DeepSeek releases multimodal model DeepSeek-V4-Flash-Vision-Exp
要闻
DeepSeek has officially launched DeepSeek-V4-Flash-Vision-Exp, a new multimodal vision-understanding model, on the DeepSeek API platform. According to official data, its text-only capabilities are on par with DeepSeek-V4-Flash, while its multimodal Agent capabilities are close to Opus-4.8. The model supports three mainstream API formats. Images are converted into tokens and billed by token, with each image using up to 384 tokens at the same pricing as the V4-Flash model. DeepSeek has also launched Files at no charge.
DeepSeek has officially released DeepSeek-V4-Flash-Vision-Exp, an experimental multimodal vision understanding model, which is now available through its API platform. The multimodal API supports three formats: Chat Completions, Messages, and Responses. Official data shows that its text-only capabilities are on par with the production version of V4-Flash, its multimodal Agent capabilities are close to Opus-4.8, and its performance on vision understanding Agent benchmarks represents a substantial leap over V4-Flash. The model has a context length of 1M and a maximum output of 384K. Each image consumes no more than 384 tokens, and a single request can contain up to 600 images. Images are converted into tokens based on their dimensions and billed together with text. Pricing is the same as DeepSeek-V4-Flash. At the same time, DeepSeek has launched a free Files API, allowing users to upload images first and then reference them in requests, thereby saving bandwidth.
Tibo responds to abnormal Codex usage limits: subscription sharing via sub2api is flagged
要闻
Tibo responded to recent reports of abnormal Codex usage limits, saying that many affected users he spoke with were using sub2api. The system flags the practice of converting a subscription into API access for sharing among multiple people. However, some users said they had never used sub2api and still experienced abnormal usage.
Codex lead Tibo responded on X to recent claims of anomalies in Codex usage limits, saying that the limits would not be changed without engaging with the community and maintaining transparency. He said that, in conversations with affected users, the team found that many were using sub2api to convert subscriptions into API traffic and then resell it or share it among multiple users. Such usage is unsupported and will be flagged by anti-fraud systems. Using a subscription via “Sign in with ChatGPT” is permitted, including through official clients and OSS clients such as Pi and OpenCode that support account login and use of the included allowance. However, multiple users replied that they had never used sub2api and had only used official apps, yet still experienced abnormal usage consumption.
Tibo says Codex has reached 20 million active users and grants paid users a banked reset
要闻
Codex lead Tibo said Codex has reached 20 million active users and will grant one banked reset to paid Codex and ChatGPT Work users, which they can use whenever convenient. Addressing reports that usage limits were being consumed too quickly, Tibo said no anomaly had been found so far, though the investigation remains ongoing. The resets have now been distributed.
Codex lead Tibo said on social media that Codex had reached 2000 万 active users this week. To celebrate the milestone, the team will grant every Codex and ChatGPT Work user a one-time banked reset that can be used whenever convenient. He later provided an update saying that the banked reset would cover all paid ChatGPT Work and Codex users and had already been distributed. Addressing reports that usage allowances were being consumed too quickly, Tibo said no anomalies had been identified so far, but the investigation was still ongoing.
DeepSeek Harness updated to v0.1.1-rc.1 with official multimodal support
开发生态
DeepSeek Harness has received three consecutive updates since the initial public release of its developer preview. The latest version adds an option for the DeepSeek-V4-Flash-Vision-Exp multimodal model, while commands such as /goal and /plan can now accept text-and-image input.
The DeepSeek Harness team announced that since the first-day release of developer preview v0.1.0-rc.6, it has completed three consecutive updates—v0.1.0-rc.7, v0.1.0-rc.8, and v0.1.1-rc.1—focused on multimodal capabilities, subagent collaboration, cross-platform compatibility, and the everyday user experience. The model adapter now includes DeepSeek-V4-Flash-Vision-Exp as a multimodal model option and supports configuring native image requests. Commands such as /goal and /plan can accept text-and-image input, while the @ menu can reference files and sessions. Codex and Claude Code subagents have been integrated into the Job Panel, can be installed on demand as Profile Bundles, and support non-interactive permission modes and multiple named instances. Plugins can also register their own settings cards. The three updates also addressed community-reported issues involving long sessions, terminal compatibility, model requests, and sandbox permissions, while refining interaction details such as question cards, Markdown tables, multiline input, and session switching.
