AI News 12 November 2026
Jensen Huang steps down as Nvidia CEO
Nvidia announced on 12 November that co-founder Jensen Huang had stepped down as chief executive and left its board. Chief financial officer Colette Kress takes over on an interim basis and keeps her finance role while the board searches for a permanent successor. Huang had led the company since its founding in 1993, turning it into a major supplier of chips and software for AI.
The company's regulatory filing says the board accepted Huang's resignation on 11 November and appointed director Harvey C. Jones as board chair. Kress has been chief financial officer since 2013. The board says it will consider both internal and external candidates for the top job.Read moreShow less
The company's regulatory filing says the board accepted Huang's resignation on 11 November and appointed director Harvey C. Jones as board chair. Kress has been chief financial officer since 2013. The board says it will consider both internal and external candidates for the top job.
The departure follows reports of a relationship with a former executive at a company that partners with Nvidia. US authorities are investigating, according to the Associated Press. Nvidia has said that it has so far seen no reason to believe national security or company secrets were compromised.
- NVIDIA · Form 8-K (Stock Titan) · Original source
- Associated Press · Jensen Huang resigns
Anthropic releases Claude Fable 5.2 and Mythos 5.2
Anthropic released Claude Fable 5.2 and Claude Mythos 5.2 on 9 November for demanding work in areas including coding, engineering and research. They share an underlying model. Fable is widely available with stricter safeguards on cybersecurity and biology tasks, while eligible organisations can apply for Mythos access.
Both cost $8 per million input tokens and $40 per million output tokens on the API. Anthropic still recommends Opus 5.5 for most coding and knowledge work. It positions Fable and Mythos for work such as large software migrations and open research problems.Read moreShow less
Both cost $8 per million input tokens and $40 per million output tokens on the API. Anthropic still recommends Opus 5.5 for most coding and knowledge work. It positions Fable and Mythos for work such as large software migrations and open research problems.
In a pre-release test, a team of agents spent about 30 hours formalising an existing proof of Fermat's Last Theorem. The agents expressed the proof in a machine-checkable form. Anthropic also reports fewer unnecessary refusals for Fable 5.2 than for 5.1 in its own tests.
- Anthropic · Launch, evaluations and access · Original source
Google opens Gemini 4 Carbon to paying users and developers
Google opened Gemini 4 Carbon to paying API customers and Google AI Ultra subscribers on 10 November. The model had been available to selected cybersecurity defenders since late October. Google positions Carbon for work that takes many steps and verification, such as debugging across a codebase.
Developers get access through the paid Gemini API. Ultra subscribers can try Carbon in places including the Gemini app and AI Mode in Search. Introductory API pricing is $2 per million input tokens and $10 per million output tokens. Google says the prices rise to $4 and $20 respectively after 30 November.Read moreShow less
Developers get access through the paid Gemini API. Ultra subscribers can try Carbon in places including the Gemini app and AI Mode in Search. Introductory API pricing is $2 per million input tokens and $10 per million output tokens. Google says the prices rise to $4 and $20 respectively after 30 November.
- Google · Gemini 4 Carbon availability · Original source
Norway's Parliamentary Ombud says agencies must be able to check AI assessments
On 11 November, Norway's Parliamentary Ombud clarified what is required when public agencies use AI to support decisions. Agencies must be able to assess whether the AI result is suitable as a basis for the decision. Having a person make the final decision is insufficient on its own if the caseworker has no way to check the assessment.
The Ombud gathered information from the Norwegian Labour and Welfare Administration, the Labour Inspection Authority, the Tax Administration and Customs. How much insight an agency needs depends in part on how the system is used and what the caseworker can check by other means. The statement also calls for procedures that ensure checks actually take place.Read moreShow less
The Ombud gathered information from the Norwegian Labour and Welfare Administration, the Labour Inspection Authority, the Tax Administration and Customs. How much insight an agency needs depends in part on how the system is used and what the caseworker can check by other means. The statement also calls for procedures that ensure checks actually take place.
- Sivilombudet · Uttalelse om KI-systemer i forvaltningen · Original source
Alibaba opens a limited-time preview of Qwen4-Max
Alibaba opened Qwen4-Max-Preview on 11 November. The model can take text, images and video as input and has a one-million-token context window. For developers, the preview is tied to Alibaba Cloud's Token Plan and ends at the close of 25 November, China time.
The model supports tool calls and structured responses, which developers can use to let an agent retrieve information and pass results to other software. Alibaba says this particular model version will be retired when the preview ends.Read moreShow less
The model supports tool calls and structured responses, which developers can use to let an agent retrieve information and pass results to other software. Alibaba says this particular model version will be retired when the preview ends.
