DeepSeek-V4-Pro production API launches with enhanced Agent capabilities
要闻
DeepSeek has launched the production API for DeepSeek-V4-Pro, updating the model to DeepSeek-V4-Pro-0813. The new version enhances Agent capabilities and supports the Responses API and Codex integration. The invocation method remains unchanged, and pricing has not yet been officially adjusted. According to official data, DeepSeek-V4-Pro-0813 performed strongly across multiple benchmarks, with its DeepSWE score rising to 62.7. No further details had been disclosed as of publication.
DeepSeek’s official API documentation and notices show that the production API for DeepSeek-V4-Pro is now live, with the model version updated to DeepSeek-V4-Pro-0813 and the invocation method remaining unchanged. According to DeepSeek, the new version offers enhanced Agent capabilities and improved scores across multiple evaluations in its official benchmark tests. It delivered leading performance on Cybergym and AutomationBench, while its DeepSWE score rose from 12.8 in the preview version to 62.7. DeepSeek said the production release supports integrations including the Responses API and Codex. The company had previously indicated that API prices could rise significantly, but no official price adjustment has yet been made. As of publication, DeepSeek had not disclosed any further information.
SpaceXAI launches Grok 4.6, reaching frontier-level performance on multiple benchmarks with unchanged pricing
要闻
SpaceXAI has released its next-generation Grok 4.6 model, which it says has reached frontier-level intelligence on multiple Agentic coding and knowledge-work benchmarks. Pricing remains the same as Grok 4.5 at $2 per million input tokens and $6 per million output tokens. It is available starting today through Cursor, Grok Build, the API, and other channels, with double usage offered in Grok Build and Cursor during the first week.
SpaceXAI has officially released its next-generation model, Grok 4.6. According to the company, the model has achieved frontier-level intelligence across multiple benchmarks for Agentic coding and knowledge work, scoring 61 on the Artificial Analysis Intelligence Index and tying GPT-5.6 Sol. The model is specifically optimized for long-running Agents and more complex interactive and visual tasks. Pricing remains the same as Grok 4.5, at 2美元 per million input tokens and 6美元 per million output tokens, while the fast version costs twice as much. Starting today, the model is available in Cursor, Grok Build, Grok Bot, and the API, as well as through partners including OpenRouter, Vercel, and Cloudflare. Grok Build and Cursor will offer 2倍 included usage during the first week.
Codex surpasses 15 million active users as Tibo announces quota reset
要闻
Codex lead Tibo previously promised to reset usage quotas for every additional one million active users, with the plan originally set to continue until the service reached 10 million users. Codex has now far surpassed that milestone, exceeding 15 million active users. Tibo announced another Codex quota reset for users, saying it would take effect in about an hour.
OpenAI Codex has surpassed 1500万 active users. Tibo had previously promised to reset usage allowances for the community for every additional 100万 users until Codex reached 1000万 active users; the initiative was temporarily suspended after that milestone was surpassed. On 8月13日, Tibo once again announced an allowance reset for users, saying it would take effect in around one hour. He had also teased a “small surprise,” which ultimately turned out to be this allowance reset.
豆包 launches student verification discount for university students
要闻
豆包 has officially launched a student promotion. University students who complete identity verification can receive 2.5 times the free usage quota and subscribe to the Professional Standard plan for a discounted recurring price of 38 yuan per month, down from the regular 68 yuan.
Doubao has announced a special student offer for currently enrolled university students. After verifying their student status, university students can unlock two benefits in the Doubao App: a student-exclusive free allowance and a discount on the standard Pro plan. The student-exclusive allowance provides 2.5倍 the standard free allowance, while the standard Pro plan, normally priced at 68元 per month, is available to students for 38元 on a recurring monthly subscription. To claim the offer, users can update the Doubao desktop app to the latest version, open Work Task Mode, and send Doubao a message such as “我要领取学生优惠.” Alternatively, they can log in to the Doubao desktop app, access the Doubao Pro subscription page through Settings, select the student benefits option in the upper-right corner, and verify their student status through a Douyin account.
