Anthropic launches Claude Opus 5.5, adjusts pricing and subscription limits
要闻
Anthropic has launched Claude Opus 5.5, which it says costs 40% less than Opus 5 on typical workloads at default settings, generates output more than 30% faster, and raises five-hour usage limits across several subscription plans.
Anthropic released Claude Opus 5.5, the first model in the Claude 5.5 family, while changing pricing and subscription allowances at launch. Input and output cost $4 and $20 per million tokens, respectively, both 20% below Opus 5, while cache reads cost $0.20 per million tokens, 60% less. According to Anthropic, total costs on typical workloads at default settings are 40% lower, output generation is more than 30% faster, and performance on most tasks is comparable to Claude Fable 5.1. Fast mode is available in Claude Code and the Claude Platform at up to 2.5x speed, priced at $8 per million input tokens and $40 per million output tokens. Five-hour usage limits have been increased for Pro, Max, Team, and seat-based Enterprise plans, and subscription users receive one rate limit reset that can be saved and used at a time of their choosing.
Anthropic says Opus 5.5 leads its benchmarks for agentic coding, computer use, and knowledge work, while benchmark margins have become less reliable indicators of differences in real-world use. Unless otherwise noted, results used adaptive thinking at max effort with production safeguards enabled. In early testing, one tester completed a 680,000-line code migration in less than a day. In a test that optimized load times across every page of a Web App, the model succeeded in 39 of 40 attempts. Anthropic also says Opus 5.5 achieved its highest score to date in the company’s automated behavioral audit and underwent pre-release external evaluation by organizations including Frontier Design and METR. Vetted organizations can apply to the Life Sciences Verification Program, while access to the Cyber Verification Program will expand in the coming weeks. Claude Sonnet 5.5 and Claude Haiku 5.5 are also scheduled to follow in the coming weeks.
OpenAI launches GPT-6 Sol and GPT-6 Luna, cuts API prices and improves caching
要闻
OpenAI has expanded the GPT-6 lineup with Sol for complex work and Luna for cost- and throughput-sensitive tasks, while opening API access at the same time. API prices are 50% below GPT-5.6 promotional pricing, and cached inputs qualify for discounts of up to 90%.
OpenAI has released GPT-6 Sol and GPT-6 Luna and begun offering them in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise, and Edu users, while also opening API access. Free and Go users can use Luna in the desktop app. At launch, neither model was available in the standard Chat interface, and access through ChatGPT is being rolled out in stages. Sol targets complex work, while Luna is designed for everyday tasks where cost and throughput matter more; GPT-6 Astra remains the highest-capability option. Sol and Luna use some of Astra’s training methods and incorporate improvements in professional work, factual accuracy, coding, computer use, and alignment.
API pricing is 50% lower overall than GPT-5.6 promotional pricing. Sol costs $2 per million input tokens and $10 per million output tokens, while Luna costs $0.10 and $0.50, respectively. According to evaluations published by OpenAI, Sol and Luna can match or approach some more expensive models in professional work, long-running coding tasks, and computer use at a lower cost per task. Its internal factuality evaluation showed that Sol made about half as many errors as the previous generation. OpenAI says GPT-6 also has a higher default cache hit rate, offers discounts of up to 90% for cached inputs, and adds cache monitoring, diagnostics, explicit breakpoints, and warm-up features. Existing caches can also be retained where possible when reasoning effort or tool availability changes. GitHub says these improvements have reduced the prompt tokens that Copilot needs to reprocess by more than half compared with its previous baseline.
Tibo grants manual reset allowances to Plus, Pro, and Business users
要闻
Codex head Tibo said every Plus, Pro, and Business user account will receive one banked reset that can be used manually. He did not disclose when it will reach individual accounts.
Codex head Tibo said one manually usable banked reset is being added to every Plus, Pro, and Business user account, but did not specify when it will reach individual accounts. Tibo had previously promised to provide a reset on Tuesday local time. The latest update confirms that the distribution will take the form of a banked reset that users can activate manually and that it applies to the three subscription plans listed above. The source did not provide the full date corresponding to Tuesday local time.
智谱 launches a double-holiday promotion for GLM Coding Plan
开发生态
Zhipu is launching a holiday campaign for GLM Coding Plan. From September 25 to October 7, all plans will consume quota at off-peak rates throughout the day, while GLM-5.3-Flash will be available free and without limits in ZCode each night from 11 p.m. to 9 a.m. the following day.
Zhipu will change the promotional quota rules for GLM Coding Plan from September 25 to October 7, with all plans consuming quota at off-peak rates throughout the day during the campaign. The nighttime campaign for GLM-5.3-Flash will also be extended through October 7. According to the company, users can access the model free and without limits in ZCode each night from 11 p.m. to 9 a.m. the following day, while the Agent quota included with other plans will be doubled. GLM-5.3-FlashX will be offered for trial on a rolling basis within Coding Plan, and approved applicants will receive two weeks of access.
