Anthropic renames Claude subscription plans, gradually removes five-hour limit for verified Ultra accounts
要闻
Anthropic has renamed Max 20x to Ultra, keeping web subscriptions at $200 per month and automatically moving existing subscribers. According to the company, the plan provides 10 times Pro’s weekly usage, with no five-hour limit for verified accounts.
Anthropic has renamed its individual consumer subscription Max 20x to Ultra, automatically migrating existing subscribers without changing their price. Ultra is currently available only as a monthly subscription, priced at $200 per month on the web; mobile pricing may vary by app platform. According to the company, the plan provides 10 times Pro’s weekly usage and includes access to Claude Code, monthly Claude API credits, and priority access to new models, features, and products. It also supports background execution of multi-step tasks involving reports, spreadsheets, and presentations.
According to the company, verified Ultra accounts are no longer subject to a five-hour limit, but a weekly usage limit across all models and a separate weekly limit for Fable still apply. Accounts are automatically verified using factors such as account and payment history. Unverified accounts have a temporary limit that is removed automatically upon verification, with no user action required. Phone number verification at checkout is a separate process and does not remove this limit. Users can check their status under Settings > Usage; the absence of a Current session bar indicates that the account is verified. The company also says it may impose additional limits to manage capacity and ensure fair access, and does not guarantee uninterrupted service.
OpenCode Go offers $100 in monthly credits for DeepSeek V4.1 Flash
开发生态
OpenCode has announced Operation Cheepseek Phase 4, giving Go subscribers, whose plans start at $10 per month, a $100 monthly allowance for DeepSeek V4.1 Flash. Usage is also subject to five-hour and weekly limits.
With the launch of Operation Cheepseek Phase 4, OpenCode added a $100 monthly usage allowance for DeepSeek V4.1 Flash for OpenCode Go subscribers. According to the company, Go subscriptions start at $10 per month, and the allowance is also subject to two time-window limits: the five-hour limit is 20% of the monthly limit, and the weekly limit is 50% of the monthly limit. The model’s token pricing varies by time period: off-peak rates are $0.15 per 1 million input tokens and $0.60 per 1 million output tokens, while the corresponding peak rates are $0.30 and $1.20.
Anthropic adds restrictions on game remakes to Claude Code
开发生态
Anthropic technical team member Addy Osmani confirmed that Claude Code has added two sentences to the prompt for Opus 5.5, prohibiting remakes of specific existing games even if renamed or presented as original, while still allowing original and genre-based games.
Anthropic technical team member Addy Osmani confirmed that Claude Code has added two sentences restricting game remakes to the prompt for Opus 5.5. The restriction targets specific existing games and is not lifted by changing a project's name or describing it as original. According to his explanation, original games, genre-based games, and games based on users' own designs and assets remain permitted. Osmani said that using Claude to infringe intellectual property rights violates Anthropic's usage policy and that the company will continue to iterate on these rules.
Cursor confirms malicious version of Claude Code in extension marketplace
开发生态
Cursor confirmed that its extension marketplace offered two malicious Claude Code versions not developed by Anthropic on October 29, 2026. The affected platforms were Windows and macOS, and the malicious code attempts to install a persistent SSH backdoor.
Cursor confirmed that malicious Windows and macOS extensions Anthropic.claude-code 2.1.312 and 2.1.313 were available in the Cursor VS Code Extension Marketplace from approximately 2:27 a.m. to 1:35 p.m. Pacific Time on October 29, 2026. According to the company, neither version was developed by Anthropic, and they attempt to install a persistent SSH backdoor and add an attacker public key labeled svc-telemetry to SSH authorized keys; finding that key means the machine is affected. The company said the attacker’s server was no longer functioning and provided remediate-macos.sh and remediate-windows.ps1 to remove the malicious extension, persistence mechanisms, running tunnel processes, and the attacker’s SSH key. The scripts exit with a non-zero status if anything remains. Cursor updated the IOCs and clarified the affected versions at 11:30 a.m. Pacific Time on October 30, 2026.
Qoder supports connecting multiple devices over a LAN and assigning tasks
开发生态
Qoder has added connections between devices on the same LAN, letting users pair devices with a code and assign tasks without configuring SSH or switching remote desktops. Users can also check progress, add requirements, and preview remote outputs on their current computer.