Kimi changes legacy plan renewal rules, barring renewals more than seven days after expiration
要闻
Kimi has announced changes to its legacy plan renewal rules. Starting August 20, 2026, accounts that fail to renew within seven days after their subscriptions expire will no longer be able to purchase legacy plans. Existing users whose subscriptions had already expired when the notice was issued may purchase a legacy plan within seven days of the announcement.
Kimi has issued an official notice announcing changes to renewal rules for legacy plans due to current constraints on computing resources. Starting from 北京时间2026年8月20日, accounts whose subscriptions have expired and have not been renewed within 7天 will no longer be able to purchase legacy plans again. Users with active subscriptions and properly enabled automatic renewal will not be affected, and both their membership benefits and automatic renewals will remain unchanged. Existing users whose subscriptions had already expired when the notice was issued may purchase a legacy plan within 7天 of the announcement. Upgrades within legacy plans, such as moving from the 199元 tier to the 699元 tier, will not be affected by this adjustment. Kimi also said it will continue to evaluate the policy as computing capacity expands and relax the 7天 restriction wherever possible.
Manus AI makes Manus 1.6 Lite and Manus 1.6 free for all users through August 25
要闻
Manus AI has announced that Manus 1.6 Lite and Manus 1.6 will be available free of charge to users worldwide from today through August 25, Singapore time. Paid users will receive priority access in the queue.
Manus AI has launched a free-access promotion. From now until 11:59 p.m. on August 25 (Singapore time), free and paid users worldwide can use Manus 1.6 Lite and Manus 1.6 at no cost within the promotion’s daily quotas. Paid users receive priority access in the queue for both models. Free users can generate up to 20 images and 1 video per day, while paid users can generate up to 200 images and 10 videos per day. Manus 1.6 Max is not included in this promotion. The promotion is subject to mechanisms including availability only while quotas last, risk controls, and dynamic queuing.
Claude Cowork arrives in the Chrome sidebar with cross-device session continuity
要闻
Anthropic has announced that the Claude in Chrome browser sidebar is being upgraded to Claude Cowork sessions. Conversations and context can now sync across devices. The feature is currently available to Max and Team subscribers and will gradually roll out to Pro users over the coming weeks.
Anthropic has officially announced that the browser sidebar in the Claude in Chrome extension has been upgraded to a full Claude Cowork session. Users’ conversation history, skills, and connectors can now sync seamlessly and carry over between the browser and Claude’s desktop, web, and mobile apps. Because sessions are linked directly to users’ accounts, users can start long-running tasks in a browser tab and later continue working on them from other devices. The feature is currently available to Max and Team subscribers and will be rolled out gradually to Pro users.
Qwen officially releases Qwen3.8-Max model weights
模型发布
The Qwen team has officially released the model weights for Qwen3.8-Max. The open-source version is named Qwen3.8-2.4T-A95B. The released weights are for a text-only model that requires thinking mode during interactions. It does not support multimodal input or the option to disable thinking.
The Qwen team has released the model weights for Qwen3.8-Max on relevant platforms, naming the open-source version Qwen3.8-2.4T-A95B. The released weights are for a text-only model that requires thinking mode during interactions. It does not support multimodal input or the option to disable thinking. Its native context length is 256K and can be extended up to 1M. Developers can currently deploy it using frameworks such as vLLM and SGLang, or access the official Qwen3.8-Max API through Qwen Cloud, which offers more features and supports a 1M context window by default.
微信 introduces the WeLM model family as 腾讯 prepares to launch Hy4 soon
模型发布
The 微信 team published an article introducing its internally developed WeLM large language model family, including the already deployed WeLM-80B and the in-development WeLM-617B. 腾讯 also stated in its earnings report that Hy4, which further improves model performance and multimodal capabilities, will be released soon.
The WeChat team introduced its self-developed WeLM large language model family in a post on its X account, highlighting resource efficiency as a key feature. The family includes WeLM-80B (80B parameters, 3B activated) and WeLM-617B (617B parameters, 23B activated, MoE architecture, under development). WeLM-80B has been deployed in Xiaowei, WeChat’s native AI assistant, supporting chat-based search, native WeChat features, and Mini Program services. Designed for complex tasks across the WeChat ecosystem, WeLM-617B is currently under development. In addition, Tencent’s Q2 2026 earnings report stated that Hy4 will be released soon to further improve model performance and multimodal capabilities.
