Google has launched Gemini 4 Neon and begun opening access to API and subscription users. The model offers a 1 million-token context window, with input and output priced at $3 and $15 per million tokens respectively; Google says it is designed for multi-step Agent workflows.
Google launched Gemini 4 Neon and began rolling it out at launch through the Gemini API, Google AI Studio, and Gemini for business, as well as to paid Google AI Pro and Ultra subscribers in Google Antigravity and Pro and Ultra subscribers in the Gemini app. The model has a 1 million-token context window. Pricing is $3 per million input tokens, $0.15 per million cached input tokens—a 95% discount—and $15 per million output tokens. Google says the model can process spreadsheets, PDFs, charts, audio, and video within a single context and turn those inputs into structured deliverables.
Neon offers four thinking levels, with high as the default. According to figures published by Google, its cost per task at that setting is 32% lower than Carbon's; at xhigh, it scores 61 on the Artificial Analysis Intelligence Index and ranks #3 overall. For development tasks, Google reports a #1 ranking on the Artificial Analysis Coding Agent Index with 58.6, a score of 65.2% on Terminal-Bench 4.1, and 46.2% on FrontierCode 2.0 at the default high setting. According to the company, Neon is designed for long-running Agent workflows and can act as an Orchestrator directing parallel Hydrogen subagents in Google Antigravity, while supporting mid-task user corrections and context compaction in long sessions.
Google makes Genie 4 available in the updated Project Genie
要闻
Google has begun opening Genie 4 to eligible Google AI Ultra subscribers, alongside the Project Genie agent and a community gallery. According to Google, the new model supports real-time environment editing through prompts and native audio, while Pro subscribers can join a waitlist.
Google has begun rolling out Genie 4 to eligible Google AI Ultra subscribers as part of this Project Genie update; the source does not specify a release date. According to Google, users can enter prompts while exploring to change the weather, add characters, or adjust the action in a scene. Native audio changes in real time with the scene, encompassing ambient sounds and character dialogue. Genie 3 remains in the model picker, supporting longer runs and spaces grounded in 360-degree panoramas. Its updates also include connecting different worlds through real-time promptable events. Google AI Pro subscribers can join a waitlist.
The update also adds the Project Genie agent and a community gallery. According to Google, the agent coordinates Google’s multimodal models through natural-language conversation to generate a world’s initial visuals and accompanying background music. During exploration, users can also bring in Gemini models to observe the screen, offer context-aware options, or automatically trigger in-world events. Users can publish creations for others to explore and remix, or modify the world structure, soundtrack, or rules of gallery creations. Google says all generated video and audio carry SynthID watermarks, and all model inputs and outputs undergo safety screening. Camera latency, intricate physical interactions, real-time audio precision, and long-term scene permanence remain research challenges.
DeepSeek ends separate charges for model-triggered searches and crawling
开发生态
DeepSeek’s official documentation states that text searches, image searches, and webpage fetching performed by models through server-side tools incur no separate charges; only model Token usage is billed. Direct client calls to the Search and Fetch APIs remain billed per successful request that returns results.
DeepSeek states in its official documentation that model use of server-side search and fetching tools is no longer billed separately; the documentation does not specify an effective date. This rule covers the online tool, web_search in the Responses API, Web Search in Claude Code, and DeepSeek Harness when its webpage search provider is set to DeepSeek server-side search or Auto, the default. Only model Token usage is billed in these cases.
Direct client calls to the Search and Fetch APIs are billed per successful request, rather than by Token usage or the number of results. Billing requires HTTP 200 and either a nonempty results array for Search or webpage body content returned by Fetch. Failed requests, searches with no results, and timeouts are not billed, and count does not affect the per-request price. DeepSeek Harness with DeepSeek Search API selected, as well as Codex Alpha Search, falls into this category. Search and Fetch have separate account-level rate limits, independent of API Keys and model concurrency limits. Fetch also has per-target-domain QPS limits; exceeding a limit returns HTTP 429.
Antigravity Pro free trial now offers only 3.8 Flash
开发生态
Antigravity’s official documentation lists 3 Orchestrator configurations that pair different Gemini 4 planning models with Hydrogen for execution. Ultra includes all configurations, while paid Pro subscriptions and trials tied to eligible devices include 2.
