OpenAI releases Images 2.6 with animated GIF generation and a 50% speed boost
要闻
OpenAI has released Images 2.6 across ChatGPT, Codex, and the API. The model can generate animated GIFs from text or photos, while the API image models are, according to the company, 50% faster to generate.
OpenAI deployed Images 2.6 to ChatGPT, Codex, and the API on the day of its release. The model can create animated GIFs from text prompts or uploaded photos, then lets users adjust playback speed and edit the result frame by frame. Users can also save images of themselves, pets, or characters as reference images and reuse them in later generations without uploading them again.
The API now also offers GPT-Image-2.6 Flare and Sunburst. OpenAI says both models generate images 50% faster than their predecessors. At low quality, Flare can generate an image in as little as 6 seconds and improves layout quality for visuals such as posters and product images. Sunburst targets higher image quality and maintains more consistent detail across multiple edits at medium and high settings. Developers can mark areas for modification on an image, use the complexity parameter to choose low, medium, or high planning levels, and export SVG. The default complexity is high and carries no additional charge. Images 2.6 is available from the day of release across all subscription tiers on web, desktop, and mobile. The two API models are priced per token: $8 per million tokens for image input, $30 for output, and $2 for cached input, the same as version 2.5.
Tencent releases the official version of Hunyuan Hy4
要闻
Tencent Hunyuan has released the Hy4 production model, adding vision and computer-use capabilities for the first time. Its hallucination rate on financial tasks fell to 5.45%, while API pricing remains $0.834 per million input tokens and $2.501 per million output tokens.
Tencent Hunyuan has released the Hy4 production model. According to the company, unlike Hy4 preview, which had 770B total parameters, 49B active parameters, and a 1M context window, Hy4 introduces vision and computer-use capabilities for the first time and can combine command-line and graphical interfaces to handle professional tasks. Its game-development evaluation score rose 19.1% from preview. In testing conducted with Tencent fintech organization FiT, 213 practitioners evaluated 579 real-world financial tasks; Hy4 received an overall score of 3.04, ranking first among open-source models, while its hallucination rate fell from 15.76% to 5.45%.
Through WorkBuddy, Hy4 can connect to WeCom chats, email, documents, and calendars, identify pending tasks, and remind users of deadlines. It ranked first among open-source models in the TenureTrail research evaluation, and its combination with Hyra 2.0 increased internal pretraining throughput by 20.1%. HyCell and HyPDE, developed through the collaborative research effort, have also been open-sourced. Hy4 is available on Tencent Cloud and OpenRouter, with API pricing unchanged from preview at $0.834 per million input tokens and $2.501 per million output tokens.
Report: OpenAI executive Tibo takes over an internal safety project
要闻
Reuters reports that Thibault Sottiaux, known as Tibo and an OpenAI executive responsible for product development, recently took over an internal security project; a source said he had previously assigned one-quarter of production engineers to improve security infrastructure.
OpenAI product-development executive Thibault Sottiaux, known as Tibo, has recently taken responsibility for a project protecting the company’s internal systems. Reuters cited a source familiar with the matter as saying that Tibo had previously assigned one-quarter of the production engineering team to work on improving security infrastructure. The report also said that Tibo had earlier posted that he had retired, then reposted a post by OpenAI employee Andrew Ambrosino. Ambrosino said he would receive feedback on all applications and work on improving their simplicity and performance. Based on these developments, some users speculated that Tibo’s focus may have shifted and that Ambrosino could take over part of the user-facing work.
Cat Wu announces she is joining OpenAI to lead product management for Core Products
要闻
Cat Wu announced on X that she has joined OpenAI to manage product work for Core Products. She will help build agents that preserve user control, and said OpenAI has brought AI to more than 1 billion people.