Google DeepMind launches SL2T sign language translation model on Pixel 11
模型发布
Google DeepMind has released a new sign-language-to-text (SL2T) translation model. American Sign Language (ASL)-to-English transcription is now available free of charge in Gboard and the Live Transcribe app on Pixel 11.
Google DeepMind and the Android team have jointly released SL2T, a multilingual sign-language-to-text translation model, and deployed it in a consumer product for the first time. The initial feature supports transcribing American Sign Language (ASL) into English, allowing deaf and hard-of-hearing users to use sign language for searches, drafting messages, and everyday conversations in Gboard and Live Transcribe on Pixel 11. The feature is currently available free of charge on Pixel 11, with support for more devices and languages planned for the future.
Liquid AI releases on-device vision model LFM2.5-VL-3B
模型发布
Liquid AI has released the open-source vision-language model LFM2.5-VL-3B. Designed for on-device deployment, the model focuses on fast, direct answers and improves on its predecessor in screen understanding and function calling.
Liquid AI has released the LFM2.5-VL-3B vision-language model, combining the LFM2.5-2.6B language backbone with the SigLIP2 vision encoder. The model uses a direct-answer generation mode rather than reasoning, ensuring real-time responsiveness on edge devices. Compared with its predecessor, the new model delivers four major improvements: screen and UI understanding, object detection, multi-image input processing, and function calling. LFM2.5-VL-3B is now available on Hugging Face and supports multiple frameworks, including llama.cpp, vLLM, and ONNX, for cross-platform or GPU inference.
Cohere releases 2.4B vision-language model North Micro Vision
模型发布
Cohere has released North Micro Vision, an open-source vision-language model with 2.4B parameters. The company says the model outperforms Gemma 4 E2B and Ministral 3 3B on visual understanding benchmarks.
Cohere has released North Micro Vision, a vision-language model with 2.4 billion (2.4B) total parameters. According to the company, the model outperforms Gemma 4 E2B and Ministral 3 3B on visual understanding benchmarks, with particularly strong results in document understanding and visual question answering. North Micro Vision uses native-resolution visual processing and supports multi-turn image-and-text conversations, spatial reasoning, and multilingual image understanding. The model weights are now openly available.
SuperGrok Heavy users can claim a free one-month Cursor Ultra trial
开发生态
SuperGrok Heavy users can now claim a free one-month trial of Cursor Ultra. According to the official Grok account, users must sign in to Cursor with the same email address associated with their SuperGrok Heavy account to redeem the offer.
SuperGrok Heavy subscribers can now receive a free one-month trial of Cursor Ultra. The official Grok account said users must log in to or register for Cursor using the same email address as their SuperGrok Heavy account, then claim the free one-month Ultra trial when the offer appears. The benefit is not a permanent subscription. The trial has a separate usage quota pool and does not affect Grok's usage limits.
Zed launches Delta, a DeltaDB-based environment for collaborative Agent coding
开发生态
The Zed team has launched Delta, an application designed for collaborative coding with Agents, and opened a closed beta to its first group of users. Built on DeltaDB, Delta links conversations with code changes in real time and integrates third-party Agent tools such as Claude Code.
The Zed team has launched Delta, a multiplayer environment for writing and reviewing code collaboratively with Agents, and has begun inviting its first users to a closed beta. Delta is built on DeltaDB, the version control system previously developed by the Zed team, and synchronizes developers’ conversations with Agents, code edits, and review workflows in real time within the same thread. DeltaDB captures every action between commits rather than recording only snapshots, keeping code and conversations continuously linked. Delta allows threads to be opened directly in a browser via a link and already integrates with third-party Agent tools such as Claude Code. The Zed team said DeltaDB will eventually be integrated into the Zed editor, with Delta serving as its first client for now.
Users question apparent cuts to OpenAI Codex subscription quotas
开发生态
Multiple community users have reported reductions in OpenAI Codex Pro subscription quotas, with some saying their allowance was cut by nearly one-third. OpenAI has not yet responded.
According to multiple forum users, OpenAI Codex Pro subscription quotas appear to have been reduced. One user said that after subscribing for a week, their quota was more than $2,400 in the first week but recently fell to about $1,700, a reduction of nearly one-third. Another user reported an anomaly in which $5 of usage was shown as 5% and questioned whether it was an account issue or an OpenAI bug. The situation has not yet been officially confirmed.
