Google has released the Gemini 3.7 Flash model. According to the company, it delivers significant improvements in software engineering, web development, and complex knowledge workflows, outperforming its predecessor across multiple benchmarks. The model is now broadly available via the API, AI Studio, and other platforms, with a 50% discount offered through the end of the year. Pricing for 3.6 Flash has also been reduced.
Google announced the release of the Gemini 3.7 Flash model, just around 3 weeks after the previous 3.6 Flash version. According to the company, the model delivers significant improvements in software engineering, Web development, and complex knowledge workflows, outperforming its predecessor across multiple benchmarks. Gemini 3.7 Flash is now available to developers, enterprises, and select individual users through the Gemini API, Google AI Studio, Antigravity, Android Studio, Gemini Enterprise, and other platforms. Through the end of this year, the model is priced at US$0.75 per million input Tokens and US$3.75 per million output Tokens, while pricing for 3.6 Flash has also been reduced to the same levels.
DeepSeek releases developer preview of open-source agent framework DeepSeek Harness
要闻
DeepSeek has released a developer preview of DeepSeek Harness, an open-source agent framework. Built around an everything-is-a-plugin architecture, it allows components such as models, tools, and UIs to be freely swapped out. The framework supports four modes—Standard, PTC, Minimal, and Creative—and is now available to install and use.
DeepSeek has launched a developer preview of its open-source agent framework, DeepSeek Harness, and released the source code under the MIT license. The framework uses an “everything is a plugin” architecture, in which all components—including models, tools, skills, sessions, sandboxes, storage, loops, scheduling, and UI—exist as plugins. Developers can independently replace or extend them without modifying the source code. It also supports four distinct runtime modes: Standard, PTC, Minimalist, and Creative.
DeepSeek V4 Pro officially launched; API price increase takes effect August 17
要闻
DeepSeek has officially released DeepSeek V4 Pro, confirming a simultaneous rollout across all platforms, with model weights also published on relevant platforms. Meanwhile, new pricing for the DeepSeek API will take effect at midnight on August 17, along with a peak and off-peak pricing mechanism.
DeepSeek has officially released the production version of DeepSeek V4 Pro (DeepSeek-V4-Pro-0813), with updates now live simultaneously on the APP, Web, and API. Users can access the new model through “Expert Mode,” while the API model name remains deepseek-v4-pro. According to the company, its Agent capabilities have been substantially enhanced. The API now natively supports the OpenAI Responses API format and is compatible with Codex, while reasoning effort can be set to low/high/max. API prices will increase starting at 0:00 Beijing time on August 17, 2026, with peak and off-peak pricing also being introduced; prices will double during peak hours.
OpenAI and Cerebras launch Ultrafast mode for GPT-5.6 Sol
要闻
OpenAI has introduced Ultrafast, a new API service tier powered by Cerebras compute. It enables GPT-5.6 Sol to generate up to 750 output tokens per second, as much as 14 times faster than the standard mode. The service is currently in limited preview and available only to select customers.
OpenAI has launched Ultrafast mode, a new service tier in the OpenAI API designed specifically to run the GPT-5.6 Sol model. The mode is powered by Cerebras’ Wafer-Scale Engine architecture. According to official OpenAI data, Ultrafast mode can generate up to 750 output tokens per second, making it as much as 14 times faster than standard mode. Ultrafast mode is currently in limited preview and available only to select customers. OpenAI will evaluate customers for inclusion based on workload fit and available capacity, and will gradually expand access as capacity grows.
Computer History launches in the ChatGPT desktop app
要闻
OpenAI has launched Computer History in the ChatGPT macOS desktop app. The feature records users' interactions across apps and websites and turns them into memories. It is now available to Pro and Enterprise users.
OpenAI has officially launched the Computer History feature for the ChatGPT desktop app. By recording users’ interactions within authorized apps and websites, the feature converts them into memories and a timeline, helping ChatGPT and Codex understand users’ recent work and provide relevant context. Unlike the earlier Chronicle preview, which relied on screenshots, the new system does not capture screenshots, screen recordings, or audio. The feature is currently off by default and has been rolled out globally to Pro, Business, and Enterprise users, but users must opt in and enable the Memories feature.
MiniMax releases and open-sources Music 3.0 music generation model
模型发布
MiniMax has officially released and open-sourced the MiniMax Music 3.0 music generation model. Using creative descriptions and lyrics, the model can generate a complete song of up to five minutes, including vocals and arrangements, in a single run.
MiniMax has released and open-sourced its next-generation music generation model, MiniMax Music 3.0, with the model weights now publicly available on Hugging Face. The model uses a hybrid architecture consisting of an 8B Global LLM and a 0.6B Local LLM, outputs 32 kHz stereo audio, and is designed to improve structural stability and vocal naturalness in long-form music. It currently supports local deployment through multiple frameworks, including diffusers.
