OpenAI releases GPT-6.2 Astra for complex tasks and Sol for everyday work
要闻
OpenAI has launched GPT-6.2 Astra for complex tasks and Sol for everyday multistep work, with API access available at launch. Standard short-context output pricing is $40 and $10 per million tokens, respectively.
OpenAI has released GPT-6.2 Astra and Sol, making them available simultaneously in ChatGPT Work, Codex, and the API. Access in ChatGPT Work and Codex covers Plus, Pro, Enterprise, and Edu users. Astra targets complex programming, scientific research, and long-horizon tasks, while Sol targets everyday multistep work. Standard short-context API rates per million tokens for Astra are $8 for input, $0.40 for cached input, and $40 for output; the corresponding rates for Sol are $2, $0.10, and $10.
According to OpenAI, Astra scored 5.2 points above its predecessor on the company's multistep work evaluation, Agents’ Last Exam, at less than one-third of the predecessor's cost per task. On Cognition's FrontierCode 2.0 programming evaluation, with both models using their highest reasoning setting, Sol scored within 0.5 points of Astra at approximately one-fifth of Astra's cost per task. For access permissions, Pro 500 can use Ultrafast with both models, while Pro 200 can initially use it with Sol in Codex; eligible enterprise and education workspaces require administrator activation. OpenAI has also tightened restrictions on high-risk cybersecurity assistance and plans to provide both models to approved defensive organizations through Daybreak.
Microsoft announced seven new models and MAI Licensed Weights at Ignite 2026. The company says MAI Code 2 delivers 98% of GPT 6.1 Sol’s performance at 40% of its cost, averaged across four benchmarks for everyday enterprise coding.
Microsoft unveiled seven new models at Ignite 2026 and introduced MAI Licensed Weights, which lets customers run models in their own Azure environments. According to the company, customers can optionally customize MAI models with proprietary data, control access themselves, and retain ownership of their data and model adaptations. Microsoft will work through Frontier Company to help deliver custom models and embed them in customers’ products and Agents. Previous adopters of Custom Models include Alaska, AT&T, Docusign, Rakuten and Woven by Toyota.
The coding model comparisons focus primarily on everyday enterprise tasks. Microsoft says that, averaged across four relevant benchmarks, MAI Code 2 delivers 98% of GPT 6.1 Sol’s performance, at 40% of GPT 6.1 Sol’s cost and 5% of Opus 5.5’s cost. Code 2 Flash matches the performance of 6 Luna and Haiku 5.5, at 65% and 20% of their respective prices. According to the company, these models have not yet been tuned for fundamental science research or very long-horizon Agent tasks. In security, Microsoft says MAI Cyber 2 Flash outperforms GPT 6 Astra on Cyber Gym end-to-end testing, with a per-task cost below 2% of its counterpart’s.
Gemini Tasks launches for free in the US, Japan, South Korea, Canada, Australia and India
要闻
Google has begun rolling out Gemini Tasks free to hundreds of millions of users in the U.S., Japan, Korea, Canada, Australia, and India. Powered by the Gemini 4 Hydrogen model, it handles appointments, bills, and shopping, with some advanced features launching first in the U.S.
Google announced that its Gemini task-handling update is rolling out free from the time of the announcement in the U.S., Japan, Korea, Canada, Australia, and India. It covers hundreds of millions of users and uses the Gemini 4 Hydrogen model, with some advanced features launching first in the U.S. Users can initiate tasks by switching from Chat to Task in the Gemini app, sending messages on Android or iOS, or forwarding emails to gemini@google.com. According to Google, Gemini can connect to email, calendars, and payment methods to handle appointments, bills, and shopping, while processing data in isolated environments.
For shopping and external service connections, Gemini already connects with more than 60 apps, and developers can apply to join the Connected Apps Platform. New Google Password Manager and Google Pay integrations support authorized account access and checkout, respectively. Google says Gemini uses the Agent Payments Protocol (AP2) for purchases and works with Mastercard and Visa to use digital tokens on supported cards, with American Express coming soon. Users can set spending caps in advance, and receipts and tracking information go to Google Wallet. Google also plans to integrate the Universal Commerce Protocol (UCP) into Gemini.