- Alibaba Cloud · Qwen4-Max-Preview documentation · Original source
Meta releases Muse Ember for lower-cost agent work
Meta released Muse Ember on 9 November for tasks agents repeat frequently, such as reading documents, choosing tools and making small code changes. The model takes text, images, audio and video as input and is available in Muse Code and the Meta Model API.
For requests of up to 200,000 input tokens, list pricing is $0.10 per million input tokens and $0.50 per million output tokens. The context window is one million tokens, with higher rates above 200,000 input tokens. The price difference can matter when a workflow makes many model calls per task.Read moreShow less
For requests of up to 200,000 input tokens, list pricing is $0.10 per million input tokens and $0.50 per million output tokens. The context window is one million tokens, with higher rates above 200,000 input tokens. The price difference can matter when a workflow makes many model calls per task.
- Meta · Muse Ember launch and pricing · Original source
MiniMax releases weights for a 16-megapixel image model
MiniMax released the weights for H3-Image on 10 November. The model generates and edits images at around 16 megapixels, including 4096 × 4096 square images. It generates at full resolution directly and accepts up to 16 reference images.
The weights are on Hugging Face. The standard licence excludes use of the weights in the EU, the UK, South Korea and the US. It also requires separate permission for businesses with more than $1 million in annualised revenue or that offer the model as a hosted service to third parties.Read moreShow less
The weights are on Hugging Face. The standard licence excludes use of the weights in the EU, the UK, South Korea and the US. It also requires separate permission for businesses with more than $1 million in annualised revenue or that offer the model as a hosted service to third parties.
- MiniMax · Model card and weights · Original source
- MiniMax · H3-Image licence · Original source
Gemini Decision 1 scores the options you give it
Google launched Gemini Decision 1 and the Gemini Decisions API on 11 November. A developer submits an input such as a document, image or audio clip alongside a set of options. The model returns a probability for each option. This can help route a request to the right department or choose which tool an agent should use.
A single call can ask up to 1,000 independent questions about the same input. Google charges $0.10 per million input tokens and bills no output tokens. The reported response time of 150–250 milliseconds comes from Google's measurements on contexts of up to 16,000 tokens.Read moreShow less
A single call can ask up to 1,000 independent questions about the same input. Google charges $0.10 per million input tokens and bills no output tokens. The reported response time of 150–250 milliseconds comes from Google's measurements on contexts of up to 16,000 tokens.
- Google · Gemini Decisions API · Original source
Claude gets spreadsheets, whiteboards and interactive apps
Anthropic launched Claude Sheets, Claude Whiteboards and Claude Data on 10 November. Users on paid plans can ask Claude to build a spreadsheet they can edit or create a visual whiteboard they can continue working on in the same conversation.
Sheets and Whiteboards launch in beta. Claude Data builds on Dashboards for creating and sharing data views. The update also lets partners embed parts of their apps in Claude through interactive connectors, so users can make changes without switching tabs.Read moreShow less
Sheets and Whiteboards launch in beta. Claude Data builds on Dashboards for creating and sharing data views. The update also lets partners embed parts of their apps in Claude through interactive connectors, so users can make changes without switching tabs.
- Claude · Sheets, Whiteboards and Data · Original source
- Claude · Interactive connectors · Original source
DeepMind uses AI to prove security properties of Signal code
Google DeepMind introduced AlphaProof Code on 10 November. Its researchers used Gemini agents to produce machine-checked proofs of selected security properties in Signal's cryptographic code. The work went down to compiled code, the instructions that can run on a device.
Human experts reviewed the formal requirements the agents had to prove. DeepMind says the verification took four weeks. When researchers reintroduced two previously fixed bugs into isolated versions of the code, the system produced proofs that both variants were insecure.Read moreShow less
Human experts reviewed the formal requirements the agents had to prove. DeepMind says the verification took four weeks. When researchers reintroduced two previously fixed bugs into isolated versions of the code, the system produced proofs that both variants were insecure.
The proofs cover the stated properties for the parts of Signal the team analysed. That makes the method relevant to other security-critical components, where reviewing source code alone may miss errors in what actually runs.
- Google DeepMind · AlphaProof Code and Signal · Original source
AlphaGenome 2 predicts how DNA changes affect specific cells
Google DeepMind opened AlphaGenome 2 to non-commercial researchers on 11 November. The model predicts how a DNA change may affect gene activity in a particular cell type or state. It can help researchers choose which variants to investigate in the lab.
The model trained on more than 100 million single-cell profiles across 68 tissues. DeepMind says it identified nearly twice as many variants that alter gene activity as its predecessor in the team's tests. It is a research tool and has no clinical or diagnostic approval.Read moreShow less
The model trained on more than 100 million single-cell profiles across 68 tissues. DeepMind says it identified nearly twice as many variants that alter gene activity as its predecessor in the team's tests. It is a research tool and has no clinical or diagnostic approval.
