AI News 6 November 2026

Google lists 720p video at $0.07 per second, 30% below 1.1 Flash. A 15-second clip comes to $1.05 at that rate. Google describes a workflow where you draft in 360p or 720p and upscale selected clips to 1080p or 4K.

You can edit the same video through multiple text prompts, for example changing the background and then adjusting the lighting in the same clip. Google also says the model follows image, video and audio references more closely. Google Cloud says audio references require separate approval.

Initial access is through Codex and the API for approved Daybreak Red customers. OpenAI says access will expand gradually. The programme is intended for authorised security testing, and API requests generally have to originate from approved networks.

On OpenAI’s ExploitGym evaluation, the score rose from 32.1% to 48.5% compared with GPT‑5.6 Cyber. OpenAI also says its researchers used the model to identify 81 previously unknown high-severity vulnerabilities in Firefox, and that Mozilla has released fixes for many of them. The evaluation score reflects a controlled setting.

METR had access to internal model evaluations, data on computing resources and staff interviews. Anthropic’s internal benchmark index rose at an estimated 1.9 times its earlier rate after October 2025. Alternative analysis choices produced estimates from 1.1 to 2.9 times.

METR looked at several possible explanations, including better training data, hiring and more computing power. It found that the computing power used for training did not start growing faster during the period. METR judges that AI likely explains much of the current acceleration, but found no decisive evidence for its exact contribution. In interviews, staff described AI primarily as supporting researchers in their own work.

The benchmarks measure tasks with clear criteria for a good answer. METR is less certain whether the models improved as much on tasks with less clearly defined goals. Anthropic also redacted parts of the evidence before publication, limiting what others can independently check.

The weights come from the end of pretraining. Developers can use them as a starting point for adapting the model to their own tasks. Baidu lists support for up to one million tokens during fine-tuning and inference, compared with up to 128,000 tokens during pretraining.

Baidu’s licence defines a territory that excludes the EU, the United Kingdom and South Korea. It also sets specific terms for commercial use. Teams assessing the weights for a project should read both the technical report and the licence.

The model takes text and images and returns text. Around 17 billion of its 594 billion parameters are active per token. The company shows a browser-based city simulator the model built with Three.js, illustrating how a coding agent can use the model to assess and refine what it has built.

Developers can try the model in Tinker Playground or fine-tune it on the hosted service. The weights remained closed at launch. The team says the weights and downloads of custom fine-tuned models will come with the planned Apache 2.0 release.

Mistral’s Python SDK lets developers define the model, instructions and tools once, then run the same code locally, on their own infrastructure or on Mistral’s managed service. Apps give colleagues an interface to the agent. Workflows suit processes where the team wants to define the steps in advance.

Studio can also show the model and tool calls an agent made during a run. Developers can use those traces to build evaluations. Individual features have their own availability and some are still in preview.

When many agents push to the same repository, each push can also trigger reads by other agents, builds and tests. Rubicon separates Git object storage from updates to references, such as the commit a branch points to. That lets GitHub scale reads and writes more independently.

GitHub reports 7,200 pushes per second to one repository in its internal test while serving read traffic. That figure describes the test setup. The company says a broader rollout is planned for early 2027.

Agility says the merger and related financing raised about $465 million before costs, with some of the proceeds going towards expanding production. It says the current Digit has moved more than 100,000 totes at GXO’s site in Georgia. The first customers are expected to get access to the next generation, Digit 5, in the first half of 2027.