OpenRouter launches Batch API with asynchronous batch processing for over 70 models
开发生态
OpenRouter introduced its Batch API on September 22, 2026, enabling asynchronous processing across more than 70 models. Providers can complete requests within 24 hours, generally charging 50% or less of their standard per-token price.
OpenRouter made the Batch API available on September 22, 2026, adding asynchronous batch processing for more than 70 models and workloads that can tolerate variable response times. After a batch is submitted, the provider can choose when to complete the requests within a 24-hour window and generally charges 50% of its standard per-token price, sometimes less. Users submit a request list through POST /api/v1/batches and specify the endpoint shape. Chat completions, responses, messages, and embeddings are supported. They can then poll GET /api/v1/batches/:id until the status becomes completed, failed, expired, or cancelled; completed results are included directly in the same response.
According to OpenRouter, among more than 230,000 batches completed during the two-week beta, the median completion time was 7 minutes, the 90th percentile was 1.0 hour, and the 99th percentile was 10.3 hours; submission time affected latency more than request count. Batches submitted from 5:00 to 12:00 Pacific time were slower, with the slowest 10% taking 2 to 4.5 hours. At other times, the 90th percentile was under 1.1 hours, and after 18:00 it was under 50 minutes. A single-request batch took 5 to 11 minutes depending on the hour, while batches containing 1,000 or more requests took 12 to 21 minutes. Among batches with more than 100 requests submitted from 0:00 to 12:00 Pacific time, the slowest 10% took as long as 6.8 hours.
阶跃星辰 has open-sourced Step Code v0.1.0 for real-world development tasks. Developers can use it in a terminal to read, write, debug, test, validate, and deliver code, with installation available on macOS, Linux, and WSL under the MIT License.
阶跃星辰 has released the terminal development tool Step Code v0.1.0 as open source under the MIT License and provides an official installer for macOS, Linux, and WSL. The tool covers real development workflows including reading, writing, modifying, debugging, executing, testing, validating, and delivering code. According to the company, Step Code reduces token consumption through context management, tool calls, and parallel subagent execution, while also providing long-running task hosting, scheduled execution, static website publishing through StepPage, MCP, Agent Skills, plugins, and multi-Agent orchestration. The product includes only the Step provider by default, and users can sign in with a Step Plan subscription or a Step Platform API Key.
腾讯混元 launches professional-grade image generation model Hy Image3.5 preview
模型发布
Tencent Hunyuan has released Hy Image3.5 preview, a professional image-generation model supporting up to 2K output and five reference images. Enterprises and developers can access its API through Tencent Cloud services, with each 2K image priced at RMB 0.15.
Tencent Hunyuan has launched Hy Image3.5 preview for everyday life, learning, productivity, and professional visual creation, and made it available in Yuanbao, WorkRally, OnSolo, Miora, WorkBuddy, and ima. Enterprises and developers can also access the API through Tencent Cloud TokenHub, Media Processing Service (MPS), and Video on Demand (VOD). A 2K image costs RMB 0.15, with charges applied only to output images. According to the company, the model supports text-to-image, image-to-image, multi-turn conversational editing, up to five reference images, multiple aspect ratios, and output at resolutions up to 2K. Tencent Hunyuan said it improved resource utilization through architecture simplification, efficient distillation, decoupled deployment of text and image branches, and attention-operator adaptation, raising overall capabilities by approximately 30% compared with Hy Image3.0. Miora will offer a limited-time free trial from September 22, 2026, through October 7, 2026.
蚂蚁百灵 launches two 6-billion-parameter Ming-Image models
模型发布
Ant Bailin has open-sourced Ming-Image-0.1-Design, a series containing two 6B-parameter models for visual-design generation and editable layer decomposition. Its stated minimum validated setup is one GPU with at least 80 GiB of memory running in BF16.
Ant Bailin released Ming-Image-0.1-Design as an open-source series comprising two 6B-parameter models, respectively designed for visual-design generation and editable layer decomposition. According to the official documentation, the validated default and minimum deployment is one GPU with at least 80 GiB of memory running in BF16, and this configuration supports end-to-end execution of both model families. The CLI accepts either a local directory or an HF Hub repo ID through --model. A Hub model is resolved to an immutable local snapshot before any component is loaded, while --revision can pin a branch, tag, or commit. The default attention implementation is eager, with flash_attention_2 available as an option. The diffusion transformer always uses PyTorch SDPA internally, but setting --attn-implementation to sdpa causes loading to fail.
