智谱 officially releases and open-sources GLM-5.3-Flash
要闻
Zhipu has released GLM-5.3-Flash under the MIT License, with weights now available on HuggingFace. The model has 320B total parameters and 18B activated parameters, and supports text, image, and video inputs with a 1M context window.
GLM-5.3-Flash has been formally launched by Zhipu, with model weights available from HuggingFace under the MIT License. It has 320B total parameters and 18B activated parameters, natively accepts text, image, and video inputs, and provides a 1M context window. According to the company, it is the first open-source frontier model to combine sparse attention with linear attention, while manifold-constrained hyperconnections are used to improve scaling. IndexPool compresses the indexer's 4 cache vectors into 1 to reduce latency and memory use at a 1M context length. Visual capabilities are integrated into the Coding loop, allowing the model to coordinate across code, browsers, and graphical interfaces for front-end development, game creation, Blender 3D scenes, and real-world operations driven by BUA and CUA. Thinking mode cannot be disabled, and thinking.type supports only enabled.
Before the formal release, Zhipu tested the model on OpenCode and OpenRouter under the codename Ox Alpha. The company said GLM-5.3-Flash scored 57 on Artificial Analysis Intelligence Index v4.1.1 and outperformed GLM-5.2 across 6 coding and agentic benchmarks. Zhipu said this was its first use of a domestic-chip cluster to serve large-scale traffic, and that domestic chips also supplied all compute during the Ox Alpha test. Compared with the initial baseline on the same hardware, end-to-end serving performance increased by 3 times, while hardware efficiency and cost per token reached levels comparable to mainstream NVIDIA GPUs, according to the company. Under the GLM Coding Plan, the model's available quota is 3 times that of GLM-5.3, and user quotas have been reset. The regular domestic API price is one-tenth that of GLM-5.3, with a 50% discount offered for a limited period of 2 weeks.
Alibaba’s Qwen released Qwen3.8-Flash and opened the Qwen3.8-Flash-Next weights. The model activates 6B parameters per Token, natively supports a 262K context, and charges 1 yuan and 3 yuan per million Tokens for input and output, respectively.
Alibaba’s Qwen has launched Qwen3.8-Flash on the Qwen AI platform and simultaneously published the weights of the open-source Qwen3.8-Flash-Next on Hugging Face and ModelScope. The latter is positioned as an early preview of the new Qwen4 architecture. The model backbone contains 125B parameters, with an additional 51B N-gram Embedding, and activates 6B parameters per Token. Its native context length is 262K and can be extended to 1M with YaRN. The architecture changes cover Attention, Residual, Embedding, and Optimization, introducing GDN+QSA hybrid attention, Gated Residual, N-gram Embedding, and Muon Optimizer. According to the official statement, its training cost is about 1/9 that of Qwen3.7-Plus, with stronger performance in coding, office, and multimodal Agent tasks. Input and output on the platform cost 1 yuan and 3 yuan per million Tokens, respectively.
OpenAI says 5.5-mini misrouting issue has been fixed and apologizes
要闻
OpenAI staff member Adam Fry said a routing regression had mistakenly sent about 3% of Pro and Thinking requests to 5.5-mini. The issue has been fixed, and he apologized for it.
When Adam Fry posted about the incident, he said OpenAI had fixed the routing regression and apologized for it. According to his explanation, the system had previously sent about 3% of Pro and Thinking requests to 5.5-mini unintentionally. Before that, multiple users had reported that selecting a higher reasoning effort in Chat mode caused their requests to switch to GPT-5.5-mini without notice.
Google releases Gemini 3.5 Transcribe speech-to-text model
模型发布
Developers can now use Google’s Gemini 3.5 Transcribe speech-to-text model through its services; according to Google, its time to final transcription is 70% shorter than Chirp 3’s.
Google has released Gemini 3.5 Transcribe and made it available to developers through the Gemini API in Google AI Studio and the Gemini Enterprise Agent Platform. The model is intended for voice agents, real-time captioning, and post-call analysis, converting raw audio directly into cleaned-up and formatted text. According to Google, it supports live language switching, streaming transcription, multi-speaker attribution, word-level timestamps, custom vocabulary recognition, and the removal of speech fillers.
Citing measurements from Artificial Analysis, Google said the time to final transcription is 70% shorter than with Chirp 3. On the FLEURS benchmark covering major languages and locales, the model recorded a WER of 5.50% in streaming mode and 5.04% in non-streaming use cases. According to Google, the model is also used in Gboard, Antigravity, the Gemini app, Chrome, Rambler on Android, and the Gemini app on macOS. Agora, Fishjam, LangChain, LiveKit, Pipecat, Vercel, and Vision Agents provide related development and deployment capabilities through the Gemini Live API.
Antigravity CLI launches voice mode for conversations with agent
开发生态
The latest Antigravity CLI adds a voice mode that lets users talk to an agent by running /voice or pressing F5. Powered by Gemini 3.5 Transcribe, it also supports using a local microphone over SSH.
The latest Antigravity CLI now lets users speak directly with an agent, with the feature accessible through the /voice command or the F5 key. According to the official announcement, both the CLI voice mode and the microphone button in the Antigravity interface use the new Gemini 3.5 Transcribe speech-to-text model, whose transcription automatically removes filler words such as “um” and “uh.” The /voice command also adds a mic-serve subcommand for using a local microphone over an SSH connection.
Claude Code can now automatically draft feedback reports and send them after user approval
开发生态
Claude Code can now automatically draft a feedback report when a run fails, it detects its own error, or a user flags a problem, then send it to Anthropic after the user reviews, edits, and approves it. Users can also disable or adjust the feature through /config.