Antigravity resets usage quotas for all models and reduces Carbon's time to first token
要闻
Antigravity head Varun Mohan announced plans to reset usage quotas for every model for all users to mark the service’s first anniversary. He also said Carbon now returns its first token at 1.5 times its previous speed, with the same optimization planned for Argon.
Antigravity head Varun Mohan announced in a post that the team will reset usage quotas for all models for all users as part of the service’s first anniversary. The reset covers every user and every model, rather than being limited to Carbon or Argon. In a separate post about model responsiveness, Mohan said Carbon on Antigravity now returns its first token at 1.5 times its previous speed. According to his explanation, Argon will also receive this optimization. The two models are at different stages: Carbon’s speed improvement is already in place, while Argon’s optimization remains a planned update.
OpenCode Go raises the GLM 5.3 usage allowance to $200
要闻
OpenCode has increased the GLM 5.3 usage allowance available to all OpenCode Go subscribers from $100 to $200, while keeping the price at $10. The change is already in effect.
OpenCode announced that the second phase of Operation GLM is now live, giving all OpenCode Go subscribers $200 in GLM 5.3 usage allowance for $10. According to the company, this phase is accessible through the Go page in OpenCode Console. The previously launched first phase also cost $10 but provided $100 in usage allowance; the second phase changes the allowance while leaving the price unchanged.
Kling AI has released Kling Image 4.0 and opened API access to enterprise customers and developers. According to the company, the model supports native 8K image generation, up to 2000 characters in a single image, and up to 10 images in a series from one description.
Kling AI has released its image creation model Kling Image 4.0, simultaneously opening API access to enterprise customers and developers. According to the company, the model can generate up to 10 images following a shared set of creative requirements from a single description. It also supports generating e-commerce product detail pages and long-format images for web design, with up to 2000 characters in a single image. Regarding its architecture, the company says the model combines visual understanding and image generation within one architecture and supports native 8K image generation for high-resolution display and subsequent cropping.
SenseTime released the model weights and code for SenseNova-Loop-9B on November 17, 2026. The approximately 9B-parameter model reuses middle layers to perform 56 decoder-layer computations, accepts text and image inputs, and supports a context length of up to 512K tokens.
SenseTime released and open-sourced SenseNova-Loop-9B on November 17, 2026, making its model weights and code available after releasing the Preview on October 31, 2026. The official release retains the looped architecture and two-stream mHC, with approximately 9B parameters across 32 physical decoder layers. It executes middle layers 4–27 twice, resulting in 56 decoder-layer computations per forward pass. The model accepts text and images, generates text, and supports a context length of up to 512K tokens. According to the official description, model weights occupy approximately 4.9 GB with 4-bit weight quantization, while an 8-bit or 4-bit KV cache can further reduce memory requirements for long-context inference.
In the published official evaluations, the release outperformed the Preview on all 7 benchmarks shown. AutomationBench increased from 14.33 to 36.67, Terminal-Bench 2.1 from 40.45 to 59.55, and GAIA from 37.86 to 69.90. Against Qwen3.8-27B, the release matched its AutomationBench score of 36.67 and scored 65.31 on AIDABench versus 65.01; on NL2Repo, it scored 41.36 versus 44.60. Both the release and the Preview execute two loops by default. The official description states that a compatible inference implementation must load the loop range, number of passes, and mHC routing parameters, and preserve the same computation order.
Google AI Studio updates its Android app development features
开发生态
Google AI Studio has updated its Android development features, adding support for cross-device data synchronization, user sign-in, and Gemini calls through Firebase. Its browser-based emulator now supports 1200p visuals and real-time audio, while publishing to a personal Play Library is currently limited to the United States.
Google AI Studio has updated its Android app development features with Firebase integration support and upgrades to its browser-based Android emulator; publishing apps to a personal Play Library is currently limited to the United States. Developers can use Firebase to add cross-device data synchronization, user sign-in, and Gemini calls to their apps. The emulator now offers 1200p visuals and real-time audio output, and developers can send emulator screenshots to an agent to help modify the interface. After publishing an app to their personal Play Library, users can download and install it on their phones.
Google AI Studio Build adds support for custom domains
开发生态
Google AI Studio has introduced custom domains for Build web apps, requiring paid billing to be enabled on the project and charging separately for domain registration and Cloud DNS. The existing free ai.studio subdomain remains available.