小红书 dots studio releases open-weight dots3-note preview model
模型发布
小红书 dots studio has released dots3-note preview, a model with 280B total parameters and 16B active parameters. It supports a 512K context window and can understand images, video, and audio. The model weights have been published on relevant platforms under the Apache 2.0 license.
Xiaohongshu's dots studio has released dots3-note preview, the first open-weight model in the dots3 series. It uses an MoE architecture with 280B total parameters and 16B activated parameters, supports contexts of up to 512K tokens, and can understand text, images, video, and audio while producing text output. The model has been open-sourced under the Apache 2.0 license on Hugging Face, ModelScope, and GitHub, with both BF16 and FP8 versions available. The company also says it has open-sourced two accompanying evaluation environments for real-life scenarios: VibeSearchBench and VibeLifeBench.
OpenRouter offers an exclusive additional 50% discount on Gemini 3.7 Flash through August 27
开发生态
OpenRouter has announced an exclusive additional 50% discount on Gemini 3.7 Flash on its platform through August 27.
OpenRouter has officially announced that the Gemini 3.7 Flash model is available on its platform at an exclusive additional 50% discount through August 27. During the promotional period, pricing is approximately $0.375 per million input tokens and $1.88 per million output tokens.
Claude Code desktop has added an auto-continue checkbox. When enabled, the system automatically resumes from where it left off once the usage limit resets.
ClaudeDevs announced that Claude Code desktop now includes a new auto-continue checkbox. This feature is primarily intended for situations where users reach their usage limits. Once enabled, users no longer need to restart tasks manually. As soon as the system's usage limit resets, Claude Code desktop will automatically resume from where it left off.
快手 streamlake to discontinue KwaiKAT models and subscription plans
开发生态
快手 streamlake has announced that the KwaiKAT model series and Coding Plan will be discontinued on August 31, 2026, as part of changes related to a business upgrade.
Kuaishou streamlake has issued its latest notice announcing that, due to business upgrades and adjustments, it has decided to officially discontinue the KwaiKAT model series and its associated subscription plan. The specific offerings affected are the KAT-Coder-Pro V2.5 and KAT-Coder-Air V2.5 models, as well as the KwaiKAT Coding Plan. These models and the subscription plan will officially cease service at 23:59:59 on August 31, 2026, after which the related capabilities will no longer be available to users.
NousResearch launches public beta of Hermes Agent Bot Mode plugin
开发生态
NousResearch has launched a public beta of Bot Mode, a desktop plugin for Hermes Agent. The plugin turns configuration files into named bots, each with its own chat, avatar, persona, and routine tasks, while also supporting communication between bots.
NousResearch has announced the release of the Bot Mode desktop plugin for Hermes Agent and launched its public beta. As a new alternative mode, Bot Mode provides a list of bots in the interface sidebar, converting existing agent profiles into separate, named bots. Each bot has its own chat history, avatar, SOUL.md persona file, skills, and memory. Users can switch between bots directly by clicking them in the list, eliminating the need to manually switch configuration files. The plugin currently requires no patches to the core code, but users must update to the latest version of the Hermes desktop app.
Artificial Analysis has launched Optima, a model benchmarking platform. Users can build custom tests with their own data and compare multiple models in terms of quality, cost, and completion time.
Artificial Analysis has officially launched Optima, a custom benchmarking platform designed to help users identify the model best suited to their specific tasks. The platform allows users to build custom benchmarks by uploading files, importing Agent trajectories, installing programming environment plugins, or directly describing their use cases. Users can run multiple frontier models for comparison with a single click and evaluate their performance in terms of quality, cost per task, and time required using methods such as standard scoring or pairwise scoring. Optima uses token-based pricing and is also offering a $10 trial credit to new users who register for a limited time.
ChatGPT on the web can now open and process Google Drive documents directly
产品应用
ChatGPT on the web can now directly open documents, spreadsheets, and presentations from Google Drive. The feature is being gradually rolled out to Plus, Pro, Business, and Enterprise subscribers of ChatGPT and ChatGPT Work.
ChatGPT now allows users to open documents, spreadsheets, or presentations from Google Drive directly on the web, enabling side-by-side work without switching browser tabs. The new feature is currently rolling out gradually on the web and is available to Plus, Pro, Business, and Enterprise users of ChatGPT and ChatGPT Work.
Anthropic improves proactive responses from Claude Tag in Slack
产品应用
Anthropic has updated Claude Tag in Slack, enabling it to use the full context of a channel to determine when to proactively join a conversation. The feature is now available to Claude Teams and Enterprise customers.
Anthropic has announced an update to its Claude Tag feature integrated into Slack, removing the previous classifier that could only assess messages individually and instead using the full channel context and persistent user-provided instructions to determine what action to take. The updated Claude Tag delivers faster initial responses and can enter sleep mode when there is repeatedly nothing substantive to add, remaining dormant until a user @mentions it. The update is now live in Claude Tag and is available exclusively to Claude Teams and Enterprise customers. Anthropic explicitly stated that there is no additional charge for the update today and that the extra context usage will not count toward the usage or spending limits of any subscription plan.