Cat Wu has joined OpenAI’s Core Products product management team, with the appointment announced by her on X. She said she will work with Tibo, the product management team, and teams across OpenAI to build agents that support users and teams while keeping users in control. She also said the company has shown the world that scaling works and has brought AI to more than 1 billion people, adding that excellent products are central to OpenAI’s mission. OpenAI co-founder and president Greg Brockman and team member Tibo later quoted the post to welcome her. According to Cat Wu’s profile, she previously worked on Claude Code at Anthropic and also worked at Dagster and Scale AI.
Users discover that ChatGPT subscriptions through Google Play come with API credits
开发生态
On November 5, 2026, multiple community users reported that ChatGPT Pro accounts subscribed through Google Play could claim API credits from the billing overview, including accounts without a payment card on file.
On November 5, 2026, multiple community users reported that accounts with ChatGPT Pro subscriptions purchased through Google Play could claim API credits from the billing overview page. A subscription page shared by users stated that an API organization must be linked before the credits can be used. Some users said the claim option appeared only after refreshing the billing page several times. Accounts without a linked payment card were also reported to display the same option.
Claude Code adjusts game restrictions: games with cash-out features are no longer rejected, but recreating original games remains prohibited
开发生态
Anthropic engineer Addy Osmani announced that Claude Code 2.1.317 has loosened Claude Opus 5.5’s game-development restrictions: original games may now reference existing titles, while direct remakes remain blocked.
Anthropic engineer Addy Osmani announced that Claude Code 2.1.317 changed the game-development restrictions for Claude Opus 5.5. Users can name an existing game and ask Claude to develop an original game of the same type by referencing its gameplay mechanics; merely mentioning an existing title no longer triggers an automatic refusal. When a request explicitly asks for a copy of the original, Claude will still refuse to reproduce it and will instead suggest creating a game with a similar experience.
Anthropic had previously added two rules to the Opus 5.5 system prompt to prevent users from copying existing games with Claude Code. Those rules could also lead to refusals when users cited an existing game only as a gameplay reference. Osmani said the change is intended to let Claude provide as much help as possible when users build original content while protecting developers’ work when others attempt to copy it.
Antigravity CLI 1.5.0 released with agy acp built in
开发生态
Antigravity CLI 1.5.0 is now available with built-in agy acp, remote agent sessions, and local project management, while enabling Google login from terminals without a browser.
Antigravity CLI integrated agy acp into the CLI in version 1.5.0, allowing it to run over stdio as an Agent Client Protocol (ACP) server. Users connecting through SSH or working in a terminal without a browser can also run agy acp --login to complete Google login. The release additionally enables /btw sidebar questions to read files or call tools during an ongoing conversation and adds the --settings parameter for overriding configuration at startup. It also introduces remote agent sessions through --remote-agent and local project management through /projects, and fixes an issue in which some models failed immediately and stopped retrying after rate-limit or server errors.
Google launches TTS Voice Remixing for Gemini 3.8 Flash
开发生态
Google has launched Voice Remixing for Gemini 3.8 Flash TTS in Google AI Studio and the Gemini API. Users can start with more than 2,000 preset voices or a custom-designed voice and use natural-language prompts to adjust timbre, pitch, speed, and accent.
Google announced that Voice Remixing for Gemini 3.8 Flash TTS is now available through Google AI Studio and the Gemini API. Users can select from more than 2,000 preset voices or begin with a custom-designed voice, then use natural-language prompts to change the timbre, pitch, speaking rate, and accent. The official examples include adding a slight Southern American accent.
Google AI Studio generates three versions for side-by-side auditioning, allowing users to select one and continue remixing it. The completed voice can be saved and reused across different scripts and scenarios with multiple speakers.
ChatGPT dots updated with improved experiences on iOS and Windows
产品应用
OpenAI ChatGPT team member samir announced updates to dots in the latest app. The iOS features ship with ChatGPT for iOS 1.2026.279, while Windows also improves task continuity and tool workflows.