Antigravity lists 3 Orchestrator configurations and their plan eligibility in its model documentation, without specifying an effective date. Google AI Ultra includes all 3, while paid Google AI Pro subscriptions and trials with eligible premium devices include Neon and Argon. Each configuration uses its namesake model—Gemini 4 Neon, Gemini 4 Carbon, or Gemini 4 Argon—for planning and reviews, with Gemini 4 Hydrogen handling implementation in every configuration. According to the company, this is not a router that first classifies a request and sends it entirely to one model; instead, two models automatically hand control back and forth within the same task, without requiring users to switch manually between planning and execution. Orchestrator runs with a fixed configuration and does not support choosing a thinking level; the CLI’s --effort and /effort also do not apply. To specify a thinking level, users can directly select a standalone model supported by their plan.
OpenCode has raised its limited-time GLM 5.3 usage allowance for Go subscribers to $500 in value, with Go priced at $10 per month. The promotion initially offered $100 in usage and was later increased to $200.
OpenCode announced that subscribers to its $10-per-month OpenCode Go plan can receive $500 worth of GLM 5.3 usage through a limited-time promotion, with the official deadline stated as “before the end of this week.” According to the company, the $500 allowance reflects an adjustment to the promotion rather than its initial offering. At launch, $10 provided $100 worth of usage; that allowance subsequently rose to $200 before reaching the current limited-time offer of $500. The promotion is available to OpenCode Go subscribers.
Perplexity has launched Context API, which reads up to 10 URLs per call and returns cleaned page content, links, and query-related snippets. Pricing is $1 per 1000 successfully retrieved URLs, with no charge for failed retrievals.
Perplexity has launched Context API for retrieving web page content, supporting up to 10 URLs in a single call. Results for each page include cleaned Markdown, HTML, links, and query-based snippets. According to the company, the API defaults to prioritizing copies crawled within the past two days in Perplexity’s web index, fetching pages live when no copy is available. Calls can also force live retrieval or use only cached content. Billing is based on the number of successfully retrieved URLs, at $1 per 1000 URLs, with no charge for URLs that fail to be retrieved.
ChatGPT adds real-time voice transcription and removes the 10-minute recording limit
产品应用
OpenAI has begun rolling out Live Dictation to all ChatGPT subscription plans, displaying transcribed text as users speak and removing the 10-minute recording limit. The update covers iOS, Android, the web, and desktop apps.
OpenAI has started rolling out Live Dictation to all ChatGPT subscription plans on iOS, Android, the web, and desktop apps, while removing the 10-minute recording limit. Usage remains confined to ChatGPT, and voice team member Atty Eleti said the feature does not currently support system-level input or background dictation. Users can see transcribed text while speaking and tap the waveform to show or hide it. According to OpenAI, the Dictation update improves speed and reliability, and displaying text in real time reduces the wait for transcription.
ChatGPT launches an accounting plugin to assist with closing the books and reconciliation
产品应用
OpenAI announced that ChatGPT for Accounting is available, using the Accounting plugin and partner plugins to handle month-end reporting, bank reconciliation, and data entry. It can connect to tools such as QuickBooks, but entries require user confirmation before posting.
OpenAI has made ChatGPT for Accounting available to support accounting close workflows through the Accounting plugin and partner plugins. Supported connections include QuickBooks, Xero, NetSuite, and BILL, with tasks covering month-end reporting, bank reconciliation, and data entry. ChatGPT can draft entries from invoices and receipts, compile documents from email and accounting apps, and check for potentially duplicate records and unusual discrepancies; user confirmation is still required before posting. OpenAI states that data from ChatGPT Business and Enterprise is not used to train its models by default.
Artificial Analysis adds filtering by reasoning effort
产品应用
Artificial Analysis has added reasoning-effort filters to every chart across its website. Its homepage highlights now let users switch between each model’s highest reasoning effort and its API default settings, while the model release comparison tool supports side-by-side reasoning-effort views and score comparisons across effort levels.