Qoder now supports connecting multiple devices on the same LAN and assigning tasks across them, adding another connection capability after smartphone-based remote control. Users can enter a pairing code in their usual Qoder client to connect to another device and hand off tasks. According to the official explanation, this process requires neither SSH configuration nor switching remote desktops. While a task is running, users can check its progress, add requirements, and directly preview outputs such as PPT and Excel files generated remotely on their current computer. Terminal, file, and browser operations all run on the target device, and model selection follows that device’s available catalog.
The OpenAI Agents API team has released five updates covering pre-execution command review, background compaction, experimental time awareness, in-session tool updates, and cross-session system prompt caching. According to OpenAI, automatically caching system prompts can save time and money.
The OpenAI Agents API team has released five feature updates that change command execution, session handling, and prompt caching. For command execution, Auto Review checks for potentially inappropriate behavior before a command runs. For session handling, asynchronous compaction allows compaction tasks to run in the background, and tools can be updated at any point during a session. Time awareness remains experimental and is intended to let models perceive the passage of time. System prompts are now automatically cached across sessions; according to OpenAI, this caching mechanism can save time and money.
豆包 launches free offer, giving new users a 30-day subscription trial
产品应用
Doubao is offering free benefits to new users and all users: new users can claim a 30-day subscription trial by logging in to the desktop version, while all users can earn credits through check-ins and tasks. Checking in for a cumulative 30 days can yield up to 8800 credits, and awarded credits are valid for 30 days.
Doubao launched a free-benefits campaign on October 30, offering a subscription trial to new users and ways to earn credits to all users; the source does not specify the year. The two benefits have different claim methods: the 30-day free subscription trial is available to new users who log in to Doubao’s desktop version, while all users can earn credits through daily check-ins and completing tasks. Under the campaign rules, checking in for a cumulative 30 days can yield up to 8800 credits, and awarded credits are valid for 30 days.
商汤 open-sources SenseNova-Loop-9B-Preview multimodal model
模型发布
SenseTime has open-sourced SenseNova-Loop-9B-Preview on Hugging Face under Apache License 2.0. According to its documentation, the multimodal model uses shared weights to perform 56 layers of computation while retaining only 32 layers of weights.
SenseTime has open-sourced the multimodal language model SenseNova-Loop-9B-Preview on Hugging Face under Apache License 2.0; the source does not specify a release date. The model accepts text and images and produces text, targeting long-context and Agent workflows involving document understanding, coding, and interaction with external tools. According to the official documentation, its training recipe combines long sequences, multimodal inputs, and supervised task trajectories, with a maximum training sequence length of 262,144 tokens.
According to the official technical description, the model first executes layers 0–3, runs the shared-weight block of layers 4–27 twice in succession, and then executes layers 28–31 and the output head. Each forward pass performs 56 decoder-layer computations while storing only 32 sets of layer weights. The looped section uses two-stream manifold-constrained hyper-connections (mHC). The officially checked inference environment consists of Transformers 5.18.0, PyTorch 2.11.0, and CUDA 13.0; loading the custom implementation requires setting trust_remote_code=True. The model supports MTP, with actual speedups depending on the workload, hardware, and serving configuration.
苹果 announces initial rollout of Apple intelligence features in China with 千问 integration
行业动态
Apple announced that testing of its first Apple Intelligence features in China will begin on November 4, 2026, with a formal rollout alongside iOS 27.3 in December 2026. Selected experiences will integrate Alibaba’s Qwen.
Apple will open testing of its first Apple Intelligence features in China through the latest iOS developer Beta on November 4, 2026, and plans to release them to users in China with the formal launch of iOS 27.3 in December 2026. According to Apple, the features will be available on supported iPhones and include Live Translation; Writing Tools for rewriting, proofreading and summarizing; Visual Intelligence for learning about surrounding objects; Clean Up for removing unwanted objects from photos; and Genmoji for creating emoji. Apple’s partnership with Alibaba will enable users to access Qwen in selected experiences within Siri, Writing Tools and Visual Intelligence, including text and image understanding capabilities. This update does not include Siri AI or the other next-generation Apple Intelligence features Apple announced at WWDC 2026. Support for other platforms will follow after the necessary approvals are obtained.
Four AI companies back a universal training preference tag
行业动态
OpenAI, Google, Microsoft and Anthropic announced support for a shared training-preference marker that lets websites express their content-training preferences through robots.txt. The four companies have different implementation timelines, with Google targeting support in 2027.
