Kimi.ai apologizes for K3.1 Pro rate limits and resets usage
要闻
Kimi.ai apologized for K3.1 Pro rate limits and unnecessary quota consumption and issued a one-time usage reset for subscribers across all tiers, restoring up to 25% of their monthly quota. The company also said it is reducing unnecessary subagent calls and output length.
Kimi.ai has apologized for rate limits and unnecessary quota consumption affecting its K3.1 Pro subscription plans and implemented a one-time usage reset for subscribers across all tiers; the source does not specify a date. The adjustment restores usage equivalent to at most 25% of the monthly quota, while users who have consumed less than 25% have their recorded usage reset to zero. The company said it was responsible for the problems and should have fixed them earlier. It described the restored quota as equivalent to one week of usage and said the usage bar would reflect the adjustment by the end of the announcement day in US Pacific Time. In addition to having adjusted its configuration, the company said it is reducing unnecessary subagent calls and output length to lower quota consumption. Kimi’s official developer account separately said it had optimized subagent calls for K3.1 Pro in Kimi Code CLI and the Kimi Code desktop application, with the aim of improving orchestration and collaboration on complex tasks.
OpenAI’s 28-day plan gets a full reset on day 20; Tibo says changes have fully propagated
要闻
Tibo of OpenAI’s Codex and ChatGPT teams said day 20 of the 28-day plan brought a full usage reset for everyone, with all related changes fully propagated. The plan delivers a product improvement or a usage reset each day, sometimes both.
Tibo of OpenAI’s Codex and ChatGPT teams said a full usage reset for everyone took place on day 20 of the 28-day plan, and all related changes had subsequently finished propagating. According to his explanation, the daily schedule provides either one improvement for Codex or ChatGPT Work users, or a full usage reset for everyone, with both sometimes taking place on the same day. Tibo also invited users to use Codex the day after day 20 and said the following week would bring interesting developments, without specifying what they would be.
CapCut plugin to integrate with ChatGPT, with public testing starting in November
产品应用
CapCut has unveiled a plugin workflow connecting Codex with video generation and editing tools and announced testing for November, without specifying a year. According to the company, users can generate drafts and modify timelines through conversation, then switch to CapCut for manual editing.
CapCut announced that its plugin will connect to ChatGPT and open for testing in November, although the page provides neither a specific year nor an exact date. According to the company, the workflow, called CapCut × Codex, requires users to install the plugin in Codex and create a task. Users can generate images, video material, and drafts with natural-language instructions, or upload existing drafts and supporting media to continue production. Drafts remain editable, with conversational controls for shortening scenes, reordering clips, replacing visuals, and adjusting text timing and music. Users can also open the CapCut editor directly from the conversation to modify individual tracks manually. The company says the plugin already includes creation capabilities powered by CapCut Skills, supporting caption generation and revision, preparation of different language versions, and matching with hundreds of thousands of creative assets, including effects, transitions, filters, stickers, and music. Users can also select a template, replace its media, and then refine the details on the timeline.
OpenCode discloses operational data for DeepSeek inference racks
开发生态
OpenCode team member dax reported that a DeepSeek inference configuration on a new GPU rack processed 6.77 trillion tokens over the past 24 hours, with p50 time to first token of 1.16 seconds and p50 throughput of 143 tokens per second.
OpenCode team member dax disclosed operating data for a DeepSeek inference configuration on a new GPU rack, covering the 24 hours before the disclosure; the source did not specify the disclosure date. The configuration processed 6.77 trillion tokens, with p50 time to first token of 1.16 seconds and p50 throughput of 143 tokens per second. One tray in the rack was unstable, affecting p95 performance. Asked whether quantization was used, dax said the service was identical to DeepSeek’s official API. He had previously stated that, in the first test, the configuration delivered better price-performance than the other providers he had found.
BOSS Zhipin’s LLM team Nanbeige has open-sourced Nanbeige4.3-33B under Apache-2.0 with a 512K context window; the team reports a Terminal-Bench 2.1 score of 80.9.
BOSS Zhipin’s LLM team Nanbeige has released the open-source model Nanbeige4.3-33B, though the source does not specify a release date. The model uses a 33B-parameter dense architecture, supports a 512K context window, and is available under Apache-2.0. For deployment, the community has already provided GGUF quantized versions based on the original BF16 weights, with disk usage ranging from 13.7GB to 35.2GB. The model will also serve as a teacher for subsequent smaller models, starting with Nanbeige4.3-3B. According to the team, its Agent performance reaches SOTA among open-source models with approximately 30B parameters. The official evaluation reports a Terminal-Bench 2.1 score of 80.9, compared with 73.0 for Qwen3.8-27B.
Anthropic recommends that enterprises fold some subagents into the main agent loop
技术与洞察
Anthropic is encouraging enterprises to evaluate folding some subagents into the main agent loop and is providing a migration kit. The company says comparisons across several enterprise deployments found that a single agent equipped with skills frequently beat subagent architectures on quality, often with lower cost and latency per task.
Anthropic has proposed folding some routers, planners, and domain subagents into the main agent loop in an article on the Claude website, alongside a migration kit that includes a mock-system demo; the source does not specify a publication date. According to the company, the alternative uses skills and tool search to load context and tools as needed rather than maintaining multiple layers of coordination, while subagents remain suitable for verification, parallel work, and modular permissioning. Migration assessment starts with a shared end-to-end eval set owned by a central team, supplemented by smaller eval sets that individual teams supply with their skills; 10 golden examples are enough to begin. Before consolidation, evaluation checks whether routing selected the correct subagent; afterward, it checks whether the agent loaded the correct skill or called the correct tool. The company also advocates having one central team own end-to-end quality and the system prompt, reviewing every new tool and eval before release.
Elvis Presley song rights holders sue Anthropic for about $41 million in damages
行业动态
According to Billboard, the rights holders to Elvis Presley’s songs have sued Anthropic, seeking approximately $41 million in damages. The report’s headline frames the dispute as alleged theft of his music, an allegation made by the plaintiffs.
The rights holders to Elvis Presley’s songs have filed a lawsuit against Anthropic seeking approximately $41 million in damages, but the supplied Billboard material does not specify the filing date. The headline alleges that Anthropic stole his music, but the material does not identify the works involved, the alleged manner of use, or Anthropic’s response, nor does it state whether a court has ruled on the allegations.
