indexing and hosting handled by OpenAI. According to the company, the product supports granular citation tracing that links charts and claims to specific paragraphs and tables, and it can use a company’s own Excel, Word, and PowerPoint templates to generate valuation models, research reports, and pitchbooks. OpenAI said Morgan Stanley and Evercore participated as design partners, with the initial use cases focused on investment banking and equity research.
Google Labs makes Dreambeans free for users across the US
产品应用
In September 2026, Google Labs made Dreambeans free for all Android and iOS users aged 18 or older in the U.S. The experiment generates personalized story collections each day and can use images from a connected Google Photos account.
Starting in September 2026, all U.S. accounts belonging to users aged 18 or older can use Google Labs’ Dreambeans for free on Android and iOS. Google Labs first introduced the experiment in June 2026 to generate personalized collections of stories each day. When users connect Google Photos, Dreambeans can incorporate images of them and their loved ones into stories and provide ways to engage with that content. Users can also open any story for more information, bookmark it, or share it with others. According to Google Labs, submitted feedback will be used to improve stories generated over time.
千问 launches AI teaching tools for Teachers' Day, including automatic test creation and lesson-plan assistants
产品应用
Qwen launched multiple AI Skills for Teachers’ Day, including test-paper assembly, past-paper search, question generation, and a lesson-plan assistant. According to the company, they cover lesson preparation, test creation, and analysis and can be installed after updating the App or PC client.
On Teachers’ Day, Qwen released a set of AI Skills for teachers, available after users update their client. The lineup includes test-paper assembly, past-paper search, question generation, and a lesson-plan assistant, all accessible through the Work Assistant Skill marketplace. The supported clients are the Qwen App and PC version; users must update to the latest release before finding and installing the required Skill. According to the company, the features can handle lesson preparation, test creation, and analysis in one place, with the goal of reducing repetitive work for teachers.
Cursor launches CursorBench 4.0 with increased difficulty across the board
技术与洞察
Cursor released CursorBench 4.0 with a higher overall benchmark difficulty. Average cost per task is calculated from each model’s published per-million-token pricing and the actual number of tokens it used on the task.
Cursor increased the overall benchmark difficulty in CursorBench 4.0. Average cost per task is calculated using each model’s published per-million-token pricing together with the number of tokens that model actually used on each task. The pricing categories are input, cache read, cache write, and output, with each of the four rates applied to the corresponding token usage. According to Cursor, results are subject to variance, and small score differences may not be statistically meaningful.
Anthropic tests AI capabilities for intelligence geolocation and weapons development
技术与洞察
Anthropic’s Frontier Red Team built evaluations for tactical intelligence targeting and conventional weapons development. The intelligence component contains 200 tasks, and Mythos Preview ranked first among the tested models in both account linking and individual classification.
Anthropic said its Frontier Red Team’s new evaluations show that models are making consistent progress on simulated tactical intelligence targeting and conventional weapons development tasks. The tested open-weights models remained behind the frontier overall, typically performing between Sonnet and Mythos-class models. The tactical intelligence component includes finding people from fragmentary information, while the conventional weapons component includes engineering drones to strike moving targets. Anthropic’s Threat Intelligence Team separately documented cases in which AI was used for surveillance and conventional weapons development; according to the team, threat actors already perceive the models as useful.
The intelligence evaluation tests cross-platform account linking and the classification of individuals into categories. It uses model-generated simulated content from WhatsApp, Telegram, Instagram, and Facebook, with 200 tasks set in fictional protest scenarios in Mexico City and Kolkata: 68 easy, 68 medium, and 64 hard, all scored using F1. In account linking, Mythos Preview came closest to the theoretical maximum across all three difficulty levels. Kimi K3 performed about as well as the frontier on easy and medium tasks but fell behind on hard samples with more noise and stronger operational security by the target personas. In classification, Mythos Preview scored highest and Sonnet lowest, while K3 performed similarly to Mythos 5 and Opus 5. Anthropic also noted that the synthetic social media data does not fully reflect real-world conditions.
Kimi has formally launched an enterprise partner program to build frontline deployment engineer (FDE) teams with IT service providers and integrators and deploy AI in companies’ core operations. It named five signed partners: Teamsun, Kingsoft Cloud, Asiacom, AsiaInfo Technologies, and Chinasoft International.
Kimi has formally launched its enterprise partner program, opening it to systems integrators, software service providers, and consulting firms across industries while jointly building frontline deployment engineer (FDE) teams with partners to deploy AI in customers’ core operations. According to Kimi, partners can access its model capabilities and conduct custom development on the Agent runtime platform Kimi Hosted Agents. Project results can be turned into reusable industry solutions, with both parties jointly pursuing similar customers, while the program will also support traditional software service providers in becoming AI software service providers. Kimi will additionally provide model capabilities, engineer training and certification, and joint delivery support. The signed companies it has disclosed so far are Teamsun, Kingsoft Cloud, Asiacom, AsiaInfo Technologies, and Chinasoft International, spanning software and information services, cloud computing, infrastructure, and telecommunications.
ElevenLabs and Universal Music Group enter multiyear strategic partnership
行业动态
ElevenLabs has entered a multi-year strategic partnership with Universal Music Group, its first with a major record company. Their first AI platform is in development, enabling fans to make remixes, mashups, and personalized vocals from participating musicians’ work.
ElevenLabs has entered a multi-year strategic partnership with Universal Music Group, with the first joint product now in development. This is ElevenLabs’ first partnership with a major record company. The product is a new AI platform that will let fans use music from participating artists and songwriters to create remixes, mashups, and new renditions, while adding personalized vocals. ElevenLabs will also develop AI audio products for those artists and songwriters. According to the company, its music models have been built with artists, record companies, and publishers from the start.
OpenAI reportedly makes substantial progress on another Millennium Prize Problem
前瞻与传闻
Blogger Andrew Curran said OpenAI had told The New York Times that it made substantive progress on another Millennium Prize Problem within five days and was preparing an announcement. The related claims have not been officially confirmed.
According to Andrew Curran’s secondhand account, OpenAI is preparing to disclose progress on another Millennium Prize Problem; the company reportedly told The New York Times that the relevant research had made substantive progress within five days. Earlier, leaker leo said OpenAI was very close to completing verification of the Hodge conjecture, while either OpenAI or Anthropic was also close to solving the Birch and Swinnerton-Dyer conjecture. None of these claims has received official confirmation, and some users have questioned the vague wording of the rumors.
