Google has launched the Gemini Decisions API, powered by Gemini Decision 1, to evaluate up to 1,000 questions in one request. Pricing is $0.10 per 1M input tokens, with no charges for output tokens.
Google has launched the Gemini Decisions API, available through the Gemini API in Google AI Studio and powered by Gemini Decision 1 (gemini-decision-1). The endpoint handles bounded decisions such as tool selection, rule checks, model routing, and context grading, allowing yes/no questions, multiple-choice questions, and rubric-based scoring to be combined against one shared input. Each request supports up to 1,000 independent questions and a shared input of up to 256K tokens, combining text, images, audio, video, and PDFs. According to Google, the model scores candidate options in a single forward pass rather than generating answers token by token, returning probabilities and confidence scores from 0 to 1.
According to Google, with multimodal inputs and contexts of up to 16K tokens, median model-compute latency is 50–100 ms and end-to-end latency is 150–250 ms, remaining nearly unchanged as the number of questions per request rises from 1 to 100. Pricing is $0.10 per 1M input tokens and $0 for output tokens. The shared input is billed once per request, with each question adding its own instruction and scoring-criteria tokens; there are no separate charges for thinking tokens, audio transcription, or OCR. Google AI Studio offers a no-cost tier for prototyping, and the documentation includes Python, JavaScript, and REST guides. Ecosystem integrations with OpenRouter, LangChain, Vercel, and others are rolling out, while support in Gemini on Google Cloud is still to come.
通义千问 announces Qwen4 is coming soon, with a preview already available
要闻
Alibaba’s Tongyi Qianwen has announced that Qwen4 is coming, with a preview already available through several channels. Qoder has also switched its anonymous test model to Qwen4-Max-Preview for individual, team and enterprise users, temporarily lowering its billing multiplier to 0.1×.
Alibaba’s Tongyi Qianwen has announced the upcoming release of Qwen4 and made the Qwen4-Max preview available through Alibaba’s Token Plan, although the source provides no release date for the official version. Individual subscriptions have four tiers, ranging from Lite to Pro, priced at 39 yuan to 312 yuan per month. Users can also get early access through QwenWork, Qoder, Qwen APP and QwenStudio. According to the official announcement, the new model improves knowledge understanding, world perception, coding and Agent capabilities, as well as its ability to operate professional software such as EDA, CAD, Blender and UE5.
Qoder’s rollout covers its full international and Qoder CN product lines, with access available across individual, team and enterprise editions. Qoduck, the model previously used for anonymous testing, has been upgraded to Qwen4-Max-Preview. Pricing is temporarily reduced to 20% of the previous rate, with the billing multiplier falling from 0.5× to 0.1×. According to the official explanation, this preview is a test version of Qwen4, and collected user feedback will be used to improve the official release.
OpenAI member hints at a possible Codex quota reset on Thursday
要闻
OpenAI Codex team member Tibo hinted on X that Thursday might be a suitable time for a Codex banked reset. The remarks appeared in two posts, but he did not confirm a reset or specify what would happen on Thursday.
OpenAI Codex team member Tibo mentioned Thursday and a banked reset in two posts on X, without announcing specific plans for Thursday or confirming whether a reset would take place. He described Thursday as a day that felt associated with Codex and suggested it might be a suitable time for a banked reset. That remark appeared in a follow-up post; in the earlier post, he noted that it had been a while since a reset and closed by suggesting that perhaps he should leave the basement.
Google DeepMind has released AlphaGenome 2 and opened API access for non-commercial research. The company says the model was trained on more than 107 million individual cell profiles and identifies nearly twice as many variants affecting gene expression in cells as its predecessor.
Google DeepMind has released the DNA sequence model AlphaGenome 2, with access through the AlphaGenome API for non-commercial research and a private preview for commercial research. According to the company, the new model shifts predictions from tissue averages to specific cell types and states. Its training used human and mouse bulk-tissue and single-cell data, covering 68 tissues and more than 107 million individual cell profiles. The company says AlphaGenome 2 identifies nearly twice as many variants that change gene expression in cells as its previous state-of-the-art model.
Architecturally, AlphaGenome 2 replaces simultaneous predictions across multiple tissues with inputs combining a one-million-letter DNA sequence and a specified tissue, cell type or state, predicting DNA effects for one condition at a time. It adds DNA methylation to its prediction targets alongside gene expression, histone modifications and chromatin accessibility. The company says researchers can upload single-cell RNA-seq data from just a few dozen cells, with a cell encoder generating a representation of the cell state, to predict variant effects in previously unseen states without retraining. It can also infer unmeasured properties from existing measurements.
