OpenRouter and OpenCode launch free stealth model Ox Alpha
要闻
OpenRouter and OpenCode have launched Ox Alpha, a stealth model whose provider has not been disclosed. Ox Alpha supports a 1M context window and text, image, and video inputs. OpenCode says it will be free for the next week.
OpenRouter and OpenCode have launched the free stealth model Ox Alpha, with OpenCode stating that the model will be free for one week. According to its OpenRouter page, Ox Alpha is a reasoning model designed for coding, sustained agentic work, and production workloads. It features a 1m-token context window, accepts text, image, and video inputs, and returns text. Notably, the OpenRouter page states that prompts and completions are retained by the provider but not used for training, while OpenCode says the model follows a zero-data-retention policy.
Anthropic launches AI learning platform Claude Academy
要闻
Anthropic has launched Claude Academy, an AI learning platform offering free courses and tutorials to everyone. In addition to guidance on using Claude, it provides general AI learning content not tied to any particular product or model.
Anthropic has announced the official launch of its AI learning platform, Claude Academy, which offers three types of learning content: courses, tutorials, and use cases. In addition to teaching people how to use Claude, the platform provides general AI learning materials that are not tied to any specific product or model. The content is organized around real-world problems users need to solve at work and in their personal lives, and the platform can recommend courses based on users’ interests while tracking completion progress and badges. All courses and tutorials are available to everyone for free.
Ollama rolls out Kimi K3 to more than half of its cloud subscribers
开发生态
Ollama says Kimi K3 has been rolled out to more than half of its cloud subscribers. Usage is included in their plan allowances, and availability will continue to expand.
Ollama is rolling out the Kimi K3 model in phases to its cloud subscribers. It is currently available to more than half of all subscribers, with usage included in their plans, and the company says it will continue expanding access. The service is hosted in the United States and Europe and follows a zero-data-retention policy. Ollama also said it is improving the pricing transparency of its cloud service to highlight the best price-to-performance ratio.
OpenCode has announced that Muse Spark 1.2 is now available for free on its platform. Users can currently access the model directly through OpenCode at no cost.
OpenCode has announced that Muse Spark 1.2 is now available for free on the OpenCode platform. Users can currently access Muse Spark 1.2 directly through OpenCode at no additional cost. The free offering is now officially in effect and applies across the OpenCode platform.
Google has officially launched Antigravity IDE extensions for VS Code, Visual Studio, JetBrains, and Zed. Individual developers can sign in with any Antigravity subscription plan, including the free tier.
Google has announced through its official blog the launch of the Antigravity IDE extension, which supports Visual Studio Code, Visual Studio, the full range of JetBrains IntelliJ-based IDEs, and Zed, and is now available to install. Once installed in an IDE, it gives developers access to sessions, customization, and multi-agent orchestration capabilities that share context with the Antigravity ecosystem. Individual developers can sign in with a Google account and use it with any Antigravity subscription plan, including the free tier, while enterprise teams can sign in using Gemini Enterprise or Google Cloud project credentials.
ChatGPT desktop now supports read-only conversation sharing
开发生态
OpenAI has announced that ChatGPT Work and Codex now support sharing conversations through read-only links. The feature is currently limited to the desktop app.
The official ChatGPT account announced that users can now share read-only conversations from ChatGPT Work and Codex to showcase their work process and provide background context. The OpenAI Developers account later explained that shared threads in Codex and ChatGPT Work can present the build process through read-only links, providing context and reasoning for pull requests, in-depth analyses, or project handoffs. The feature is currently available on desktop.
ChatGPT desktop has added an Apple Messages plugin that can search messages, organize conversations, and send replies on Mac.
ChatGPT officially announced the launch of an Apple Messages plugin, allowing users on Mac to use ChatGPT to search for information, catch up on conversations, and draft and send replies. The feature is now available in ChatGPT Work and Codex on desktop.
OpenAI adds collaborative editing with teammates to ChatGPT Sites
开发生态
OpenAI's developer account announced that ChatGPT Sites now lets users add teammates as editors. Multiple people can build and publish a site together, with collaborators pushing changes to the same project.
OpenAI Developers officially announced that ChatGPT Sites now supports adding teammates as editors, enabling multiple people to build and publish together. Collaborators can push changes to the same project, while git management and CI are handled automatically by Codex in the background.
OpenAI expands Computer History to the EEA, UK, and Switzerland
开发生态
OpenAI has expanded Computer History to the European Economic Area, the UK, and Switzerland. It is currently available to ChatGPT Pro, Business, and Enterprise users on Mac.