OpenAI ChatGPT team member samir announced that a set of dots updates is available in the latest app; the iOS features correspond to ChatGPT for iOS 1.2026.279 and require dot access permission. According to the explanation, iOS now lets users reply to dot directly from a notification, share content with dot, and open dot through the iPhone Action Button. Loading after opening from a notification is faster, and Gmail links can open the corresponding message directly in the Gmail app. Call connection time has been reduced, and saying goodbye ends a call without manual disconnection. On Windows, task continuity between dots and local threads was improved, along with the desktop and browser experience and the reliability of workflows using local MCP tools and plugins.
Anthropic updated the Claude desktop app to v3.1623.0, adding a Storage panel and the /storage command. Users can inspect local disk usage and remove some data that has been inactive for more than 30 days.
Anthropic updated the Claude desktop app to v3.1623.0 and added a Storage panel under Settings. The panel breaks down local disk usage into Cowork tasks, Code sessions and worktrees, Runtime, cache, and logs. Users can remove Cowork tasks and Code sessions that have been inactive for more than 30 days, as well as worktrees with no uncommitted changes. Deleting a session permanently removes its conversation history, while deleting a task also removes its attachments and output files. Runtime required to run the app cannot be removed from this panel.
The new /storage command is available on the app’s Cowork and Code pages and can be used to view the storage information described above. The command is not supported in the terminal, and the Storage feature is not available for third-party deployments.
Ricursive Intelligence releases the 65B dense model Ripple-65B Preview
模型发布
Ricursive Intelligence has released Ripple-65B Preview, a general-purpose dense model pretrained on 12 trillion tokens with up to 1M context length, scoring 80.52 for U.S. open-weight SOTA in Agentic Coding.
Ricursive Intelligence has released Ripple-65B Preview as a general-purpose model. It was pretrained on 12 trillion tokens and supports up to 1M context length. According to the company, it scored 80.52 on TerminalBench 2.1, ahead of Reflection Beam 501B at 80.1, Meta Muse Spark Open 816B at 76.4, Thinking Machines Inkling 975B at 63.8, and NVIDIA Nemotron 3 Ultra 550B at 56.4. The model is available through Hugging Face, Together AI, GMI Cloud, and OpenRouter, while Fireworks is scheduled to make it available later the same day. Its dense architecture is designed for low-latency token generation and efficient execution on local devices such as NVIDIA DGX Spark.
According to the company, the total cost of developing and training Ripple-65B Preview from scratch was $32 million, including pretraining ladder development, data mixing, evaluations, debugging, synthetic data generation, posttraining recipe development, and the final training runs. The team did not purchase posttraining data; it built synthetic data pipelines, quality filtering, and coverage metrics instead. Its mixed RL recipe reached a TerminalBench 2.1 score of 80.5 with 1.4 million rollouts, roughly one-seventieth of the number used by Reflection Beam. The model uses a modernized dense Transformer rather than Mixture-of-Experts, maintaining 40% MFU during pretraining and posttraining. Its configuration includes a d_model of 5120, 80 total layers, global NoPE, gated attention, and value residuals. The company also states that, at the 4B model scale, the final model ladder improved code-training compute efficiency by up to 16x compared with the Olmo 3 ladder.
Cognition introduces SWE-2 Flash, claiming it outperforms every U.S. open-weight model
模型发布
Cognition launched the coding model SWE-2 Flash, saying it outperforms every American open-weight model on leading evaluations. It is available in Devin Desktop and CLI at one-fifth the price and three times the speed of SWE-2.
Cognition made SWE-2 Flash available today in Devin Desktop and CLI. According to the company, the coding model costs one-fifth as much as its larger model, SWE-2, and runs three times faster. Its score on Terminal-Bench 4.0 rose from 0.9% for the base model Inkling-Small to 20.0%, while its score on FrontierCode 2.0 rose from 3.2% to 24.0%. Cognition also said that SWE-2 Flash outperforms every American open-weight model on leading evaluations, including models three times its size.