Shanghai Innovation Institute has open-sourced Nex-N2.6, a multimodal Agent model with approximately 321B parameters that accepts text, image, and video inputs. Its FP8 quantized weights are available on GitHub and Hugging Face, and its API provides a 256K context window.
Shanghai Innovation Institute has released and open-sourced Nex-N2.6, built on GLM-5.3-Flash, while also making its API available and launching it on OpenRouter and the MOSS open platform. The model has approximately 321B parameters and supports text, image, and video inputs. Its API offers a 256K context window, and its FP8 quantized weights have been published on GitHub and Hugging Face. According to the institute, compared with Nex-N2.5-Pro, the new model's repetition rate during reasoning fell from 16.0% to 1.6%, a relative reduction of 90%, while the overall non-looping call rate rose to 97.8%. In tests using 956 real-world contexts, both effective tool-selection metrics reached 100%, matching GPT-6-Sol. In evaluation results self-reported by the development team, Nex-N2.6 scored 83.0 on the computer-control benchmark OSWorld and 84.1 on the command-line-only benchmark Terminal-Bench 2.1.
百度 launches ERNIE 5.5 Flash Preview, with API now available on 千帆
模型发布
Baidu has released ERNIE 5.5 Flash Preview and made its API available on Qianfan. The capabilities listed by the company include a 1M-token context, along with Coding, 3D modeling, knowledge work, and content creation.
Baidu has released ERNIE 5.5 Flash Preview, with its API now available on Qianfan; the research page labels the product update with the date November 11, 2026. According to the company, the model offers a 1M-token context, with listed use cases including Coding, 3D modeling, knowledge work, and content creation. The page categorizes the item as Featured / Product updates, and the supplied source text does not specify API pricing, invocation methods, or detailed performance test results.
Runway releases weights for its Praxis-1 robotics model
模型发布
Runway has released Praxis-1’s weights and inference and fine-tuning code to robotics developers and researchers under Apache 2.0, allowing commercial use. The model has 27B parameters and achieved a 46.2% success rate on RoboLab-120, according to the company.
Runway has released Praxis-1’s weights and inference and fine-tuning code under Apache 2.0, without specifying an exact release date in the source article. The release permits commercial use and supports full fine-tuning and LoRA. Developers can access resources through GitHub, Hugging Face and vLLM-Omni, with native LeRobot support on Hugging Face and hosted inference on Reactor. According to the company, the model runs out of the box on NVIDIA H100, B200, RTX 5090 and RTX PRO 6000 hardware, as well as Jetson Thor on the robot itself.
Praxis-1 is a 27B-parameter mixture-of-experts diffusion transformer that predicts video and actions together. According to company-published data as of November 2026, it achieved a 46.2% success rate across 1,200 episodes on RoboLab-120, compared with 42.9% for FLUX 3 Action, 36.8% for Cosmos3 Nano and 28.0% for π0.5. The company also says a single checkpoint has been adapted to different partner robots through light fine-tuning. After fine-tuning on Google’s RT-1 robot, which was absent from its training data, Praxis-1 achieved an 80.0% success rate on SimplerEnv, compared with 52.4% for the original RT-1 policy.
小红书 upgrades its dots.tts.edit speech editing model
模型发布
Xiaohongshu released dots.tts.edit V1.1 in October 2026, adding free-form style editing, and open-sourced its training-data production Data Pipeline. The team says the update improves prosody and emotion editing.
Xiaohongshu released dots.tts.edit V1.1 in October 2026 and open-sourced the Data Pipeline used to produce training data for the speech editing model that same month. The update introduces free-form style editing. According to the team, it improves prosody and emotion editing and achieves state-of-the-art results for instruction following and local preservation across editing categories including text, emotion, and prosody. Developers access speech editing through the separate dots.tts.edit entry point, which requires source audio, a tagged instruction, and an output path. Source and target transcripts are optional and are derived from the instruction when omitted; explicitly supplied non-empty transcripts override those derived results. For evaluation, doteBench V1.0, released in September 2026, is an XML-first bilingual Chinese-English Benchmark containing 2,081 cases. It covers text, emotion, prosody, pause, and compositional edits, evaluating instruction following, local preservation, and audio quality.
DeepL unveils Aure, a speech model that reproduces tone and emotion in real time
模型发布
DeepL released its Aure voice model on November 11, 2026, with initial support for 8 languages, including Japanese. The company says the model is automatically applied to DeepL Voice and follows changes in a speaker’s tone, emotion and pace during real-time translation.