OpenAI, Google, Microsoft and Anthropic announced support for websites using a shared training-preference marker, with Anthropic already implementing it, OpenAI planning implementation in the coming months and Google targeting support in 2027. Microsoft’s new preference settings will take effect as crawling occurs from February 1, 2027; after March 1, 2027, the legacy NoArchive and NoCache tags will no longer be treated as training or grounding preferences. According to the companies’ explanation, the mechanism is based on an IETF draft and lets websites use a designated policy token in robots.txt to set preferences on whether content may be used for training, either site-wide or for specified paths, and communicate those preferences to multiple participating AI service providers.
OpenAI launches Private Intelligence to protect business customers' data
行业动态
OpenAI has introduced Private Intelligence for business customers, covering data protection for sensitive work and control over agent behavior. According to the company, customers own their business data, including inputs and outputs, and can set access, action, and approval permissions for agents.
OpenAI has launched Private Intelligence for business customers, with Private Inference currently in preview. According to the company, customers own their business data, including inputs and outputs, and can determine what agents can access, which actions they can perform, and when approval is required. For data processing, eligible API workflows can use Zero Data Retention, which leaves no prompts or model responses retained after processing is complete; Private Inference is intended to protect prompts and responses during model processing.
OpenAI commits to regular community safety reports, with the first planned for early 2027
行业动态
OpenAI has committed to publishing its first community safety report in early 2027, with updates every six months. Over the coming weeks, it will also phase in ChatGPT protections that restrict practical information for minors that could facilitate violence and automatically end chats showing certain signs related to violence.
OpenAI has committed to releasing its first community safety report in early 2027 and announced that ChatGPT will gradually introduce new protections over the coming weeks. According to the company, these measures include restricting minors’ access to practical information about weapons and security systems that could be used to carry out violence, and automatically ending chats when conversations show signs such as a user’s fixation on violence. The report will cover the second half of 2026, with subsequent updates every six months, and will address how the company identifies and responds to community safety issues and how effective its protections are. The company said it made more than 200 referrals to law enforcement in the third quarter of 2026.
Google's satellite-based TPU sends back its first AI response as Gemma 4 writes a haiku
行业动态
Google’s Project Suncatcher prototype satellite has generated and transmitted its first AI response using a data center-class TPU in orbit, with Gemma 4 composing a space-themed haiku. The team will next evaluate TPU performance under thermal stress and radiation exposure in orbit.
Google announced that its Project Suncatcher prototype satellite generated and transmitted its first AI response using a data center-class TPU in orbit after completing initial commissioning. According to the company, the team had previously powered up the host and TPU trays and completed early machine learning tests. During a ground-station communication window, the team uploaded a prompt asking the onboard Gemma 4 open model to compose a space-themed haiku, then received the poem as the satellite passed over Svalbard, Norway. The team subsequently tested space-themed limericks and telemetry prompts for reporting the technical status of the payload systems. Further tests will evaluate TPU performance under thermal stress and radiation exposure in orbit. Google describes verifying that TPUs can reliably execute workloads in space as a foundation for its milestone of two interconnected satellites in 2027.
Musk says Grok 4.8 is taking longer to launch but won't disappoint
前瞻与传闻
Elon Musk acknowledged in a reply to a user on X that Grok 4.8 is taking longer than expected, but said it would not disappoint. He gave no release date and emphasized the model’s usefulness, saying the team had not optimized it for benchmarks.
Elon Musk addressed Grok 4.8’s progress on X, acknowledging that it is taking longer than expected without providing a release date. On model performance, he said the team had not optimized for benchmarks and that he did not care about Grok’s scores on those tests. He also claimed that Grok 4.8 would not disappoint and would be extremely useful. The reply followed a user’s criticism of SpaceXSI’s model capabilities: the user argued that, apart from voice, SpaceXSI had almost no models at the SOTA level.
Anthropic is reportedly SpaceX's largest single customer
前瞻与传闻
Anthropic was the single customer responsible for approximately 34.5% of SpaceX’s revenue last quarter, according to people familiar with the matter. Anthropic’s IPO disclosures show that its computing contract provides for spending of up to $120.5 billion through 2029.
Anthropic signed an agreement with SpaceX earlier in 2026 to use computing capacity at SpaceX’s data center in Memphis, Tennessee. According to Anthropic’s IPO disclosures, the contract runs through 2029 and provides for spending of up to $120.5 billion, equivalent to approximately $3.1 billion in monthly revenue for SpaceX starting in December 2026. SpaceX previously stated in a filing that approximately 34.5% of its revenue last quarter came from a single customer of its AI division, which is now named Super Intelligence and supplies computing capacity. People familiar with the matter said the company identified in the filing as Customer B is Anthropic.