- Google DeepMind · AlphaGenome 2 research and access · Original source
Gemini Robotics 2.5 rolls out to robotics partners
Google DeepMind introduced Gemini Robotics 2.5 on 10 November. The model turns images and language instructions into movements a robot can carry out. DeepMind is rolling it out to robotics partners and showed a pilot in which robots pick and sort car parts at Mercedes-Benz.
In the trial, Apptronik's Apollo 2 picked metal brackets from supply bins and placed them in the right sequence. DeepMind reports a 96% success rate for this pilot. The figure applies to the specific task and production lines in the trial.Read moreShow less
In the trial, Apptronik's Apollo 2 picked metal brackets from supply bins and placed them in the right sequence. DeepMind reports a 96% success rate for this pilot. The figure applies to the specific task and production lines in the trial.
- Google DeepMind · Gemini Robotics 2.5 and factory pilot · Original source
New York opens registration for major AI developers
New York State opened a registration portal for major developers of advanced AI models on 10 November. Companies covered by the RAISE Act must register by 1 January 2027. From then on, they will have to report critical safety incidents to the state within 72 hours, among other requirements.
Developers must also publish how they manage safety risks and issue transparency reports for the models covered by the law. The state also marked the launch of a research centre at Binghamton University that will help test advanced AI systems.Read moreShow less
Developers must also publish how they manage safety risks and issue transparency reports for the models covered by the law. The state also marked the launch of a research centre at Binghamton University that will help test advanced AI systems.
Railway estimates $740,000 in unauthorised usage on stolen keys
Cloud hosting service Railway published its report on the October intrusion on 10 November. Railway estimates that the attacker used stolen keys to run up about $740,000 in unauthorised charges. The report shows how a breach at one service can let an attacker charge accounts with other providers.
Railway says the data the attacker managed to decrypt included linked keys or GitHub access tokens for 1,852 users. The company says it reset those credentials, notified users and is working with providers on refunds.Read moreShow less
Railway says the data the attacker managed to decrypt included linked keys or GitHub access tokens for 1,852 users. The company says it reset those credentials, notified users and is working with providers on refunds.
- Railway · October security incident report · Original source
Fireworks AI confirms breach also reached customer data and model weights
On 9 November, Fireworks AI expanded its account of what attackers accessed during an October breach. The affected data included account and billing information. For some customers, attackers also accessed secrets, datasets and portions of unfinished model weights.
Fireworks links the incident to a group it says attempts to sell model usage through unauthorised gateways. The update shows that material customers use to customise their models may be affected alongside the access credentials disclosed initially.Read moreShow less
Fireworks links the incident to a group it says attempts to sell model usage through unauthorised gateways. The update shows that material customers use to customise their models may be affected alongside the access credentials disclosed initially.
- Fireworks AI · October incident update · Original source
United moves to block unauthorised AI agents from booking
On 11 November, United Airlines said it would restrict access for unauthorised third-party AI agents. The airline plans to update its booking, ticketing and distribution policies. That covers its own digital channels and indirect access via travel agencies and aggregators.
United cites privacy, security and the risk of an agent making a purchase the customer did not intend. The decision affects services that aim to complete travel bookings on a customer's behalf and need access to the airline's booking system.Read moreShow less
United cites privacy, security and the risk of an agent making a purchase the customer did not intend. The decision affects services that aim to complete travel bookings on a customer's behalf and need access to the airline's booking system.
- United · Third-party AI agent access · Original source
Runway can split a flat image into editable layers
Runway added ByteDance's Seedream 5 Pro Layers to its developer API on 10 November. The model accepts one PNG or JPEG and returns a base image and transparent PNG layers. That lets users edit parts of a finished composition separately.
On 11 November, Runway changed billing from a flat fee per split to a charge per returned image, including the base image. It costs three credits per image at 1K or 1.5K and five at 2K or automatic resolution.Read moreShow less
On 11 November, Runway changed billing from a flat fee per split to a charge per returned image, including the base image. It costs three credits per image at 1K or 1.5K and five at 2K or automatic resolution.
- Runway Dev · API changelog · Original source
Anthropic starts investor meetings ahead of its planned IPO
Anthropic began investor meetings for its proposed initial public offering on 10 November. The company and existing shareholders plan to offer 825 million shares at an expected range of $70 to $78 a share. Anthropic has applied to list on Nasdaq under the symbol ANT.
Anthropic itself is offering about 801 million of the shares. Existing shareholders are selling the rest and retain the proceeds from those sales. The price range is indicative. The final price will be set later in the process.Read moreShow less
Anthropic itself is offering about 801 million of the shares. Existing shareholders are selling the rest and retain the proceeds from those sales. The price range is indicative. The final price will be set later in the process.
- Anthropic · Proposed IPO roadshow · Original source