Google AI Studio has added custom domain support for web apps in Build, with paid billing required on the project to use the feature. The publishing panel provides the interface for obtaining and connecting a domain, allowing users to purchase and connect one there. A domain previously purchased in AI Studio for the same Google Cloud project can also be reused. Domain registration and Cloud DNS are billed separately, and enabling paid billing is a condition for using the feature. The existing free ai.studio subdomain remains available following the addition of the custom domain option.
Google DeepMind has launched the Google Antigravity robotics plugin, connecting supported robotic arms and linking MuJoCo simulation practice with real-world tasks performed through conversation. An official example took about one hour to go from a whiteboard sketch to a conveyor-belt sorting system.
Google DeepMind has released the Google Antigravity robotics plugin, which works with the Gemini 4 Hydrogen model to let supported robotic arms practice in MuJoCo simulation before performing real-world tasks through conversation. In an example presented by the company, a whiteboard sketch served as the starting point for a conveyor-belt sorting system completed in about one hour. According to the company, Gemini 4 Hydrogen matches the larger Argon model on routine control tasks at a lower cost and surpasses Gemini Robotics ER 2 in capabilities such as 3D spatial reasoning.
Documentation for Claude Code’s experimental Code Mode outlines a workflow from enabling and confirming the mode to assigning a task involving multiple tool calls, followed by reading the code Claude wrote. The page also provides a path to the complete documentation index.
Claude Code’s documentation page describes how to use experimental Code Mode, without specifying a launch or effective date. The workflow starts with enabling Code Mode at startup, followed by the user confirming its use, assigning Claude a task that requires multiple tool calls, and then reading the code Claude wrote. The page also lists /docs/llms.txt as the location of the complete documentation index and explains that the file can be used to find all available pages before exploring the documentation further.
Anthropic has launched linked billing in beta on the Claude Platform, allowing third-party Claude API usage to be billed directly to a customer’s organization without sharing an API key. Customers retain their existing billing relationship, set their own spending limits, and can unlink at any time.
Anthropic has made linked billing available in beta on the Claude Platform, enabling third-party platforms to link customer organizations so that Anthropic bills end customers directly for Claude API usage. According to the company, customers authorize each platform through an Anthropic-hosted flow and can then use their existing spending commitment or pay as they go; platforms can view usage for each linked customer without storing, encrypting, or rotating customer API keys. Customers can set spending limits and unlink at any time. However, requests run in the platform’s organization, so the customer organization’s settings, including IP allowlists, encryption keys, and data retention, do not apply. Supported platforms include Asana, ClickHouse, Linear, Railway, and Warp. Third-party platforms can add support by creating an app in Console and following the documentation. Customers can link their organization from a supported platform and find platforms offering linked billing on Claude Marketplace.
Cognition has launched Devin Bots, which lets users set goals for a Bot-led Agent team to pursue until PRs are ready to merge. The company says test users saw a 4.5-fold increase in the average number of PRs merged per message.
Cognition, the developer of the AI coding Agent Devin, has launched Devin Bots, allowing users to specify metrics or goals for a Bot-led Agent team to work on. Execution can continue for days or weeks, covering planning, code review and automated testing until PRs are ready to merge. Users can check progress and adjust tasks through Slack, text messages, phone calls or group chats. According to Cognition, test users saw a 4.5-fold increase in the average number of PRs merged per message, with one individual engineer's weekly PR count rising from 50 to nearly 2,000.
Perplexity launches a unified API for over 45 models
开发生态
Perplexity has launched Orchestrator API, offering access to more than 45 models through a single interface compatible with existing OpenAI or Anthropic SDKs, with unified billing at the published per-token rate of the model that actually responds.
Perplexity has launched Orchestrator API, bringing proprietary models from OpenAI, Anthropic, Google, and xAI together with open-weight models under one interface, with more than 45 models available. The interface is compatible with existing OpenAI or Anthropic SDKs. According to the company, charges follow the published per-token rate of the model that actually responds, regardless of which provider serves it. Failover between providers for the same model is automatic: when a request encounters an error, a rate limit, or a timeout, the system retries it on another deployment of that model.
OpenAI releases an updated legal Astra, claiming leading results on multiple benchmarks
产品应用
OpenAI has released a new Astra for legal work and opened access applications to all ChatGPT Enterprise customers, with separate approval required for the API. The company says the configuration ranked first on Harvey LAB-AA with a full-score rate of 20.8%.