OpenAI cuts GPT-5.6 Sol API pricing by more than 20%
开发生态
OpenAI announced that it is cutting GPT-5.6 Sol API pricing by more than 20% for three months. The new pricing is already in effect for API users and pay-as-you-go Codex (ChatGPT Work) Credits. Usage included with Pro, Plus, and Business subscription plans remains unchanged.
OpenAI announced that it will cut GPT-5.6 Sol API prices by more than 20% over the next 3 months. The price reduction is already in effect for API users, and purchased Credits will support more calls for Codex users on usage-based billing. OpenAI emphasized that the pricing adjustment applies only to API and credit-based billing; the included usage allowances under Pro, Plus, and Business subscription plans will remain unchanged.
OpenAI Developers launches usage and spending tracking by API key
开发生态
OpenAI Developers announced that developers can now track usage and spending by API key and set hard monthly limits that stop traffic once exceeded.
OpenAI Developers announced that developers can now track usage and spending by API key in the usage and spending dashboard, making it possible to see which applications and workloads are driving costs. Developers can also set monthly spending limits at the organization or project level, including hard limits that stop traffic immediately once the limit is reached. These features are now available to developers using the OpenAI API.
Claude Security scanning adds support for Claude Mythos 5
开发生态
Claude announced that Claude Mythos 5 is now available for Claude Security code scanning in public beta for all Claude Enterprise customers. It has also established a $35 million open-source security fund and will expand the Cyber Verification Program over the coming weeks.
Claude announced that Claude Security code scanning is now powered by Claude Mythos 5 and is available in public beta to all Claude Enterprise customers. After users specify a GitHub repository at claude.ai/security, Claude Mythos 5 scans the codebase for vulnerabilities, with each finding including a CWE classification, confidence and severity ratings, and a recommended fix. Mythos scanning itself does not extend Mythos access to other interfaces. Scans are billed at standard token usage rates under existing subscription plans. The same announcement also unveiled integrations of Claude Mythos 5 with cybersecurity partners, a US$35 million Defender Advantage Fund, and plans to expand the Cyber Verification Program in the coming weeks.
ClaudeDevs announces improvements to Claude Code Remote Control
开发生态
Anthropic developer account ClaudeDevs announced reliability improvements to Claude Code Remote Control. Users can now start Claude Code sessions directly from their phones, and disconnected connections will automatically recover.
Anthropic’s developer account ClaudeDevs announced that a series of improvements addressing the Remote Control reliability issues most frequently reported by users last month are now live. Users can now start Claude Code sessions directly from their phones. Sessions remain synchronized between the phone and the machine; when a session is resumed on a laptop, the phone continues to display the live session, and after Claude Code exits, the phone shows it as offline within seconds. Sessions also recover automatically from brief disconnections. Mobile and CLI sessions remain synchronized on the model and effort level, and the behavior of slash commands such as /clear, /compact, and /diff on mobile has also been improved.
Google Antigravity launches Remote Control and headless terminal operation
开发生态
Google Antigravity has launched Remote Control, allowing users to remotely access ongoing sessions through supported browsers or mobile devices.
Google Antigravity announced the launch of Remote Control, allowing users to access active sessions from any modern supported browser or from iOS and Android mobile devices. The feature is beginning a gradual rollout to all users today, starting with Ultra subscribers.
Fish Audio extends free S2.1 Pro access for developers through November
开发生态
Fish Audio announced that it is extending free developer access to the S2.1 Pro voice model through November.
Fish Audio announced that its free access policy for developers using the S2.1 Pro voice model will continue through November. The company made S2.1 Pro available to developers for free three weeks ago, along with a free API. According to the company, the model is now the most-requested voice model on OpenRouter, ranking ahead of Google, Grok, MiniMax, and Deepgram.