OpenAI announced through the OpenAI Developers account that Computer History is now available to ChatGPT Pro, Business, and Enterprise users in the European Economic Area, the UK, and Switzerland, currently on Mac only. The feature can remember users’ activity across apps and websites, making subsequent conversations more personalized and reducing the need for repeated explanations. Users must opt in to enable it.
OpenAI adds transparent backgrounds to the GPT-Image-2 API
开发生态
OpenAI has added transparent background support to the GPT-Image-2 API. The company says it can generate reusable assets that work against any background, with more transparency capabilities coming soon.
The official OpenAI Developers account announced that transparent backgrounds are now available in preview for GPT-Image-2 through the API. Developers can generate reusable assets that can be placed directly on any background, covering use cases such as product imagery, graphic design, website mockups, and marketing campaigns.
Exa has launched Exa plugins for Codex and ChatGPT. The company says the plugin gives Codex access to more than 100 billion websites, documents, and other resources.
Exa announced on its official account the launch of an Exa plugin for use in Codex and ChatGPT. The OpenAI Developers account reposted the announcement, stating that both ChatGPT Work and Codex can access more than 100 billion websites, papers, documents, and other resources through the new Exa plugin. Neither post mentioned pricing or usage limits. Multiple users asked in the comments whether it is free, but no response has been seen so far.
Anthropic's developer account ClaudeDevs announced that Claude Code now supports setting its output style to Concise. When enabled, Claude prioritizes results and shortens its responses. Users can turn it on through /config or settings.json.
Anthropic’s developer account, ClaudeDevs, announced that Claude Code has added a Concise output style option. When enabled, Claude prioritizes delivering results and keeps responses brief, while still providing full details when requested by the user. Users can enable the feature in either of two ways: select Output style in the /config menu, or set the "outputStyle" field to "Concise" in the settings.json file.
Claude Platform makes computer use and other features generally available
开发生态
Anthropic has announced the general availability of computer use, browser use tools, the Skills API, and the Files API on Claude Platform.
Anthropic announced the general availability of computer use, the Skills API, and the Files API on the Claude platform, along with a new browser use tool in computer use. The Skills API and Files API are also now available on Microsoft Foundry, while the updated computer use and browser use tools are coming soon to Google Cloud’s Vertex AI. Existing beta integrations will remain available during the migration.
TraeWork has launched a computer control feature that lets users instruct AI in natural language to operate their desktop directly and perform tasks across applications. Background operation is currently supported on Mac but not yet on Windows.
TraeWork has officially launched computer control capabilities, allowing users to simply describe a task in natural language and have AI continuously operate across applications to deliver files or reports. The feature extends AI workflows to local files and desktop software without APIs, covering virtually anything that can be operated with a mouse and keyboard. To ensure safety, the entire execution process remains visible, and high-risk actions such as payments and deletions are paused for confirmation. Screen content is not stored persistently. Users need to upgrade to the latest version and enable the feature in Settings or the plugin marketplace. Mac supports background operation, while Windows does not currently support it and may require control of the mouse.
SenseTime open-sources multimodal model SenseNova-U1.5-8B-MoT
模型发布
SenseTime has open-sourced SenseNova-U1.5-8B-MoT, a natively unified multimodal model. The company says the release delivers improvements in six areas, including image generation, text rendering, native 4K generation, and image editing.
SenseTime released the native unified multimodal model SenseNova-U1.5-8B-MoT on Hugging Face. Built on the NEO-unify architecture, it is open-sourced under the Apache 2.0 license. According to the company, this version delivers user-perceptible improvements in six areas: image generation quality, Chinese and English text rendering and infographics, native 4K generation, image editing, complex instruction following, and visual control. The company is preparing to open-source a technical report and the complete training pipeline spanning SFT, RL, and MOPD.
Liquid AI has released DSpark draft model checkpoints for three models in the LFM2.5 series. They use speculative decoding to accelerate inference without changing output quality.
Liquid AI has released DSpark draft model checkpoints for three models—LFM2.5-1.2B-Instruct, LFM2.5-2.6B, and LFM2.5-8B-A1B—delivering substantial decoding acceleration with only a minimal increase in memory usage and no change in output quality. According to official data, throughput improves by up to 3.18× on GPUs and 2.87× on-device, while function-calling latency for LFM2.5-2.6B is reduced by an average of 57%.
Alibaba Qwen releases GUI agent foundation model Qwen-UI-Agent
模型发布
Alibaba Qwen recently released Qwen-UI-Agent, a foundation model for GUI agents. According to official data, Qwen-UI-Agent outperformed models including GPT-5.6 Sol across multiple mobile and desktop benchmarks.