DeepL launched Aure, a voice model for DeepL Voice, on November 11, 2026. Its new capabilities apply automatically to DeepL Voice for Online Meetings, DeepL Voice for In-Person Conversations and the DeepL API for Voice, without additional configuration. Initial language support covers English, German, Spanish, French, Italian, Japanese, Portuguese and Korean. According to the company, Aure preserves recognizable characteristics of a speaker’s voice while continuously combining source audio and translated text to adjust speech output. Rather than determining the voice solely from a sample at the start of a conversation, it allows translated speech to follow changes in emotion, emphasis and rhythm.
According to DeepL, Aure’s training separates voice identity, prosody and spoken content so that each can be transferred independently into another language. Reinforcement learning further adjusts the output to balance translation accuracy and speech quality. The model also needs to identify speaker characteristics amid background noise, echo and audio distortion, and adapt to a voice in real time before a sentence is finished. The company says DeepL Voice does not permanently retain meeting transcripts or translations and deletes them from memory when a session ends, limiting the real-world data available for learning how the same speaker talks across languages.
Alibaba has released Qwen Code 1.0 Preview, bringing Harness capabilities refined in Qoder into an open-source core under the Apache 2.0 license. It provides a CLI and SDK for developers to embed Agent execution capabilities in their own products.
Alibaba has released Qwen Code 1.0 Preview on a separate branch while retaining the existing stable branch; the source does not specify a release date. According to the company, this version integrates Harness capabilities refined in Qoder into a shared open-source core, keeps the code under the Apache 2.0 license, and provides a terminal CLI and SDK. It incorporates SubAgents and MCP, connects entry points and tasks through a unified event mechanism, and supports model access across multiple protocols, including OpenAI, Anthropic, Gemini, and Qwen, as well as dynamic workflows. Community PRs are not being accepted at this stage, with acceptance expected to open after December. The company also states that users who log in to Qwen Code with a Qoder account from November 11 to November 24 can claim a free Qwen4 trial allowance, valid for 30 days after it is claimed; the source does not specify the year for these arrangements.
Google Antigravity offers Browser use for Agent-driven web interaction and application inspection. Personal Google Account users on App v2.30.0+ can trigger browser subagents automatically, while the CLI connects to local Chrome v144+.
Google Antigravity offers Browser use, with the App supporting automatic browser subagent launches through natural-language requests for users signed in with a personal Google Account starting with v2.30.0+, without prior Chrome setup or a slash command. According to the official documentation, the main Agent reads pages directly for documentation questions answerable from static text, but delegates to a browser subagent for live rendering, clicks, form input, responsive layout checks, screenshots, console or network inspection, and signed-in pages. Users can also invoke it explicitly with /browser. The subagent can interact with a running application, inspect errors, and verify visual fixes before completing a task.
The App’s built-in browser sits in a side panel beside the conversation and supports multiple subagents working concurrently. Each subagent uses an independent context window and dedicated tab, and cannot operate tabs belonging to other subagents or opened by the user. According to the official documentation, permission controls combine permission presets with protections for sites, uploads, downloads, and page isolation. The CLI instead controls Google Chrome on the same machine through a Chrome DevTools MCP server launched with npx, requiring Node.js, npx on PATH, and Chrome v144+. Users must enable remote debugging, submit a task and a clear completion check through /browser, and allow the connection within 60 seconds of Chrome’s prompt.
Gemini API Managed Agents adds support for self-hosted Linux sandboxes
开发生态
Google AI Studio announced that Gemini API Managed Agents can now connect to self-hosted Linux sandboxes. Tool operations run on developers’ infrastructure, while model calls and the Agent loop remain on Google’s side.
Google AI Studio has announced support for connecting self-hosted Linux sandboxes to Gemini API Managed Agents, allowing developers to delegate shell commands and file operations to workers they operate themselves. Managed Agents is a server-side Agent service in the Gemini API that uses Gemini models and tools to handle multistep tasks; connecting a self-hosted sandbox does not relocate model calls or the Agent loop, both of which continue to run on Google’s side. According to the official explanation, the connection uses the open-source gemini-api CLI and outbound HTTPS, with no need to open inbound firewall ports. The related APIs are currently in Beta, while the environments and Managed Agents are in preview, and features and data structures may change. The official documentation requires workers to run in isolation inside a container, microVM, or virtual machine, with local mode reserved for testing.