OpenAI released a new legal Astra, allowing all ChatGPT Enterprise customers to apply through its trusted access program starting on the release date, while API access requires a separate application and approval. The product is a model configuration based on GPT-6.2 Astra that combines a dedicated legal search index with instructions for conducting research, weighing evidence and drafting answers. On evaluations, the company reports a first-place full-score rate of 20.8% on Harvey LAB-AA, nearly 40% higher than GPT-6.2 Astra, and an average of 0.03 substantive hallucinations per task, the lowest among the compared systems. According to the company, its pass rate on the Vals AI legal research benchmark also rose from approximately 46% in September to 54.81%.
OpenAI has introduced ChatGPT for Sales, including a Sales plugin and more than 30 new or enhanced partner plugins, with connections to CRM systems such as Salesforce and HubSpot for meeting preparation and drafting follow-ups.
OpenAI has launched ChatGPT for Sales, offering a Sales plugin and more than 30 new or enhanced partner plugins. Users can connect CRM systems such as Salesforce and HubSpot to their workflows, use customer records to prepare for meetings, and draw on meeting notes to draft follow-up content. The related plugins and features are not uniformly available to all users; availability depends on the plan, region, and workspace settings. Business and Enterprise administrators can enable the relevant plugins for their teams.
Google AI Mode introduces private conversations that leave no history
产品应用
Google is rolling out temporary threads in AI Mode on desktop and mobile browsers globally, covering every language AI Mode supports. According to the company, these conversations are separate from users’ accounts, are not saved to history, and do not affect future personalization.
Google is introducing temporary threads in AI Mode on desktop and mobile browsers globally, covering all languages supported by AI Mode; the source does not specify a launch date. Users can click the temporary threads icon in the upper-right corner of AI Mode to start asking questions without changing settings, then return to regular search when finished. According to Google, these sessions are completely separate from users’ accounts: responses do not draw on past preferences, and conversations neither affect future personalization nor get saved to history. Closing a thread makes it disappear. Google lists planning a surprise gift and looking up information that users do not want to reappear in their history or recommendations as examples of how the feature can be used.
Google Vids makes its avatar feature free for personal accounts
产品应用
Google announced that personal avatars in Google Vids is available at no cost to everyone with a personal Google account. Users can create custom digital avatars to appear in videos, but the avatars are restricted to the account holder’s own likeness.
Google has announced free access to the personal avatars feature in Google Vids for all users with a personal Google account. According to the company, the feature lets users create custom digital avatars and have them appear in videos. These avatars are linked to the user’s Google Account and are restricted to the account holder’s own likeness. Google says users can try the feature by visiting vids.new on desktop.
Isomorphic Labs raises $5 billion in Series C funding
行业动态
Isomorphic Labs, which focuses on designing and developing new medicines with AI, announced a $5 billion Series C round co-led by Greenoaks and Sequoia Capital, with further investment from Alphabet and participation from all existing investors.
Isomorphic Labs has announced the completion of a $5 billion Series C funding round, without specifying its closing date. Greenoaks and Sequoia Capital co-led the round, joined by Andreessen Horowitz (a16z) and several new investors, while Alphabet made a further investment. All existing investors—Thrive Capital, MGX, GV, CapitalG, Temasek and the UK Sovereign AI Fund—also participated. The company focuses on designing and developing new medicines using AI. According to the company, it is scaling its proprietary drug design engine, IsoDDE, and expanding its pipeline of AI-designed therapeutic candidates to develop next-generation medicines for patients.
US discloses progress in export control investigation into Megaspeed
行业动态
The U.S. Departments of Commerce and Justice disclosed that Justice is now leading the criminal investigation into suspected export-control violations by Singapore-based Megaspeed. Commerce said its investigation began in 2024, and NVIDIA placed the company on its no-sale list in June 2025.
The U.S. Departments of Commerce and Justice disclosed in a joint statement that the investigation into suspected export-control violations by Singapore-based Megaspeed began in 2024, with Justice now leading the criminal investigation. According to Commerce, the department had asked NVIDIA to provide records of sales made directly or indirectly to Megaspeed. NVIDIA placed Megaspeed on its no-sale list in June 2025, after which U.S. law enforcement personnel and Singaporean law enforcement authorities conducted an on-site end-use check at Megaspeed.
