AI News 5 November 2026

The API offers two variants. Flare is designed for fast drafts and can generate an image in as little as six seconds at low quality, according to OpenAI. Sunburst is aimed at more detailed images and better consistency across edits to characters, products and other visual elements.

Developers can now mark the area to edit directly on the image, a feature already available in ChatGPT. They can also export SVG for use on websites, in apps or in print. That makes it easier to revise one part of an illustration or take it into another tool.

The API has separate settings for planning effort and image quality. A simple colour change can use low planning effort even if the final image needs to be high quality. A poster with several elements and strict placement constraints may need more planning.

K3.1 Pro takes text and images as input and supports just over a million tokens of context. The model card lists 2.8 trillion parameters in total, with 104 billion active for each token. The model selects only parts of the network at each step. That reduces computation compared with activating the whole network, though the full set of weights still has to be available to the serving system.

For self-hosting, Moonshot points to vLLM, SGLang and TokenSpeed. The file size conveys the scale of the release, but it does not tell you how much GPU memory a particular setup needs. Moonshot also offers the model through its own API, where users can choose how much reasoning the model does before answering.

K3.1 Pro builds on Kimi K3. Moonshot changed how the model retrieves information from long inputs to reduce the work needed to find relevant passages. That matters when an agent needs to keep source code, documentation and earlier tool results in view during a single task.

Moonshot’s tests show why a single ranking misses part of the picture. Kimi scored 65.7 on Terminal-Bench 4.0, close to Claude Opus 5.5 at 66.4 in the same table. On the science version, Kimi scored 38.1 and GPT-6 Astra scored 68.1. The model card records the tools and settings used, which matters when judging the model for a particular task.

Downloadable weights give teams control over where the model runs and how it is configured. The licence has specific terms for large model-as-a-service providers and separate attribution requirements for very large products. Those terms are worth checking before commercial use.

Tencent says the model can analyse complex charts and use tools across multi-step tasks. The company says it trained Hy4 specifically on Blender and CAD software, and highlights development in Unreal Engine, Unity and Godot. That shows where Tencent focused its training. How well the model performs a given task also depends on the tools and environment available to it.

In WorkBuddy, Hy4 can use information from WeCom, including emails, meetings and documents, to keep track of tasks and deadlines. Tencent also tested the model on 579 finance tasks with input from 213 professionals. Hy4 had the highest overall score among the models compared. Tencent led the evaluation, and the result is most informative for the tasks it covered.

WorkBuddy, CodeBuddy and Tencent Cloud TokenHub were among the first ways to access Hy4. Tencent also announced a lower-cost variant called Hy4 Flash. At launch, teams considering self-hosting still had to wait to see which weights, deployment requirements and licence terms Tencent would publish.

Conductor stopped new sign-ups and app downloads on 4 November. It says existing Pro users can keep using Pro features for free and active subscribers will receive refunds. During the read-only period, users can still read old chats and export data. Cloud data deletion begins when the service shuts down.

On a Mac, users can right-click a workspace and move it to the local app, along with its chats. That is the option for continuing work there. To keep a full archive, users can instead export their data from account settings.

The local app will receive new models and critical fixes until early 2027, when it too starts winding down. The team says it is working on something new, but the announcement gives no details about what it will be.

Preview is a general-purpose variant of the model the company uses internally for chip design. Ricursive released both a base model and an instruction-tuned version. There is also a compressed NVFP4 release. The company points to NVIDIA DGX Spark as one example of hardware that can run it locally.

Ricursive reports a score of 80.52 on Terminal-Bench 2.1, which tests agents on terminal tasks. In the company’s comparison, that puts it ahead of some substantially larger models. The model card says different agent tools and context settings changed the scores, and its table mixes publicly reported results with in-house measurements. The score is a result for a particular setup.

Ripple uses all 65 billion parameters for each token. Models such as Kimi K3.1 Pro activate only part of the network at a time. Ripple’s model card also places it behind several of the models it compares against on SWE-bench Verified, which tests fixes to existing code. The terminal score alone tells you little about which model is the best fit for that work.

Ricursive puts the total development cost at $32 million, including experiments, evaluations, data generation and the final training run. At launch, users could download the model from Hugging Face or access it through hosted services including Together AI and OpenRouter.

The agents started with 2,308 candidate findings and checked which flaws they could reproduce in isolated machines. Researchers used the demonstrations to assess severity and prioritise reporting. The run also produced 291 proofs of concept for information leaks. OpenAI puts the cost of the main run at about $1.2 million in API credits.

The researchers first had agents scan parts of the kernel reachable by a user with limited privileges. Once the agents reproduced a flaw, they checked what it could be used for. Multiple demonstrations could stem from the same bug. The researchers grouped 1,169 working privilege-escalation demonstrations into 948 unique findings based on the fixes they required.

An early run appeared to show 200 cases where an agent gained administrator privileges. On review, researchers found that 20 demonstrations had used a different vulnerability from the one assigned. The team added an independent check of which flaw each demonstration actually exploited.

By the end of the October run and review, 61 of the local privilege-escalation flaws had already been fixed upstream and 43 had pending patches from other contributors. OpenAI says it sent hundreds of reports and candidate fixes. The figures show why findings need to be checked against the latest code before they are reported.

These shells, called vault particles, were already known in organisms including mammals. Researchers had also identified related proteins in bacteria and their viruses. The question was whether those proteins could assemble into complete shells. The team used Claude to search a database of roughly 15 billion predicted proteins before researchers selected candidates to test. Microscopy confirmed the shape for four of them.

The shell consists of 96 copies of the same protein. The team used cryo-electron microscopy, which images frozen samples from many angles to reconstruct a 3D model. Anthropic says particle purification took about 90 minutes instead of the usual three days. A first full structural model was ready after 34 hours of processing and the final maps after about a week.

The team also modified the protein to see what makes the shell close. Some changes at the ends produced open particles instead of sealed shells. Adding two extra repeats to the wall increased the enclosed volume by about 20%. These experiments show how the shell’s shape and size can change.

Researchers directed the work and corrected mistakes. One protocol from Claude delayed the project by two weeks. The model also skipped a standard analysis step and blended the structure with its mirror image. The team had produced the shells in E. coli and had yet to test whether they form where the proteins occur naturally. At the time of the announcement, the work had been submitted as a preprint ahead of peer review.

Kino uses Remotion to represent edits and motion graphics in code. That gives a coding agent such as Codex or Claude Code access to the edit while a person can make visual changes on the timeline. The team says it used this setup to let several agents work on different parts of its launch video at the same time.

Kino transcribes and describes footage so editors can search for what was said as well as what was on screen. A colleague can comment on a particular moment from a browser. The note appears on the timeline, where the editor can give the agent a link and ask it to make the change in the right place.

Kino says it plays footage in formats such as ProRes, BRAW and R3D locally, while uploading a lower-resolution copy for collaboration and cloud processing. The app is available for macOS, Windows and Ubuntu. Editors can work with original footage locally while colleagues and agents work with a smaller copy.