Alibaba's Qwen team recently released Qwen-UI-Agent, a foundation model for GUI agents spanning mobile, desktop, web, and deep-research environments. According to official data, Qwen-UI-Agent outperforms flagship models including GPT-5.6 Sol and Claude Opus 4.8 on multiple benchmarks such as MobileWorld, OSWorld-Verified, and WebArena.
Grok Build mode opens to all SuperGrok and X Premium users
产品应用
Grok has announced that Build mode is now available to all SuperGrok and X Premium subscribers across grok.com, iOS, and Android.
Grok's official account announced that Grok Build mode is now available on grok.com, iOS, and Android to all SuperGrok and X Premium subscription plans, allowing users to turn an idea into a published product with its own domain using a single prompt. According to Grok, Build mode runs a full coding Agent with support for subagents, browser use, databases, and secrets. Apps can be published to a grok.me domain or the user's own domain, exported as source code for further development, or shared only with a team.
Black Forest Labs launches FLUX Upscale for upscaling video to 4K
产品应用
Black Forest Labs has launched FLUX Upscale, a video upscaling tool that can regenerate any video at up to native 4K. It is now available through the BFL API and playground.
Black Forest Labs has officially launched FLUX Upscale, a video upscaling tool offered as a standalone FLUX Tool and endpoint, now available through the BFL API and playground. The tool accepts any video at 480p or above and regenerates it at resolutions up to native 4K. With native understanding of FLUX 3 outputs, it can restore blurry faces that general-purpose upscalers often miss and remove grid-like artifacts from textures such as water and grass.
MiniMax's former head of R&D, 阿岛, has reportedly left the company. He previously oversaw core business lines including the M-series models, Agent, and 海螺AI.
According to reports, Adao (Miao Yuhang), head of MiniMax Agent's engineering department, has left the company, with his Feishu status now showing that he has departed. He previously served as head of engineering R&D, overseeing multiple development tracks including MiniMax M3.x, Code, Agent, Audio, and Hailuo AI. His profile information on personal social platforms has not yet been updated, and there is currently no public information about his next destination or who will take over his former responsibilities.
Meta details internal evaluation framework WildArtifactBench
技术与洞察
Meta has detailed its internal evaluation framework WildArtifactBench and released 10 of its tasks. The framework evaluates multimodal agents on complex real-world tasks using win rates and Elo scores based on human and agentic preference judgments.
Meta introduced an internal evaluation framework called WildArtifactBench to assess agent performance on complex, real-world tasks and across multiple deliverable formats. Rather than relying on rigid ground-truth scoring rubrics, the framework uses win rates and Elo scores derived from human preference judgments and Agentic preference judgments, expanding task coverage in practical multimodal workflows. The framework as a whole remains internal, with only 10 of its tasks released publicly so far.
Anthropic could publicly file for an IPO as early as late August
前瞻与传闻
According to Bloomberg, citing people familiar with the matter, Anthropic's IPO is expected to match or exceed SpaceX in size, and the company could publicly file for a listing as early as late August.
According to Bloomberg, people familiar with the matter said Anthropic expects the size of its IPO to match or even surpass SpaceX's record-setting IPO. The sources said Anthropic is modeling the relevant figures in preparation for a potential mega-IPO and could publicly file as early as the end of August. They also said a recent investor briefing led by Chief Financial Officer Krishna Rao avoided the question of valuation. None of these claims have been officially confirmed by Anthropic.
Anthropic plans changes to its advanced AI data retention policy
前瞻与传闻
Anthropic staff say the new enterprise Fable safeguards will run on customers' own infrastructure, giving them control over their data while Anthropic retains none of it. The system is planned for release this fall, while the current 30-day retention requirement remains unchanged.
Anthropic staff said the company is rolling out new Fable safeguards that can run on enterprise customers' own infrastructure, giving customers control over where their data is stored and who can access it. Thariq said the solution has been co-developed with roughly 100 companies for some time and is expected to see a broader rollout in the fall; Boris Cherny said Mythos-class models require additional safeguards, customers can own and control their own data, and Anthropic retains no data. Anthropic previously required prompts and outputs from Covered Models such as Fable 5 and Mythos 5 to be retained for 30 days for safety monitoring, even for customers with ZDR agreements. The current official documentation still specifies this requirement, and the policy has not yet been formally changed.
Note: This content was created with AI assistance and may contain hallucinations and errors.