OpenAI launches ChatGPT credit gifting, allowing users to buy credits for others
产品应用
OpenAI has launched a ChatGPT credit gifting feature that allows eligible individual users to purchase credits for others. The feature is rolling out gradually and is open to all individual accounts.
OpenAI’s official Help Center has published documentation for ChatGPT’s Gifting credits feature. According to the documentation, eligible users with individual ChatGPT Free, Go, Plus, or Pro accounts can purchase a standalone bundle of prepaid credits for someone else on the ChatGPT website. After completing the payment, the purchaser will receive a one-time redemption link, which can be sent to the recipient by email or by copying and sharing the link. After opening the link, the recipient must sign in to or create an eligible individual ChatGPT account and redeem the credits within 30 days. Once redeemed, the credits will be added to the account’s individual credit balance and can be used for credit-based features supported by its subscription plan, including ChatGPT Codex.
OpenAI gradually expands ChatGPT ad testing to more markets
产品应用
OpenAI has announced that ChatGPT ads have expanded to the United Kingdom, Mexico, Brazil, Japan, and South Korea. Adult users on the Free and Go plans in these markets, as well as in the United States and other regions, will now see ads.
OpenAI is expanding its advertising test in ChatGPT, which is currently live in the United States, Canada, Australia, New Zealand, the United Kingdom, Mexico, Brazil, Japan, and South Korea. The feature is shown to signed-in adult users on the Free and Go subscription plans, while Plus, Pro, Business, Enterprise, and Education plans remain ad-free.
Google to integrate Gemini with more third-party apps in the coming weeks
产品应用
Google has announced that Gemini will integrate with more apps over the coming weeks. The new services span work and everyday life, allowing users to complete tasks such as summarizing meetings and booking restaurants directly through Gemini.
Google announced that it will roll out a new batch of connected apps to Gemini over the coming weeks. These additional services will allow users to handle tasks directly within Gemini across four main categories: productivity and creativity, local and entertainment, music, and home health and lifestyle. By simply enabling these services in Gemini, users will be able to complete tasks such as summarizing meetings, booking restaurants, playing playlists, and scheduling doctor’s appointments, all in one place. These features are currently being rolled out gradually.
面壁智能 begins A-share listing guidance with 中信证券 as its adviser
行业动态
According to reports, 面壁智能 has begun guidance procedures for an A-share listing, formally moving toward an IPO, with 中信证券 serving as the adviser. The company is valued at more than 20 billion yuan, making it the highest publicly valued unicorn in China's on-device AI sector.
According to reports, ModelBest first submitted its listing counseling registration report to the Shanghai Securities Regulatory Bureau on August 12, formally beginning counseling for an A-share listing, with CITIC Securities serving as the counseling institution. Previously, on July 15, ModelBest completed a new funding round involving national-level funds, centrally administered state-owned enterprises, automakers and other industry investors, as well as prominent financial investment institutions. Its cumulative financing in the first half of 2026 exceeded RMB 5 billion, with its valuation surpassing RMB 20 billion. ModelBest was founded in Beijing in August 2022 and incubated by Tsinghua University’s Natural Language Processing Laboratory. Its core business is the research, development, and commercialization of on-device large models, and its flagship product is the MiniCPM model series.
Elon Musk says Grok 4.7 is expected to be ready in three to four weeks
前瞻与传闻
Elon Musk said the initial training of Grok 4.7 has been completed and that the model is now undergoing additional training with a large volume of SpaceX company data. It is expected to be ready in three to four weeks. He claimed the model will surpass all current models and that SpaceX's training corpus will make it unmatched in real-world engineering.
Elon Musk announced that the initial training of Grok 4.7 has been completed and that a large volume of SpaceX company data is now being added as supplementary training material. The model is expected to be ready in 3 to 4 weeks. Musk said that Grok 4.7 will surpass all current models, adding that given the uniqueness of the SpaceX training corpus, he would be very surprised if any model outperformed Grok 4.7 in real-world engineering.
Note: This content was created with AI assistance and may contain hallucinations and errors.