Google has launched Sheets canvas, a new Google Sheets feature that lets users turn spreadsheet data into interactive apps using prompts. The feature is now available to Google AI Pro and Ultra subscribers, among others.
Google has launched Sheets canvas, a new feature in Google Sheets. Built on Gemini, the feature allows users to transform static spreadsheet data into visual interfaces such as interactive dashboards using simple natural-language prompts, without writing formulas or code. Sheets canvas functions as a read-write layer that stays synchronized with the source data in real time and supports collaboration and sharing just like a regular spreadsheet. The English version is now available globally to Google AI Pro and Ultra subscribers, and is also beginning to roll out to Google Workspace customers on select plans and subscribers to relevant education add-ons.
Microsoft to merge consumer and enterprise Copilot apps
产品应用
Microsoft is merging its consumer and enterprise Copilot apps and will discontinue several features, including group chats, AI podcasts, and Deep Research. Files stored in the standalone app will be migrated to OneDrive.
Microsoft is merging the consumer-facing Copilot app with the enterprise-focused Microsoft 365 Copilot app and will gradually phase out a number of underperforming AI features. According to Microsoft support documentation, regular users will lose access to group chats, AI-generated podcasts, Copilot Labs experimental features, and Deep Research by August 18, 2026. Files created in the standalone Copilot app will be migrated to OneDrive.
菜鸟智能体 launches on 千问开放平台, enabling parcel delivery through conversation
产品应用
千问 has announced the launch of 菜鸟智能体 on 千问开放平台. Users can describe their shipping needs through conversation, and the agent will recommend a suitable shipping plan and courier company.
The Qwen Open Platform has officially launched the Cainiao agent, enabling users to send packages through conversational interactions. Based on requirements such as item type, price, and pickup time, the agent can recommend suitable shipping options and courier companies. It also supports inquiries about shipping policies, recommendations for shipping bulky items, and automatic completion of previously used addresses. Users currently need to update Qwen to the latest version, after which they can invoke the agent by typing @Cainiao or simply stating their needs.
超级小爱2.0 launches 专家模式 and points-based membership plans
产品应用
超级小爱2.0 has introduced 专家模式 and points-based subscription plans in the 小米澎湃OS 4 Beta. Powered by the Xiaomi MiMo model, the mode supports complex task orchestration and document generation and includes 1,000 points per month. Users can purchase paid plans for additional points and exclusive benefits.
Super XiaoAI 2.0 has officially launched “Expert Mode,” along with corresponding membership subscription plans based on a unified points system. Powered by the Xiaomi MiMo model, the mode supports complex task orchestration, cross-device coordination, and document generation, and provides all users with 1,000 free points per month. Heavy AI users can choose auto-renewing monthly paid subscription plans ranging from 19 yuan to 99 yuan as needed to receive more points and exclusive benefits. Super XiaoAI 2.0 “Expert Mode” is currently open for test users to register and try through the Xiaomi HyperOS 4 Beta.
IBM announces strategic partnership with OpenAI to scale enterprise AI deployments
行业动态
IBM has announced a strategic partnership with OpenAI to help industries deploy AI at scale across core business processes. The partnership will focus on transforming traditional operations, modernizing applications, and strengthening cybersecurity. IBM will establish a dedicated OpenAI business unit and join OpenAI's elite partner tier.
IBM recently announced a strategic partnership with OpenAI to embed OpenAI’s frontier models and products, including GPT-5.6, Codex, and ChatGPT Work, into IBM Consulting Advantage. The partnership will help enterprises scale AI deployment across industries such as finance, government, telecommunications, and retail, as well as functions including finance, procurement, customer operations, and human resources. The collaboration focuses on three key areas: transforming legacy operational processes into AI-ready workflows, application modernization and product development, and cybersecurity and AI risk management. IBM will establish a dedicated OpenAI business unit, with thousands of consultants and engineers earning expert-level certifications through assessments conducted via the OpenAI partner network. As part of the partnership, IBM will join OpenAI’s elite partner tier, and the two companies will also jointly develop solutions for priority industries.
Databricks has reportedly raised $5 billion at a valuation of $190 billion. The funding will be used to invest in enterprise AI capabilities.
According to media reports, Databricks announced that it has completed a US$5 billion funding round, bringing the company’s valuation to US$190 billion. The round was led by Coatue, Blackstone, and other investors. Databricks said its revenue run rate exceeded US$7 billion in the second quarter, up more than 80% year over year, and that the proceeds will be used exclusively to invest in enterprise AI capabilities. CEO Ali Ghodsi revealed that although Databricks intends to go public, it has put those plans on hold for now due to market volatility and its focus on investing in AI products.
Note: This content was created with AI assistance and may contain hallucinations and errors.