NVIDIA is facing a civil investigation by the U.S. Securities and Exchange Commission into whether its disclosures misled investors, according to The Wall Street Journal, citing people familiar with the matter. The report did not establish the investigation’s full scope or outcome.
The U.S. Securities and Exchange Commission is conducting a civil investigation into NVIDIA over whether the company’s disclosures misled investors; the source text provides no start date for the investigation. The information comes from reporting by The Wall Street Journal citing people familiar with the matter, rather than a published investigative conclusion. The investigation’s full scope and outcome remain unclear, and the report does not establish that NVIDIA’s disclosures misled investors.
NVIDIA board establishes a national security and export control oversight committee
行业动态
NVIDIA’s board announced a three-director National Security and Export Control Oversight Committee that will meet at least quarterly. The standing committee will oversee national security policies, export control compliance and global trade risk management.
NVIDIA announced that its board has established the National Security and Export Control Oversight Committee as a standing committee responsible for national security policies, export control compliance and global trade risk management. The existing National Security Team and Global Trade Compliance Team will report directly to the committee and Interim CEO Colette Kress on day-to-day operations. Suzanne Nora Johnson will chair the committee, with Stephen C. Neal and John O. Dabiri serving as its other two members. It will meet at least quarterly, including meetings with management and executive sessions. According to the company, the committee will also oversee continued efforts to strengthen both teams and work with independent third-party experts in law, national security and global trade compliance. The company said establishing the committee also reflects the board’s commitment to strengthening engagement with relevant government stakeholders.
黄仁勋 says he is willing to return; 英伟达 chairman says there are no plans to reinstate him
行业动态
NVIDIA founder and former CEO Jensen Huang said he would return as CEO if the board determined that doing so served all parties’ interests. However, according to Axios, chairman Harvey Jones said the board unanimously accepted Huang’s resignation and had no intention of reinstating him.
NVIDIA founder and former CEO Jensen Huang expressed his conditional willingness to return as CEO in a post on X on November 17; the source did not specify the year. He made a return contingent on the board determining that it would serve all parties’ interests and pledged to cooperate with the board’s independent investigation. On the board’s position, Axios reported that chairman Harvey Jones said during an earnings call that the board had unanimously accepted Huang’s resignation and had no intention of reinstating him. NVIDIA had previously announced that chief financial officer Colette Kress would serve as interim CEO until the board selected a permanent successor.
OpenAI makes security tool Daybreak available to NATO and others
行业动态
OpenAI announced access to Daybreak for NATO and governments and institutions in countries including Ukraine, France, Germany and Poland for authorized cybersecurity work. Its collaboration with the NATO Communications and Information Agency also provides models to NATO’s core bodies, with a monthly limit of 300 billion input and output tokens.
OpenAI announced that it is making its cybersecurity tool Daybreak available to NATO and governments and institutions in countries including Ukraine, France, Germany and Poland for authorized security work, including software reviews and vulnerability verification. According to the company, cybersecurity models for NATO’s core bodies are being provided through collaboration between OpenAI and the NATO Communications and Information Agency. Under this arrangement, NATO can use up to 300 billion input and output tokens per month.
OpenAI strengthens protections for GPT-6.2's private reasoning
行业动态
On November 17, 2026, OpenAI updated the wording in its AI Self-Improvement Threat Modeling section to clarify the High threshold under its Preparedness Framework, positioning it as an early milestone for preparing safeguards in advance.
OpenAI revised its AI Self-Improvement Threat Modeling section on November 17, 2026, to clarify the definition of the High threshold under its Preparedness Framework. According to OpenAI, this threshold is an early milestone for proactively preparing safeguards for models that may become capable of fully automated, recursive self-improvement in the future. The previous wording focused on preparing safeguards ahead of fully automated, recursive self-improvement. The revision makes explicit that the safeguards are intended for models that may acquire this capability in the future.
Anthropic tightens vulnerability disclosure rules, shortening the deadline for high-risk flaws to 30 days
行业动态
Anthropic will replace its 90-day disclosure window with 30 days for newly reported high- and critical-severity vulnerabilities and 45 days for other issues starting January 1, 2027. Reports already in progress on that date will retain their existing deadlines.