The model was built on Inkling-Small, which has 276B total parameters and 12B active parameters, followed by mid-training on a few trillion tokens and then reinforcement learning. The training data included scientific articles, blogs, code, and other sources, while the RL stage expanded the task types and improved audits of instructions, environment setup, and verifiers. Training aimed at cybersecurity work raised its ExploitBench score from 11.3% to 34.8%. Cognition has not released the weights publicly because the same capabilities could be used by attackers. The company is building a program to share the weights securely with trusted industry partners and research labs, and said it will provide more information in a few weeks.
ByteDance has open-sourced the Axion 1.0 Agent foundation model under Apache 2.0, releasing its paper, training code, training data, and evaluation scripts alongside 35B and 122B MoE versions that activate 3B and 10B parameters respectively.
ByteDance released Axion 1.0 together with its paper, training code, training data, and evaluation scripts, with the overall release following the Apache 2.0 license. The Agent foundation model targets problem solving through scientific reasoning, information retrieval, and tool augmentation, and comes in 35B and 122B MoE versions that activate 3B and 10B parameters respectively.
Axion 1.0 was trained through two SFT stages and one RL stage, with a 128K context window used during SFT. The paper states that the model is competitive with leading commercial models on multiple benchmarks.
Moonshot AI releases the model weights for Kimi K3.1 Pro
模型发布
Moonshot has released the Kimi K3.1 Pro model weights, an natively multimodal Agent model in the 3T class with a 1-million-token context window and API access compatible with OpenAI and Anthropic.
Moonshot released the Kimi K3.1 Pro weights in this launch. According to the company, the model is the first entry in the Kimi K3.1 series and its most capable 3T-class model to date. It was post-trained from Kimi K3 with large-scale joint reinforcement learning across hundreds of downstream domains. Kimi K3.1 Pro supports a 1-million-token context window, with improvements listed in coding, 3D creation, financial research, factual accuracy, and other Agent capabilities.
The Kimi K3.1 Pro API is available on platform.kimi.ai by selecting kimi-k3.1-pro, with compatibility for OpenAI and Anthropic APIs. Thinking effort is configured through the top-level request field reasoning_effort, which supports low, high, and max; the default is max. For multi-turn conversations and tool calls, the complete assistant message returned by the API must be passed back into messages unchanged, including reasoning_content and tool_calls. Compared with Kimi K3, the model replaces full-attention Gated MLA layers with Gated DeepSeek Sparse Attention (Gated DSA) layers for more efficient long-context modeling. It is released under the Kimi K3.1 Pro License.
Anthropic says it measured the structure of a bacterial vault for the first time with Claude
技术与洞察
Anthropic said its research team used Claude to determine, for the first time, the structure of a bacteriophage-derived vault found in bacteria. The structure contains 96 L05 proteins, while purification time fell from the usual three days to about 90 minutes.
Anthropic’s research team said it used Claude to analyze the structure of a bacteriophage-derived vault, but has not yet determined the vault’s biological function. Claude was used to search protein databases, design DNA constructs, optimize the isolation method, and process microscopy data, while cryo-EM was used to produce a 3D model. The team said vault purification, which conventionally takes three days, was completed in about 90 minutes, and that the first structural model was produced 34 hours later. The vault described in the study consists of 96 copies of a single phage protein called L05, with 48 copies in each half, and has a combined mass of about 9.5 megadaltons. It is roughly 54 nanometers long and 45 nanometers wide. Mammalian vaults contain 39 copies per half and measure about 67 by 40 nanometers. The team said the structure’s ability to assemble inside bacterial cells could support research into vault assembly and exploration of nanocapsules for enzyme delivery.
OpenAI says an autonomous process generated patches for 948 Linux kernel vulnerabilities
技术与洞察
OpenAI said its security researchers built an autonomous workflow that reproduced 1,724 Linux kernel vulnerabilities in one run and generated candidate patches for 948 local privilege-escalation vulnerabilities at an API cost of $1,213,430.