Ollama makes Kimi K3 available to all Pro and Max subscribers
开发生态
Ollama announced that Kimi K3 is now available to all Pro and Max users with cloud subscriptions.
Ollama announced that Kimi K3 is now available to all Pro and Max cloud subscribers and can be used with the usage allowance included in their subscriptions. Ollama previously said that Kimi K3 had begun rolling out gradually to cloud subscribers, and its status has now been updated to fully available.
Vercel Developers launches is-agentic.com to test how Agent-friendly websites are
开发生态
Vercel Developers has launched is-agentic.com, a tool for testing how easily an Agent can read a website. It offers more than 100 audit checks, usage visualizations, one-click fix prompts, and a CLI.
Vercel Developers launched is-agentic.com, a tool for checking whether Agents can successfully read a website, which is now available through the site. According to the official announcement, the tool is based on research by @oradotai and supports audits with more than 100 checks, visualizations of how Agents use a website, one-click prompts for fixing issues, and a CLI for Agents. A user named Nick said that after running the CLI on his documentation site, the one-click fix feature rewrote the site's canonical tag. All feature descriptions come from the official announcement, while real-world performance has so far received only scattered user feedback.
NovelAI Diffusion V5 launches with two versions available immediately
模型发布
NovelAI has released the NovelAI Diffusion V5 image-generation model in two versions: V5 Curated and V5 Full. The new model is more than twice the size of V4.5 and adds features including native alpha transparency generation.
NovelAI released the image generation model NovelAI Diffusion V5, with both V5 Curated and V5 Full now available. Built on a proprietary architecture and improved from V4.5, the model is more than twice the size of its predecessor, officially supports prompts in English and Japanese, and uses a new proprietary 32-channel VAE to improve detail accuracy. New capabilities include native alpha transparency generation. The image generation UI has also been fully redesigned, with corresponding adjustments to subscriptions and usage limits.
Thinking Machines makes Inkling temporarily free on OpenRouter
模型发布
Thinking Machines announced that Inkling and Inkling-Small are now available for free on OpenRouter for several weeks, exclusively through agentic harnesses. The company said it will collect usage data to improve the models' agentic performance.
Thinking Machines announced that, to improve Inkling's agentic performance and better understand how it is used in real-world settings, Inkling and Inkling-Small are now available for free on OpenRouter for several weeks. Free access is limited to calls made through agentic harnesses. The company said it will use the collected usage data to improve the models, and that the data will be disassociated from user accounts.
GPT-Image-2 can now generate transparent-background images in ChatGPT
产品应用
GPT-Image-2 can now generate images with transparent backgrounds in ChatGPT and through the API.
An OpenAI staff member announced in a social media post that users can now generate images with transparent backgrounds using GPT-Image-2 in ChatGPT and through the API, and included an example image of a cactus they planned to print and stick on their laptop.
Grok Bot expands access to SuperGrok Plus, Cursor Pro+, and Cursor Teams users
产品应用
Grok Bot announced expanded access and is now available to SuperGrok Plus, Cursor Pro+, and Cursor Teams subscribers. All other users can try a free trial with limited usage.
Grok Bot has announced that it is now available to more users. All SuperGrok Plus, Cursor Pro+, and Cursor Teams subscribers now have access to Grok Bot, while all other users can try it for free with limited usage. The announcement does not currently specify the free trial’s usage cap, duration, or future pricing.
Kimi Code's 飞书 AI coworker Mira officially enters Beta testing
产品应用
Kimi Code announced the start of Beta testing for Mira, its 飞书 AI coworker. Mira operates within 飞书, acting as a coworker in group chats and an assistant in direct messages. It currently supports only 飞书, requires an application from Kimi Code members, and consumes membership usage allowances.
Kimi Code has announced the official Beta launch of Mira, an AI colleague embedded in Feishu. In group chats, it acts as an AI team member that takes on tasks and follows up on progress, while in private chats, it serves as a personal assistant. According to the company, Mira can learn proactively, with message organization, reviews, and scheduled checks running continuously in the background. This means it consumes model quota even when users do not actively ask questions, and overall usage may be substantial. During the Beta period, only Feishu is supported. Users must apply on the Kimi Code official website and pass a review, and Mira can only use the model quota included with a Kimi Code membership. Connecting a user’s own model API Key is not currently supported, and the company recommends using the K3 model.