Anthropic has revised its coordinated vulnerability disclosure (CVD) policy, shortening the disclosure window from 90 days to 30 days for high- and critical-severity vulnerabilities reported to vendors and maintainers on or after January 1, 2027, and to 45 days for other issues. Reports already in progress before that date will retain their existing deadlines. If a maintainer does not respond to an initial report, the company will aim to escalate it to an external vulnerability coordinator after 15 days rather than 30. The wait between patch availability and publication of full technical details will also fall from 45 days to 30. For high- and critical-severity vulnerabilities reported more than 90 days ago that remain unpublished despite exceeding the current deadline, the company is notifying maintainers that they have an additional 30 days. It will publish these reports on its dashboard under the updated policy starting January 1, 2027.
According to Anthropic, models’ improving vulnerability-discovery capabilities have been accompanied by more frequent rediscovery of the same flaws, so the company no longer believes vulnerabilities can remain private throughout a 90-day window. For hard-to-deploy software and issues requiring changes to a specification or RFC, it will work with vendors on remediation timelines. Vulnerabilities under active exploitation will remain subject to a seven-day disclosure window. The update does not apply to AI-discovered findings that have not received human review, including such findings from OSS Scanner. The company says it will continue to provide actionable guidance with each report and include a patch or mitigation where possible. Before sending a large volume of findings to a single project, it will also continue to agree on a sustainable reporting pace with maintainers.
微软 joins over 40 organizations to develop open standards for AI transparency and safety
行业动态
Microsoft announced on November 17, 2026, that it is advancing the AI Transparency Initiative with more than 40 organizations, seeking open standards across models and providers that would let enterprises independently verify AI and Agent behavior.
Microsoft Chairman and CEO Satya Nadella announced the launch of the AI Transparency Initiative at Microsoft Ignite on November 17, 2026, bringing together more than 40 organizations to advance open standards for AI transparency, security and governance. Participants include Accenture, Adobe, KPMG, Mastercard and Volvo Cars. According to Microsoft, the initiative covers closed and open-weight models from different providers and aims to preserve IT, security and compliance teams’ ability to govern and audit Agent behavior when development teams change or update the underlying models.
The work includes expanding Microsoft’s participation in the U.S. National Institute of Standards and Technology’s SI Agent Standards Initiative and drawing on its research to develop industry standards; working with partner organizations on accountability, independent verifiability and safety by design in AI systems; and incorporating the results into Microsoft AI, Microsoft Foundry and Microsoft Copilot. According to the company, customers will be able to use the resulting evidence with security and compliance tools from Microsoft or other providers. Microsoft says existing controls can observe much of the activity around models but cannot inspect their internal processes or independently verify models’ accounts of their own behavior.
dot to get its own email address; OpenAI employee says testing has begun
前瞻与传闻
OpenAI employee samir said on X that testing has begun for dot’s dedicated email address feature, which will let users request scheduling or order refund negotiations by copying it on emails or forwarding messages to it. No specific date has been announced for broader access.
OpenAI employee samir disclosed on X that dot is set to receive a dedicated email address and that testing has already begun, but no specific date has been announced for making it available to more users. According to his description, the feature will let users assign tasks by email directly from their inbox, without switching to another interface. Users will be able to copy dot on an email and ask it to find a time that works for all participants, or forward a receipt and ask it to negotiate a refund for the corresponding order. samir said he hopes to make the feature available to more users as soon as possible.
New GPT-6.2 models draw negative reviews; Altman and Tibo respond
其他
GPT-6.2 Astra and Sol received negative feedback from early testers and heavy users after release, including complaints about interrupted long-running tasks and inadequate explanations of completed work. OpenAI’s Sam Altman and Codex lead Tibo responded that the team would continue improving usability and the user experience.
OpenAI CEO Sam Altman and Codex lead Tibo responded to criticism in user discussions following the release of GPT-6.2 Astra and Sol, stating respectively that future versions would prioritize usability and that the team would keep making improvements. The feedback concerned the user experience of both models and Astra’s performance on long-running tasks: according to reports, Matt Shumer found both models difficult to use, while developer Theo Browne said Astra tended to stop during long tasks and could not clearly explain what it had done after finishing. In a reply to his own launch post, Altman described both models as highly capable; in a separate user discussion, Tibo said he considered the new models a clear improvement, as well as faster and cheaper. The source provided no corresponding performance or pricing figures and included a notice that its content was created with AI assistance and could contain hallucinations and errors.