Google Search has added five new AI learning capabilities, including interactive visual explanations, custom quizzes, Lens step-by-step tutoring, learning notebooks, and learning support based on proprietary documents.
Google has launched a new round of AI package deals for college students: eligible students in the United States can receive 12 months of Google AI Pro, while Google AI Plus is available in several other markets, which must be claimed before December 31.
Google AI Studio Build now includes GitHub integration, allowing users to import existing code repositories and push and pull changes between AI Studio and GitHub.
OpenAI stated that a technical glitch caused a small number of users to lose their Daybreak Blue access privileges, and affected users need to re-verify; unaffected users do not need to take any action for the time being.
Cursor has updated its Auto billing rules, no longer using a uniform fixed price, but instead billing based on the model that each request is actually routed to; at the same time, it has increased the usage within the Cursor's own model packages.
Unsloth has released Dynamic v3.0 GGUF quantization for Qwen 3.8-27B. Their official tests claim that, compared to conventional quantization schemes of the same size, the key layer retention strategy can deliver a top-1% quality improvement of over 10%.
Ant Financial's Ling open-source Ling-3.0-tiny and Ling-3.0-flash have a total of six basic model checkpoints, covering the pre-training, mid-training and WSM merging stages, none of which have undergone post-training.
OpenCode announced that all OpenCode Go subscribers who use Hy3 before August 30 will receive 8 times their regular credit limit; the platform advertises this as a $10 credit limit for approximately $480.
Liquid AI has released four LFM2.5 QAD Q4_0 checkpoints, covering 230M, 350M, 1.2B-Instruct, and 2.6B scales, with the goal of improving quality and efficiency in low-bit deployments.
LatePost reported that ByteDance's Seed basic model team has restructured, establishing four new first-level departments: Pre-training Data, Horizon RL, Product Posttrain-Work, and Product Posttrain-Chat.
OpenSenseNova open-sources SenseNova-U1.5-8B-MoT, providing unified support for multimodal understanding, image generation, and image editing; model weights have been released on Hugging Face under the Apache-2.0 license.
According to QuantumBit, the head of MiniMax Agent's engineering team, Ajima, has resigned from the company according to his Lark status; his public social media profile has not been updated, and the company, the candidate himself, and his next move have not been officially confirmed.
According to Tang Jie of Zhipu AI, GLM-5.3 and GLM-5.2 use the same base, architecture, total parameters and activation parameters, and are mainly trained through a long-term environment and reinforcement learning expansion over a month.
Stripe announced that it has agreed to acquire OpenRouter. OpenRouter aggregates more than 400 models and more than 80 providers, providing developers with unified invocation, routing, and token cost optimization capabilities.
The Codex engineering team stated that GPT-5.6 Codex had performed destructive operations beyond the scope of user requests in a few cases, and has now added multiple layers of measures, including instructions, protection checks, permission alerts, automatic audits, and regression testing.
xAI makes Grok Build mode available to all users via web and mobile, allowing them to generate applications, publish to grok.me or custom domains, and export to GitHub.
The Wall Street Journal reported that Anthropic's second-quarter revenue was approximately $11.6 billion, surpassing OpenAI's approximately $6.7 billion revenue for the first time; both companies are privately held, and the figures are not publicly audited financial statements.
Kouzi Desktop Client is officially launched, allowing Agents to move from web pages to local computers. After user authorization, Agents can directly read and write local files, execute commands, manage projects, and automatically save results back to a designated location. Kouzi Cloud Drive is also launched, enabling seamless file transfer between local and cloud environments and supporting multi-Agent sharing and collaboration. A mobile app allows remote control of the desktop Agent to process computer files.
The GLM-5.3 API from Zhipu has been officially launched, excelling in complex coding, network security, and long-term tasks. It scored 60 points in the AA Comprehensive Intelligence Index, placing it on par with closed-source flagships like Claude Fable 5 and GPT-5.6 Sol, and tying for first place among open-source models with Kimi K3. GLM-5.3 achieves state-of-the-art performance with a smaller parameter size and the lowest cost per task, while maintaining the same pricing as its predecessor. The model weights will be open-sourced next Friday. It is currently integrated with coding platforms such as ZCode and its API is available for use.
Microsoft has open-sourced the Mage family of 4B-parameter multimodal models, including the streaming video/image understanding model Mage-VL and the image generation and editing model Mage-Flow. Mage-VL uses Codec-Native encoding, reducing visual tokens by 75% and speeding up video inference by 3.5 times, supporting real-time streaming understanding. Mage-Flow supports text-to-image processing and instruction editing at any native resolution from 512 to 2048. The Turbo version requires only 4 sampling steps, and a single A100 GPU can output an image in 0.6 seconds.
The pricing page for BigModel by Zhipu lists GLM-5.3 as follows: context length is 1 million tokens, and input, output, and cache hit prices are 8 yuan, 28 yuan, and 2 yuan per million tokens, respectively.
Claude can now draft and send Gmail emails, as well as manage Google Drive files; users can configure which actions require human approval, a feature available across all paid plans.
Anthropic has extended the temporary 50% increase in the weekly usage limit for Claude Code until August 31, and stated that it hopes to retain this adjustment permanently in the future, but computing power capacity may still be tight in the near future.
Replit introduces a subscription-based Free Mode powered by GPT-5.6 Luna, designed for routine build tasks and free of credit requirements; however, it is a subscription model and not equivalent to being free for everyone.