豆包 adds skills, connectors, and work partners to work task mode
产品应用
豆包 announced the launch of three features in work task mode: skills, connectors, and work partners.
Doubao’s Work Task mode recently introduced three major features—skills, connectors, and work partners—upgrading its collaborative productivity capabilities. The platform currently offers more than 200 skills and connectors.
Anthropic publishes the AI-Native SDLC Practice Handbook
技术与洞察
Anthropic has published the AI-Native SDLC Practice Handbook, outlining an approach for transforming traditional linear software development processes into an Agentic AI-driven closed loop.
Anthropic has released the AI-Native SDLC Practices Guide, outlining a method for transforming the traditional linear software development process into an Agentic AI-driven closed loop. The guide divides the SDLC into six stages: planning, design, building, testing, deployment, and maintenance. It argues that code is no longer the bottleneck and that AI should be embedded into every stage, with artifacts such as intent.md and spec.md enabling automated workflows. For the building stage, it recommends using Claude Code’s plan mode and auto mode, together with CLAUDE.md, skills, and hooks to provide context and guardrails. For deployment and maintenance, it emphasizes integration through claude-code-action and MCP. Claude Design and Claude Tag are both currently in testing.
Anthropic publishes new research revealing interference weights
技术与洞察
Anthropic's interpretability team published new research identifying “interference weights” in a trained small language model that have almost no effect on outputs or even harm predictions. The study shows that a large numerical weight does not necessarily correspond to useful computation that the model genuinely relies on.
Anthropic’s interpretability research team has published a study examining the problem of “interfering weights” in a fully trained, single-layer miniature Transformer. The paper defines interfering weights as connections that are irrelevant or harmful to model behavior and proposes two metrics, effectiveness and helpfulness, which respectively measure whether a weight genuinely affects the model’s output and whether that effect improves or harms predictions. The research shows that numerically large virtual weights do not necessarily correspond to computations the model actually relies on. The authors argue that identifying interfering weights is necessary but insufficient for understanding a model’s global circuits, and that the key to further improving interpretability may lie in finding more appropriate ways to represent models.
Google DeepMind partners with Fenris Creations to advance game AI research
技术与洞察
Google DeepMind has shared details of its research collaboration with Fenris Creations, developer of the EVE universe. The research will begin with offline instances of EVE Online and use general-purpose Agents such as SIMA 2 to explore new frontiers in game AI.
Google DeepMind outlined the latest progress in its collaborations with game developers to advance AI and gaming research in an official blog post. According to the company, Gemini-powered SIMA 2 is a general-purpose Agent that requires no access to APIs or source code. Using only on-screen visuals, natural-language instructions, and keyboard-and-mouse controls, it can play complex 3D games such as No Man's Sky, Valheim, and Hydroneer in a human-like manner. The research will begin in offline instances of EVE Online isolated from real players, followed by studies of human-Agent coexistence through EVE Frontier. The technology will only be considered for introduction into EVE Online and EVE Vanguard once its capabilities are sufficiently mature.
NVIDIA says AVO Agent scored 100 on the ARC-AGI-3 public set
技术与洞察
NVIDIA announced that its general-purpose coding Agent AVO, powered by the Claude Opus 5 model, achieved a perfect score of 100.00 RHAE on the ARC-AGI-3 public set, completing all 183 levels across 25 environments.
NVIDIA announced that its general-purpose coding Agent system, AVO, achieved a score of 100.00 RHAE on the public set of the interactive reasoning benchmark ARC-AGI-3, completing all 183 levels across all 25 environments, despite receiving no descriptions, rules, or objectives for the test environments. The result was achieved using the Claude Opus 5 model and a total of 6,624 environment actions. NVIDIA clarified that the result covers only the public set included in the official scorecard and excludes the semi-private and fully private competition sets.
Note: This content was created with AI assistance and may contain hallucinations and errors.