Artificial Analysis has rolled out reasoning-effort filters for every chart across its website, allowing users to filter chart data accordingly. The update also includes two comparison options: the model release comparison tool can display different models’ reasoning efforts side by side and compare scores across reasoning-effort levels, while the homepage highlights provide a toggle between each model’s highest reasoning effort and its API default settings. According to its announcement, the website is working to display more reasoning-effort levels.
Gemini app adds support for low-memory Android devices
产品应用
Google is extending the Gemini app to eligible Android (Go edition) devices. The company says the expansion will bring its AI assistant to hundreds of millions more people, with Gemini Live supporting voice conversations in more than 45 languages.
Google has begun making the Gemini app available on eligible Android (Go edition) devices, although the source does not specify a launch date. According to the company, the team redesigned the app’s architecture so that the core Gemini experience can run on hardware with less memory. Features include voice conversations or interview practice in more than 45 languages through Gemini Live, using the camera to learn from and translate notes or forms, and generating custom images with Nano Banana. Users can download the app from Google Play or launch it by pressing and holding the power button on supported phones; it will also come preloaded on upcoming devices from partners including Transsion.
Google introduces Stitch Loop for continuous product improvement
产品应用
Google Labs has introduced Stitch Loop, an experimental product Agent, and opened a global waitlist. According to Google, it connects to codebases and live metrics, offers three default optimization loops, and requires users to review PRs before any changes can be merged.
Google Labs has introduced the experimental product Agent Stitch Loop and opened global waitlist registration, with eligible Google AI Ultra users receiving priority access. The tool targets post-launch software maintenance. Google says it continuously identifies issues, proposes fixes, and creates PRs using codebases and live metrics, running 24 hours a day until users’ goals are met. Users can connect GitHub and supply product context, including logs, user behavior, and feedback, through more than 180 integrations, custom MCP servers, or local CLI uploads. Every PR requires user review before it can be merged.
Stitch Loop includes three default optimization loops for cybersecurity, latency, and accessibility, and lets users create custom loops for goals such as improving checkout conversion. According to Google, the security loop combines Gemini 4 Carbon’s cybersecurity capabilities with Google DeepMind’s CodeMender to provide evidence for detected vulnerabilities and submit PRs with proposed fixes. The tool can also serve as a service layer for coding agents such as Google Antigravity, providing persistent product context to other coding agents through the Stitch CLI and MCP. In Slack, users can use @stitch to receive daily insight digests and create new loops. Waitlist registration is available at stitch.google.com/loop/waitlist.
Tencent Hunyuan has released weights for three models in the Hy4 flash series for developers to download and deploy. According to the company, the series has 224B total parameters, activates 12B per token, and supports text and image inputs with a 1M-token context window.
Tencent Hunyuan has announced the open-source release of the Hy4 flash series, making Hy4 flash, Hy4 flash-FP8, and Hy4 flash-Base available for developers to download and deploy; the source does not specify a release date. Downloads are available on GitHub, Hugging Face, ModelScope, and GitCode. According to the company, the series supports text and image inputs, has a 1M-token context window and 224B total parameters, and activates 12B parameters per token. Tencent Hunyuan also says it is accelerating the open-source release of Hy4 models.
OpenAI shares AI chip design progress, improving AMD GPU circuit metrics by up to 14%
技术与洞察
OpenAI disclosed two results from model-assisted chip design: optimization proposals passed silicon validation on its first custom inference chip, Jalapeño; according to the company, work with AMD improved PPA by up to 14% in selected GPU circuits.
OpenAI disclosed that circuit optimizations generated by its models have passed silicon validation on its first custom inference chip, Jalapeño, and that some design changes from its collaboration with AMD have been incorporated into designs sent for fabrication. The company estimates that these proposals reduced standard-cell area in Jalapeño’s vector and matrix units by approximately 13% and 10%, respectively. For the AMD project, the company said its models achieved improvements of up to 14% in power, performance, and area (PPA) in selected circuits of a flagship GPU, compared with highly optimized baseline designs, in less than 5 months. Beyond these chip projects, OpenAI has also worked with Synopsys to train GPT-Synopsys to use electronic design automation (EDA) tools.