OpenAI said its security researchers built an autonomous workflow that uses an upcoming cybersecurity model to find, validate, and patch Linux kernel vulnerabilities. In a single run, the workflow reproduced 1,724 vulnerabilities, generated candidate patches for 948 local privilege-escalation vulnerabilities, and produced proofs of concept for 291 information-disclosure vulnerabilities. The total API cost was $1,213,430. These local privilege-escalation vulnerabilities can allow an unprivileged local user to obtain root access.\n\nAccording to OpenAI, the Agent discovered and used at least 10 Linux vulnerability exploitation techniques that the company knew about and that experts consulted by it judged to be newly discovered. Human researchers will review the results, and OpenAI is working with Linux maintainers and Patch the Planet partners on the reports and patches.
Altman: Humanity should always control the future, not hand it over to superintelligence
技术与洞察
OpenAI CEO Sam Altman said on X that humanity should not hand control of its future to superintelligence and that AI development should be constrained by this principle. He also clarified his earlier comments about memory, saying he wants global production to increase.
OpenAI CEO Sam Altman said in an X post that humanity should continue to control the future and that OpenAI opposes reducing people to doing only the tasks an AI system considers worthy of rewarding humans for doing. He said people should not give superintelligence control over the future, adding that this requirement should ultimately serve as a constraint on the development of artificial intelligence.
After a user joked about Altman’s previous comments concerning memory, he replied that he wants everyone to have a large amount of memory. He also said he was angry that those comments had been taken out of context, and stated that he has long advocated increasing global memory production so people can install as much memory in their computers as they want.
DeepSeek discusses open-source safety and says it is studying approaches such as model tiering
技术与洞察
At the World Internet Conference Wuzhen Summit’s “AI Wuzhen Dialogue,” DeepSeek research engineer Gao Xi said the team is studying model tiering and other open-source safety measures, while noting that open source may amplify some AI safety risks.
DeepSeek research engineer Gao Xi said at the “AI Wuzhen Dialogue” during the World Internet Conference Wuzhen Summit that the team is studying proposals such as model tiering to address safety issues associated with open source. Gao also said that keeping models closed creates safety problems as well. DeepSeek chose open source to help AGI arrive sooner and to bring more people into discussion and validation by making its research methods public. The heads of Alibaba’s Qwen and MiniMax also shared their views on open-source safety and governance during the dialogue.
Artificial Analysis updates its API performance benchmark to measure models by total task duration
技术与洞察
Artificial Analysis has updated its API performance benchmark and made Time per Task the core metric. It evaluates the total time models take to complete each Intelligence Index task instead of relying solely on streaming output speed.
Artificial Analysis announced an update to its API performance benchmark, using the total time required to complete a task as the primary basis for comparison. The new Time per Task is calculated by dividing the model’s output token count for each Intelligence Index task by End-to-End Output Speed. Measurement runs from request submission through receipt of the final token, including the wait before the first token arrives, while output token counts use the figures reported by each provider’s API.
The revised benchmark incorporates Agentic workloads into all measurements and covers short prompts, document-level prompts, long-context prompts, and multi-turn sessions for programming and knowledge work. It also adds evaluations using each lab’s API default Effort setting.
OpenAI announces plans to acquire developer tools company Pierre to accelerate the development of the Codex collaboration platform
行业动态
OpenAI announced plans to acquire developer-tools company Pierre, bringing its open-source projects and Code Storage gradually into Codex to support collaborative software development with AI; the deal still requires regulatory approval.
OpenAI has announced its acquisition of Pierre Computer Company. The transaction has no confirmed closing date because it remains subject to customary closing conditions and regulatory approval. Until closing, the two companies will continue to operate independently. Pierre develops the open-source tools Trees and Diffs, along with the Code Storage code-hosting service; Trees and Diffs already support file browsing and code review in the Codex app.
According to OpenAI, the acquisition is intended to develop Codex into a platform where teams and AI can build software together from an initial idea through the final release. It will also adapt the Git workflows developers rely on for collaboration with coding Agents. After closing, Pierre’s open-source projects will remain open source, and OpenAI will continue contributing code. Code Storage customers will be able to keep using the service, while its functionality will be gradually integrated into Codex.