For both tasks, --prompt accepts raw text or a path to a prompt file, while --validate-only checks the checkpoint contract and task combination without loading model weights. Prompt enhancement is a preprocessing step outside infer.py and can use Ling-3.0-flash-VL or qwen3.8-27B to convert a short description into a structured specification covering layout, coordinates, hierarchy, colors, and text ownership. Layer decomposition reads the layer count from either “Decompose this image into N layers” or “Number of layers: N” in the prompt. The model also returns a leading composite/full-canvas image, which the CLI skips before saving files such as layer_01.png in sequence. According to the documentation, explicitly setting --resolution 512 speeds up decomposition, while 1024 is recommended for output quality. Text-to-image produces a square image at the selected bucket, whereas layer decomposition preserves the reference image’s aspect ratio.
PixVerse unveils real-time world model R2 with prompt-based control and editing
模型发布
PixVerse unveiled PixVerse R2, a real-time world model that it says lets users explore, control, and edit dynamic worlds with prompts, determine how stories unfold, and interact with characters capable of remembering and responding. R2 uses two core engines to handle capability and efficiency separately.
PixVerse announced the real-time world model PixVerse R2 on social media, saying users can use prompts to explore, control, and edit dynamic worlds, determine how stories unfold, and interact with characters capable of remembering and responding. According to the company, R2 uses two core engines to handle model capability and runtime efficiency separately, addressing the trade-off between real-time operation and model capability. PixVerse said real-time operation favors smaller, faster models, while broader understanding favors larger models.
Kimi Browser Extension launches, allowing users to record web actions as a skill
产品应用
Kimi has launched Kimi Browser Extension for AI Agents. It can record web actions as a skill and open pages, click buttons, fill in forms, and extract information on a user’s behalf.
The product formerly known as Kimi WebBridge is now available as Kimi Browser Extension, allowing web actions to be recorded as a skill for use by AI Agents. According to Kimi, the extension can open pages, click buttons, fill in forms, and extract information in the same way a user operates a browser, handling tedious web tasks. To begin, users click the Kimi extension icon in the browser toolbar and sign in with a Kimi account.
Volcano Engine has added a Draft mode to the Seedance 2.5 API, letting users create a 480P draft first and then produce a 1080P final video from its ID. The company says a specified workflow for a 5-second clip can cut costs by 57%.
Volcano Engine has launched Draft mode for the Seedance 2.5 API. Enterprise users can now set the draft parameter to true when creating a video generation task, while consumer creators can select “Seedance 2.5 (Draft Mode)” in Jimeng. The mode lets users generate a 480P draft, review elements such as composition and motion, and then use the draft ID to produce a 1080P final video instead of generating a high-resolution version for every attempt. According to the company, the final video reuses the draft’s generation conditions to keep shot structure and motion as consistent as possible, although local details may change. In an example involving no input video, a 5-second 1080P clip, and four attempts before selecting a segment, the company says the workflow reduces costs by 57% and nearly doubles generation speed.
Artificial Analysis adds text-to-speech leaderboards for nine languages
技术与洞察
Artificial Analysis has launched speech synthesis rankings for nine languages. In Mandarin, Inworld Realtime TTS-2 leads with 1185 Elo, while Cartesia Sonic 3.6 ranks second with 1146 Elo.
Artificial Analysis has added speech synthesis model leaderboards covering nine languages, with an option to view results by language. In the Mandarin ranking, Inworld’s Realtime TTS-2 is first with 1185 Elo, Cartesia’s Sonic 3.6 is second with 1146 Elo, and StepFun’s StepAudio 2.5 TTS is third with 1130 Elo. Alibaba’s Qwen-Audio-3.0-TTS-Plus and MiniMax’s Speech 2.8 Turbo rank fourth and fifth, respectively. Among open-weight Mandarin models, BreezeBlue’s Breeze TTS 2 ranks highest. The site also provides a public speech synthesis arena where users can vote, and it has published its evaluation methodology.
OpenAI outlines priorities for third-party safety evaluations and pledges in-depth access
行业动态
OpenAI published priorities and principles for third-party safety assessments, proposing that independent organizations in the private and nonprofit sectors gain deep access to model training, evaluation, and deployment to examine safety evidence and the effectiveness of safeguards. The document sets out 4 priorities for in-depth assessment; the available material details 2 of them.
OpenAI published priorities and principles for third-party safety assessments, proposing ongoing technical safety evaluations of models by independent organizations in the private and nonprofit sectors. Under the proposal, assessors could gain deep access to model training, evaluation, and deployment to examine the evidence behind safety claims, identify overlooked risks, and independently judge whether safeguards work. The document sets out 4 priorities for in-depth assessment; the available material details 2: safety cases and key safeguards. These cover evidence, testing, and the practical effects of safeguards in training and internal and external deployment. Assessments could run in parallel and are expected to last from several weeks to several months. Some work could take place before deployment and inform deployment decisions. The document frames this as a long-term technical safety assessment process independent of any single release, rather than testing open to the public.