OpenAI details the Codex sandbox: network access disabled by default, with domain-level permissions
技术与洞察
OpenAI has detailed Codex’s sandbox permissions: network access is disabled by default, workspace source files can be edited directly, and .git and .codex remain read-only. Proxy rules can allow designated domains without requiring approval for each matching request.
OpenAI published a technical article explaining how the Codex sandbox uses operating-system mechanisms to restrict commands’ access to files and networks within users’ existing environments. According to the company, default permissions let Codex edit workspace source files without confirmation for each action, while keeping the .git and .codex directories read-only and disabling network access. When a command requests additional permissions, Codex may ask the user or route the request to automated review. For domains that require recurring access, users can enable a network proxy and configure rules so that matching requests do not require individual approval. The underlying implementation varies by operating system: macOS uses Seatbelt, Linux uses bubblewrap with seccomp, and Windows is transitioning to Microsoft’s MXC library.
Arena launches a new overall agent leaderboard alongside six specialized indices
技术与洞察
Arena studied Interaction-as-a-Judge, an approach in which computer-use agents evaluate applications. According to its report, the approach raised inference-time scaling results from 55.4% to 75.5% and delivered a gain of 50 Arena points through reinforcement learning.
Arena published research results on Interaction-as-a-Judge, a paradigm that uses computer-use agents to evaluate applications. According to Arena, improvements were observed in both evaluation and downstream generation, with inference-time scaling results increasing from 55.4% to 75.5%. The research also tested a reinforcement learning approach: Arena reports that a reward model based solely on interaction can be used for reinforcement learning, delivering a gain of 50 Arena points.
Signatories from multiple sectors question Anthropic's safety image in a joint statement
行业动态
Ahead of Anthropic’s upcoming initial public offering, the Center for AI Safety (CAIS) joined figures from multiple groups on November 18, 2026, in signing a statement challenging the consistency of the company’s safety image with its approach to AI development and concentration of power.
The Center for AI Safety (CAIS) and figures with differing positions signed and published the MisAnthropic Statement on November 18, 2026, challenging the safety image Anthropic presents to the public and policymakers. The publication came ahead of the company’s upcoming initial public offering. According to CAIS, the signatories include AI safety researchers, open-source advocates, AI ethicists, and accelerationists. CAIS Executive Director Dan Hendrycks accused Anthropic of weakening its earlier safety commitments while seeking tens of billions of dollars from the public to develop recursively self-improving AI. He also criticized treating a company’s own dominance as a public good. The full statement and complete list of signatories are published at misanthropicstatement.com.
微信小微 expands testing as daily active users reach 30 million
行业动态
Tencent disclosed in its third-quarter earnings report that it had expanded user testing of Xiaowei, WeChat’s AI assistant, which has reached 30 million daily active users. According to the company, Xiaowei uses a custom WeLM model and is gradually being integrated into multiple use cases within WeChat.
Tencent disclosed in its third-quarter earnings report that user testing of Xiaowei, WeChat’s AI assistant, had expanded and that daily active users had reached 30 million. Positioned as a personal AI assistant, Xiaowei is available within the WeChat ecosystem, with integration gradually extending to chats, Moments, Official Accounts and Channels. According to the company, Xiaowei uses a custom WeLM model optimized for cost efficiency and ease of use, was designed with privacy protection in mind from the outset, and has low usage and operating costs, while inference costs are extremely low and continue to decline. Tencent expects some inference tasks to run on users’ devices in the future.
腾讯 earnings: WorkBuddy surpasses 10 million daily active users
行业动态
Tencent disclosed in its third-quarter 2026 results that WorkBuddy had more than 10 million daily active users, collectively consuming over 100 trillion large language model tokens per day. The company said it was China’s most-used AI product by token consumption.
Tencent reported in its third-quarter 2026 results that its AI application WorkBuddy had surpassed 10 million daily active users, with users collectively consuming more than 100 trillion large language model tokens each day. In describing its position in the Chinese market, the results used large language model token consumption as the metric and identified WorkBuddy as the country’s most-used AI product. Tencent Chairman and CEO Pony Ma also described it as China’s largest AI application on that basis. The results additionally stated that enterprise adoption was increasing rapidly.
