According to CoinJie.com, Google has announced several underlying infrastructure updates for the Gemini API managed proxy to improve the stability of agents running in the cloud. This update addresses three major obstacles developers face when deploying production-grade applications: long-running tasks experiencing HTTP timeouts and disconnections, the need to build intermediary proxies to connect to private data, and resetting sandboxes when updating authorization keys. The four adjustments include: 1. Asynchronous background execution, with clients able to poll the status at any time; 2. Direct connection to remote MCP services, allowing agents to communicate directly with the server; 3. Hybrid function calls, with status flagged as `requires_action`; 4. Lossless refresh of network credentials, allowing developers to directly override configurations. The managed proxy is now available in the free API tier, and developers can use the `/genai` development kit to call it directly.

According to CoinWorld, Meta has launched Muse Image, an image generation model developed by its Super Intelligence Lab. Unlike traditional models that directly map text to pixels, Muse Image exhibits behavior similar to a design agent and works in conjunction with the recently released Muse Spark language model. Before drawing, the model devises a composition plan. If the prompt requires factual information, it searches the web to ensure accuracy. If precise layout is needed, it renders clear and readable text within the artwork and constructs realistic QR codes based on custom prompts. Users can also annotate artwork using the annotation function, instructing the model to understand the context and fine-tune local details without regenerating the entire image. Muse Image is currently available on the Meta AI app; it's free for everyday creation, while more advanced creation features require a paid subscription.

According to CoinWorld, Anthropic's intelligent agent application, Claude Cowork, has ended its desktop exclusivity and is now officially available on web browsers, as well as iOS and Android mobile devices. Previously, Cowork could only run online on a computer client, and tasks would be interrupted if the device was turned off or offline. Now, tasks are fully hosted in the cloud background, allowing users to simply close their computers after setting up scheduled tasks. When the agent encounters decisions requiring human intervention during task execution (such as analyzing financial reports or compiling contracts), it will send a pop-up request to the user's phone, and will not send emails or publish any content without the user's consent before the user "signs off" via mobile phone. Beta testing will first be available to Max subscribers in the coming weeks, and the double credit promotion for Cowork will be extended to August 5th.

According to CryptoWorld, Anthropic has announced a 5-day extension to July 12th for paid subscribers to enjoy the free trial of its top-tier inference model, Claude Fable 5. During the extended period, the user's spending limit remains at 50% of their weekly total. The model was briefly removed from subscriptions due to export controls and was originally scheduled to fully transition to a pay-as-you-go model on July 7th when it resumed global access on July 1st. After July 12th, Claude Fable 5 will be officially removed from subscription limits; paid subscribers will only be able to purchase usage credits through the API at a price of $10 per million inputs and $50 per million outputs.

Zcash's native token, ZEC, surged over 12% on Tuesday after the team developing its privacy pools announced it was close to completing a mathematical proof confirming the absence of an undetectable forgery vulnerability in the latest Zcash shielded pool. This verification, spearheaded by Project Tachyon, is aimed at Zcash's upcoming Ironwood shielded pool. Zcash founder Zooko Wilcox stated that the project is on the verge of producing a mathematical proof that demonstrates the absence of an undetectable forgery vulnerability in the latest Zcash privacy pool. This progress stems from the disclosure last month of a serious forgery vulnerability in the Zcash Orchard shielded pool. At the time, this vulnerability caused market concerns about undetectable hidden forgery risks in Zcash's privacy system, causing ZEC to plummet by over 40% within two days. Developers stated that with AI-assisted formal verification, such proofs, which previously might have taken years, can now be compressed into weeks. The news propelled ZEC back above $500, its highest level since early June. (The Block)