OpenAI and Mazda share research findings on automotive design
行业动态
OpenAI and Mazda announced automotive engineering research in which they used models to simulate airflow and structural behavior. With Astra in Codex, the teams built and simulated dozens of front-end designs to support vehicle design development and evaluation.
OpenAI and Mazda shared results from research that uses models in automotive design, including simulations of airflow and structural behavior to support the development and evaluation of vehicle designs. An OpenAI employee said its Forward Deployed Engineering team worked with Mazda aerodynamic engineers to use Astra in Codex to build and simulate dozens of front-end designs. According to OpenAI, the research explores performance improvements while taking styling and cost into account, with more progress and results to be shared over the coming months.
OpenAI says ChatGPT’s weekly active users in India have surpassed those in the U.S., making India the global leader
行业动态
Prabhjeet Singh, OpenAI’s managing director for India, said India has overtaken the United States to become ChatGPT’s largest market by weekly active users, with its user base doubling since January 2026.
Prabhjeet Singh, OpenAI’s managing director for India, said India now has the world’s largest ChatGPT user base by weekly active users, which has doubled since January 2026. The metric includes free and paid users who open ChatGPT at least once a week and is the only usage figure OpenAI publicly discloses. OpenAI said in February 2026 that India had passed 100 million weekly users; at that point, it was the company’s second-largest market and had the largest population of ChatGPT student users worldwide. User growth has been accompanied by local expansion. ChatGPT Go costs less than $5 per month, and Indian users who registered from 4 November 2025 could use it free for one year. OpenAI also announced plans for its first Indian office in Delhi NCR and said it would add offices in Mumbai and Bengaluru.
Meta has acquired AI coding startup Conductor. Conductor Cloud will shut down over the next few weeks, and the team plans to release a new coding interface early the following year.
Conductor officially announced that it is joining Meta, with the related changes taking effect following the announcement. The company launched in July 2025, with its early product centered on having Agents write code and coordinating multiple coding agents. According to the company, Conductor will continue operating normally until the new product is released, except for Conductor Cloud. Conductor Cloud will shut down over the next few weeks, and users can transfer their workspaces to local environments. Users who need assistance can contact [email protected]. The team said it plans to provide a new interface early the following year.
The China Academy of Information and Communications Technology and multiple companies release a self-regulatory pact for agent development
行业动态
On November 4, the Agent Development Self-Regulation Convention was formally released at the World Internet Conference Wuzhen Summit. It was jointly drafted by China Academy of Information and Communications Technology, Tencent, Ant, Alibaba, Douyin and other ecosystem participants, and sets out eight self-regulatory measures.
China Academy of Information and Communications Technology released the Agent Development Self-Regulation Convention with Tencent, Ant, Alibaba, Douyin and other Agent ecosystem participants at the World Internet Conference Wuzhen Summit on November 4. China Network Society Organization Federation and other organizations took part in the related work, while Huawei, OPPO, Xiaomi and Baidu were also listed among the initiating organizations.
The convention contains eight self-regulatory measures covering the development of an Agent trust framework, interconnected infrastructure, stronger security protections, user-rights protection, and broader access to and sharing of technology. Under the convention, the initiating organizations will lead implementation of its provisions and accept oversight from industry organizations and the public.
ChatGPT reportedly preparing five family subscription tiers
前瞻与传闻
X user Tibor Blaho found a Family Basic family-subscription lineup in ChatGPT’s web checkout pricing configuration. It contains five plans priced from $8 to $500 per month, although OpenAI has not announced the lineup.
A Family Basic family-subscription lineup has appeared in ChatGPT’s web checkout pricing configuration, where it was found by X user Tibor Blaho. The lineup lists five plans: Go at $8 per month, Plus at $20, Pro Lite at $100, Pro at $200, and Pro Max at $500. All prices are before tax, and OpenAI has not announced the subscription lineup.
