AI news, November 9, 2026: OpenAI teaser, Codex Pro plans, secrets in Claude Code files
Short answer: The main stories of November 7 to 9: OpenAI teases an announcement for Monday without content, according to Codex team staff all Pro plans no longer have a 5-hour limit, and a summary of Apiiro research reports secrets in 1,122 of 104,185 Claude Code settings files. Claims about Nvidia, gas generators for Anthropic and autonomous research are unconfirmed.
Illustration by Claude (AI), drawn with code: a door ajar (teaser), a gauge with settings (plans) and a folder with a key (secrets).
OpenAI teases an announcement for Monday; Karpathy builds TristanTube with Claude CodeConfidence: high that the posts exist; the content of the OpenAI announcement is unknown
OpenAIDevs posted only the word 'Tomorrow.' with an image, and the post received more than 11,000 likes. The announcement is scheduled for Monday, November 9; its content is not known.
Andrej Karpathy reports that he turned the top YouTube channels of Tristan into 'TristanTube', with 47,000 videos and a TV mode, and that it took three minutes with Claude Code. The post received over 8,700 likes.
What this means: the OpenAI announcement is a teaser without content, so guesses about the topic cannot be tested. Karpathy's example shows how quickly small web apps can be made with a coding agent. Background is in our overview of agentic AI.
Sources: @OpenAIDevs op X (teaser) · @karpathy op X (TristanTube)
Codex Pro plans without a 5-hour limit; Musk announces Grok Bot fixes; Copilot sessions open in VS CodeConfidence: medium (own posts by staff and companies on X; no official plans page read)
Tibo of the OpenAI Codex team reports that all Pro plans no longer have 5-hour limits. Pro 100 gives 10 times, Pro 200 gives 20 times and Pro 500 gives 50 times the usage of Plus. Another post says usage on Pro plans and Business 100 has doubled, and he asks whether a 'Pro 50' plan would make sense. The latter is a poll and not an announcement.
Elon Musk reports that fixes for Grok Bot target token efficiency and smarter behavior. After a round table with the engineering team, fixes were to be rolled out overnight, and users should notice a 'dramatic improvement' on Monday, according to him. One user describes running heavy tasks outside Grok Bot because usage runs out quickly.
GitHub reports that a session in the Copilot app opens directly in VS Code, with task, plan, conversation and code changes, so there is no need to start over.
What this means: providers compete on usage allowance per plan. The claims about doubling and removed limits come from staff on X; the official plans page is the source to rely on. Background on the tooling around them is in our overview of MCP servers.
Sources: @thsottiaux op X (Pro-plannen) · @thsottiaux op X (gebruik verdubbeld) · @elonmusk op X (fixes) · @github op X (Copilot-app)
Research on Claude Code settings files: 1,122 of 104,185 contain secretsConfidence: low to medium (second-hand summary; primary research not read)
A summary of Apiiro research (November 5) reports that of 104,185 public Claude Code settings files, 1,122 contain secrets such as API keys. Approved commands can store values in those files. The summary comes from an account with little reach; we have not read the original research.
What this means: settings files in a repo should not be public if they contain commands with keys. Check your own repos for such files and rotate keys that have leaked. Background is in our overview of AI and cybersecurity.
Qwen4-27B: release time unconfirmed; Kakao publishes Kanana-2-v-decideConfidence: low for Qwen4-27B (no confirmation); medium for Kakao (release details via a third-party account)
Two accounts give different release times for Qwen4-27B, an open-weight model announced on September 22 at Apsara. Alibaba has not confirmed a date.
Kakao published Kanana-2-v-decide on Hugging Face on November 6 under the Kanana Open License. The model has 3.67 billion parameters, is meant for choice, yes/no and score questions, and handles text, JSON or chat with up to 10 images.
What this means: anyone waiting for Qwen4-27B is better off following Alibaba's official channels than loose timestamps on X.
Sources: @MartinSzerment op X (Qwen4-27B) · @getsuperintel op X (Kanana)
Unconfirmed: probe around the Nvidia CEO, gas generators for Anthropic and autonomous research at the labsConfidence: low (no primary source; claims from accounts on X)
Andrew Curran reports that the Nvidia board has reportedly started an investigation into claims that Jensen Huang was the target of a foreign intelligence operation. Dylan Patel expects the matter to have little effect on media and market, but that Democrats want to call Huang before Congress over GPU sales to China. A primary source is missing.
A daily data center overview reports that Anthropic bought 1.7 GW of gas generators for its own power. The post received 12 likes and names no primary source. Background on power and data centers is in our article on data centers and AI in the Netherlands.
Curran further states that the three major labs have reached autonomous research and that results follow this month, and that an OpenAI chart at the Navier-Stokes announcement showed no diminishing returns from test-time compute. These are statements by a single account without attached evidence.
What this means: treat these points as signals that still need confirmation.
Sources: @AndrewCurran_ op X (Nvidia) · @dylan522p op X (Jensen-zaak) · @SingularityRes op X (gasgeneratoren) · @AndrewCurran_ op X (autonoom onderzoek)
Primary sources: OpenAIDevs op X · Tibo (OpenAI Codex) op X · GitHub op X