阿里云 becomes a founding corporate sponsor of the Omacom Foundation
行业动态
Alibaba Cloud has become a founding corporate patron of the Omacom Foundation, pledging $1 million a year for three years, or $3 million in total. The two will also build Omarchy China for users in China and explore how Omarchy can work with the forthcoming Qwen Book.
The Omacom Foundation announced that Alibaba Cloud, as a founding corporate patron, has pledged $1 million annually for three years to support Omarchy’s development, maintenance, and promotion. The $3 million commitment matches the scale of DigitalOcean’s support. Alibaba Cloud will also help build Omarchy China, including a local CDN, hosting, a website, in-person meetups, and a community for users in China. The two plan to make Omarchy work with Qwen models and integrate it with local Chinese services, while exploring its use on the forthcoming Qwen Book. According to Alibaba Cloud, Qwen Book is intended to be an AI Agent-native computer. Its system supports ARM and x86, with an initial ARM-based version featuring Agent Engine and Continuous Context Engine; Qwen model services will be delivered through cloud-device integration. Alibaba Cloud joins DigitalOcean and Meta Superintelligence Labs as a founding corporate patron. The foundation says total support amounts to approximately $21.7 million.
TypeSafe AI pauses new user registrations for Jev amid surging demand
行业动态
TypeSafe AI has paused new registrations for Jev as demand grows, saying it needs to prioritize service quality for registered users. Existing users can still use Jev; the company says it is working to reopen registrations but has not announced a date.
TypeSafe AI has paused new registrations for Jev and has not announced a date for resuming them. The change applies only to new registrations; users who have already registered can continue using Jev as usual. According to the company, growing demand for Jev prompted the pause because it needs to prioritize service quality for existing registered users. TypeSafe AI says it is working to reopen access and hopes everyone can use Jev as soon as possible, but it has not specified when it will begin accepting new registrations again.
千问 launches Qwen Intelligence, a full-stack AI smartphone solution
行业动态
Qwen has launched Qwen Intelligence, a full-stack solution for smartphone makers that uses three types of Agent for planning, phone operation, and visual creation, and has already been deployed with Honor Magic OS.
Qwen has released Qwen Intelligence, a full-stack solution for smartphone scenarios, and is working with device manufacturers and ecosystem partners, with the solution already deployed in Honor Magic OS. It does not involve hardware production and instead consists of smartphone-optimized models, a modular platform, and scenario-specific solutions designed to help phone makers build mobile Agents that can plan, operate devices, and complete creative tasks. Its three components are Mobile Planner Agent, Mobile-Use Agent, and Mobile Creative Agent, which handle intelligent planning, smartphone operation, and visual creation, respectively. According to the company, the three solutions achieved leading results in multiple evaluations. Qwen Intelligence currently has an official website and an entry point on the Qwen AI platform.
阿里巴巴 reveals the Qwen4 roadmap, with multiple models launching soon
前瞻与传闻
Alibaba unveiled the Qwen4 roadmap at the Apsara Conference, saying the new-architecture model had entered training and four models, including Qwen4-Max, were marked as coming soon. Qwen4.5 and Qwen5 are planned to scale to 5—10T parameters.
Alibaba disclosed the Qwen4 roadmap at the Apsara Conference. According to the company, Qwen4 is based on a next-generation architecture and has entered training, while Qwen4-Max, Qwen4-Flash, Qwen4-Plus, and Qwen4-27B are scheduled to launch soon. The materials shown at the conference listed these four models as part of the Qwen4 series, with Qwen4.5 and Qwen5 included in the subsequent plan. The roadmap indicates that these two later versions are planned to scale to 5—10T parameters.
阿里QwenBook debuts in a hands-on demo, with 100 testers being recruited via the official website
前瞻与传闻
Alibaba’s Qwen Book has unveiled a live hardware demo and opened recruitment for 100 user testers. Its website says selected participants can use the commemorative edition free of charge, share one trillion tokens, and receive priority access to new developer-system versions.
Alibaba has published a live hardware demo on the Qwen Book website and started recruiting 100 user testers. Participants are required to fully test the product, complete a research questionnaire, and join co-creation discussions to provide insights and suggestions. Benefits include free use of the Qwen Book commemorative edition, shared access to one trillion tokens, and priority updates to new developer-system versions. According to the company, the device can infer user intent from content being pointed to in webpages, files, and applications, without requiring users to switch applications, copy and paste, or restate the context. YoooClaw is used to collect and organize that context. The website also lists support or capabilities covering office work, local storage, handwriting, mice, microphones, computing power, and battery life, and states that the ecosystem spans 2000+ TOOLS, operating systems including Linux, Windows, and Android, and 1000+ ecosystem partners.