Claude Code has introduced automatic feedback drafting, and reports are sent to Anthropic only after the user has reviewed, edited, and approved them. It can be triggered when a run fails, Claude detects that it made an error, or a user explicitly points out a problem; once a draft is generated, the user can inspect and revise it before deciding whether to approve sending it. According to Anthropic staff member Thariq, the capability is implemented through the new SendFeedback tool, so users no longer need to run /feedback first and then write the report manually. The related automatic drafting and sending behavior can also be disabled or adjusted through /config.
Claude in Chrome is now officially available on all paid subscription plans
产品应用
Claude has made Claude in Chrome generally available on every paid Claude plan, allowing it to carry out browser actions autonomously across tabs. Before each action runs, a safety classifier checks that it is safe and matches the user’s original request.
All paid Claude subscribers can install Claude in Chrome from the Chrome Web Store. The tool is now generally available and adds automatic approval for browser actions. Using the user’s existing logins, it can view pages, read and enter text, click links, navigate between pages, fill out forms, work across tabs, and continue conversations in the desktop, mobile, and web apps. Users can disable automatic approval in settings and return to confirming each action manually. Actions that do not match the original request are blocked.
The safeguards also include probes that scan tool results. If they detect suspected prompt injection, Claude is warned and can check with the user when necessary. According to the official evaluation using stronger attacks supplied by professional red-teamers, before additional safeguards were applied, attacks that reached the model succeeded 17.6% of the time against Claude Opus 4.5 and 3.8% against Claude Opus 5. With probes and the safety classifier enabled, the rate was 0% for Claude Sonnet 5, Claude Opus 5, and Claude Mythos 5, and 0.3% for Claude Fable 5. The official account says manual verification found that every successful break occurred in a low-severity scenario. Enterprise administrators can restrict the tool to approved domains in Organization Settings. Working with local files or other applications still requires the Claude desktop app, and the tool does not support other Chromium browsers.
Claude introduces a built-in browser in Cowork, rolling out within the next week
产品应用
Claude Cowork is rolling out a built-in browser that will reach all paid desktop plans within the next week. It can navigate websites, fill out forms, and complete entire tasks from the side panel without requiring a separate installation.
Claude Cowork is rolling out its built-in browser in stages and will make it available to all paid desktop plans within the next week. When a task involves a website, the browser opens directly in Cowork’s side panel. According to the official announcement, Claude can use it to navigate pages, fill out forms, and execute the entire task. The feature is integrated into the desktop app, requires no additional browser installation, and operates independently from the user’s own browser and login state.
Grok Bot opens to SuperGrok and Cursor Pro subscribers and resets weekly usage limits
产品应用
Grok Bot is now available to all SuperGrok and Cursor Pro subscribers, and weekly usage limits have been reset for all users; Michael Truell said the product is growing faster than any product he has seen.
Grok Bot’s official account announced that the service is now available to all SuperGrok and Cursor Pro subscribers and that weekly usage limits have been reset for all users. Michael Truell said Grok Bot is growing faster than any product he has seen. According to his description, users are delegating tasks across multiple real-world workflows, including running small e-commerce businesses, coordinating customer events, testing production software, and completing mundane daily work.
谷歌 rolls out a major productivity upgrade for Gemini Live
产品应用
Google has begun rolling out a productivity upgrade for Gemini Live, adding Spark multi-step tasks, Daily Brief, voice-based Gmail management and Personal Intelligence; according to the company, 63% of users interact with Gemini by voice.
Google is now rolling out a set of voice-based productivity features for Gemini Live, including Spark integration, Daily Brief, Gmail inbox actions and Personal Intelligence. Users can create complex, multi-step tasks in natural language, with Spark running them in the background across Google Docs, Sheets, Drive and the web and handling ongoing or scheduled work over days or weeks. Daily Brief combines important updates for the day from Gmail and Calendar into a spoken to-do summary. According to the company, 63% of users speak to Gemini aloud.
In Gmail, Gemini Live can search, summarize, star, archive or delete messages through voice commands, and it can answer whether there are new emails or urgent messages from a specified source. Personal Intelligence answers questions by combining past conversations with information from connected Gmail, Photos, Search and YouTube accounts. Within an ongoing conversation, Gemini Live can also hand requests to Spark for background processing, such as extracting event invitations from email, creating family Calendar events and adding driving times. To use the features, users must first choose to connect their apps in the Gemini Personal Intelligence settings, then tap the Live icon in the Gemini app.
OpenAI introduces $visualize for generating visual charts in ChatGPT Work and Codex
产品应用
OpenAI Developers reposted a product demo introducing the $visualize feature in ChatGPT Work and Codex. According to the demo, users can enter the command to convert information into visual content.
OpenAI Developers introduced the $visualize feature in ChatGPT Work and Codex by reposting and quoting a demo from staff member Katia Gil Guzman. The demo says users can enter the $visualize command to convert information into visual content. The presentation covers both ChatGPT Work and Codex, with the same $visualize command serving as the core interaction in each. The original post does not provide a specific launch date or list the types of charts that can be generated.
Manus enables data recovery as services return to normal, stable operation
产品应用
Manus has opened its data recovery feature and says its service is back to normal and operating stably. Users with backed-up data can restore it at any time with no deadline, while affected accounts will also receive a Welcome Back Bonus.
Manus has now enabled its data recovery feature and said through its official account that the service has returned to normal and remains stable. Recovery is available to users who previously backed up their data, and they can initiate the process according to their own schedule without completing it by a specified date. According to the company, there is no recovery deadline, and users can perform the operation whenever convenient. Manus will also provide affected accounts with a return incentive called the Welcome Back Bonus.