豆包 tops 8 million weekday daily active users as enterprise ARR exceeds 2 billion yuan
行业动态
Doubao Work officially reported that, 3 months after launch, its daily active users (DAU) exceeded 8 million and annual recurring revenue (ARR) from enterprise customers exceeded RMB 2 billion. The product targets productivity use cases and accepts tasks expressed in natural language.
Doubao Work officially announced that its daily active users (DAU) had exceeded 8 million and annual recurring revenue (ARR) from enterprise customers had exceeded RMB 2 billion at the 3-month mark after launch. The product targets productivity use cases, with tasks entered in natural language. According to the official description, Doubao Work can autonomously break down tasks assigned by users and call tools to execute them, covering work such as research and analysis and content creation. The 2 figures disclosed represent the product’s daily active user base and annual recurring revenue from enterprise customers, respectively.
小米 reportedly weighs spinning off its large-model business; 罗福莉 does not rule out founding a startup
前瞻与传闻
Xiaomi has yet to define a financing plan for a potential spin-off of its large-model business, while several venture capital funds and CVCs have made preliminary contact, according to media reports. MiMo head Luo Fuli remains at odds with Lei Jun and has also contacted a very small number of investors and financial advisers, with an independent venture still a possibility.
Xiaomi Group has not yet formulated a clear financing plan for a potential spin-off of its large-model business, and the related decision remains unresolved, according to media reports that did not specify a date. Internal discussions about spin-off financing are proceeding alongside external contacts. People familiar with the matter said Lei Jun and MiMo large-model head Luo Fuli still disagree, with ongoing discussions yet to yield positive progress. The core research team represented by Luo wants the large-model business to develop independently. Several prominent venture capital funds and CVCs have made preliminary contact with the business. The reports also said Luo has contacted a very small number of investors and financial advisers, and starting an independent venture remains an option.
White House reportedly plans to require US companies to disclose use of Chinese AI models on their websites
前瞻与传闻
The U.S. government is considering requiring American companies to disclose their use of Chinese AI models on their own websites, Reuters reported, seeking to reduce dependence through public pressure rather than new restrictions. Whether the proposal will secure internal support and advance to formal rulemaking remains unclear.
The U.S. Treasury Department and the White House Office of Science and Technology Policy are discussing a proposal for companies to disclose their use of Chinese AI models, Reuters reported, with its prospects for formal rulemaking still unclear. The plan would require U.S. companies to publish details of that use on their own websites, aiming to reduce reliance on Chinese AI through public pressure while avoiding new restrictions. Airbnb is one example of a company covered by the proposal and currently uses Alibaba’s Qwen model for AI customer service. Separately, the chairs of two U.S. House committees have launched investigations into American companies’ use of Chinese AI and sent inquiry letters to Airbnb, Cursor developer Anysphere and DoorDash.
OpenAI employee Tibo explains differing experiences with GPT 6.2
其他
OpenAI employee Tibo addressed differing impressions of 6.2 Sol/Astra on X, saying the models underwent many changes in the final 3 to 4 days after creators finished their videos. He also said he hoped to gradually bring ultrafast-like speeds to more users.
In a post on X, OpenAI employee Tibo attributed the differing impressions of 6.2 Sol/Astra among video creators and ordinary users to the timing gap between review production and model changes. According to his account, the videos had been completed the week before his post, while the models underwent many changes in the final 3 to 4 days, and he was certain the creators had not retested them during that period. He said both models improved on their GPT predecessors but remained imperfect, and that the team hoped to improve their taste and ease of use in subsequent iterations. He also encouraged Pro 200/500 subscribers to try 6.2 Sol ultrafast, describing a different user experience and expressing hope that more users would gradually gain access to this kind of speed.
OpenAI member responds to developer complaints about Codex's English phrasing
其他
Developer Theo criticized Codex’s English on X, prompting OpenAI team member Tibo to say the team was working on the issue. Theo said custom instructions had limited effect, but expressed strong confidence that the next version would fix the problem.
Tibo, a member of OpenAI’s Codex and ChatGPT teams, responded on X to developer Theo’s criticism of Codex’s English, saying the team was working on the issue; the source did not provide the date of the response. To address the wording, Tibo proposed adding a custom instruction asking for explanations in plain language by default, with more depth when needed. Theo subsequently explained that this was among the first approaches he had tried, but that its effect was limited. He also said that, after speaking with researchers at OpenAI, he was highly confident the issue would be fixed in the next version. That assessment came from Theo, rather than an officially announced commitment to a fix.
Negative GPT-6.2 reviews raise suspicions of coordination; Tibo says he suspects collusion
其他
After several bloggers published negative GPT-6.2 reviews at the same time, OpenAI employee Tibo suggested they had coordinated on Slack. Reviewer Matt Shumer attributed the timing to an embargo lifting and denied both collusion and receiving money from Anthropic.
OpenAI employee Tibo replied to the simultaneous publication of negative GPT-6.2 reviews by several bloggers, saying he believed they had coordinated their releases in a Slack channel. Matt Shumer, who participated in the reviews, attributed the shared publication timing to an embargo lifting. He said the negative assessments reflected problems with the model rather than collusion, and that the reviewers were expressing their individual views. Regarding specific model issues, Matt said in the comments that the model slacking off was the main problem for him; he also denied having received money from Anthropic. The source did not provide specific dates for these statements.
NVIDIA 621.24 driver reportedly writes excessive logs while idle
其他
User reports compiled by troublegrid.com describe repeated idle-time logging with NVIDIA driver 621.24, with one user measuring about 30 GB per hour. NVIDIA has not acknowledged the issue, and the reported effects of rolling back or clearing the log remain based on user feedback.
Users report sustained disk writes while systems running NVIDIA drivers 621.14 and 621.24 are idle, according to a compilation by troublegrid.com. The reports identify NVDisplay.Container.exe, run by the NVIDIA Display Container LS service, as the process involved. It reportedly truncates and rewrites the binary log nvlog.nvlgstg rather than appending to it. One measurement on 621.14 recorded 515 MB written per five minutes at idle, compared with under 2 MB per five minutes on a clean installation of 621.01 under the same conditions. Another user running 621.24 reported about 30 GB per hour, despite having disabled telemetry in the NVIDIA App.
The vendor has not confirmed the reported remedies. Two users independently reported that reverting to 621.01 stopped the sustained writes. Stopping the service, deleting the log and restarting the service reduced the process's writes by 91% in one test on 621.14, but the log grows again; another user on 621.24 said restarting the service and rebooting did not help. Disabling the service prevents NVIDIA Control Panel and NVIDIA App from opening. Previously applied settings remain in effect but cannot be changed until the service restarts. NVIDIA has not acknowledged the issue, and its 621.24 release notes dated November 12, 2026 (RN-08399-621.24_v01) list Open Issues as N/A and do not mention the disk-write problem.
Nginx releases a security update fixing 10 vulnerabilities
其他
nginx has announced security updates addressing 10 vulnerabilities, including 3 rated major. Its advisory lists 1.31.9+ as unaffected by all 10 and 1.30.7+ as unaffected by 9 of them; no publication date is provided.
nginx has released security updates fixing 10 vulnerabilities, although the supplied advisory does not specify a publication date. According to the official information, 1.31.9+ is unaffected by all 10 vulnerabilities, and 1.30.7+ is also listed as unaffected by 9 of them. The exception is a NULL pointer dereference when using PROXY protocol version 2, identified as CVE-2026-120993 and rated minor. Its affected versions are 1.31.4–1.31.8, and the advisory lists only 1.31.9+ as unaffected.
The official advisory rates 3 of the 10 vulnerabilities as major: memory corruption in HTTP/2 flow control (CVE-2026-123007), an infinite loop while reading a client request body over TLS 1.3 (CVE-2026-118899), and a use-after-free in the mail proxy module (CVE-2026-113425). The remaining issues involve use-after-free in resolver response processing, HTTP/2, and HTTP/3; a buffer overflow when using alias with try_files; an SNI bypass during client certificate validation in stream; and insufficient validation in the stream ssl_preread module.