Skip to content
ML2

Weekly roundup · 31 Oct – 6 Nov 2026

A.I. for builders: week ending 6 Nov 2026

6 short notes on what changed that week, each with its primary source and one line on why it matters if you build things.

See the newest tools and workflows →

  • Google makes Gemini Omni 1.2 Flash video generally available (opens in new tab)

    Why it matters: Google has released Gemini Omni 1.2 Flash, its video generation and editing model, and made it generally available on the Gemini API and Google Cloud. It generates and edits clips natively up to 15 seconds (up from 10 with Omni 1.1), lets you set the duration in one-second steps from 3 to 15 seconds, and adds multi-turn conversational editing so you can adjust a clip without starting over. Google says the price is 30% lower than Gemini Omni 1.1, at $0.07 per second of 720p video, which works out to about $1.05 for a full 15-second 720p clip. Prototype in Google AI Studio and move to the Gemini API for production; it is also rolling out in Flow, Google Vids, the Gemini app (for Google AI Ultra subscribers) and Google Ads, and OpenRouter lists it as google/gemini-omni-1.2-flash. Google itself says fast sports action, rapid script changes and complex physics still need work, so test it on your own shots before switching a video pipeline over. If you make product explainers, social ads or short clips at volume, the longer clips and lower price are the reasons to re-run your current prompts on it.

    Source: Google: Gemini Omni 1.2 Flash ↗@GoogleAI on X: Gemini Omni 1.2 Flash ↗OpenRouter: Gemini Omni 1.2 Flash ↗

    Read the full write-up about Google makes Gemini Omni 1.2 Flash video generally available

  • Mistral opens Managed Agents and Studio apps in public preview (opens in new tab)

    Why it matters: Mistral has opened Managed Agents in public preview: you write an agent once in Python with its Agents SDK (mistralai-agents) and run the same code in your own process, on your own servers with agents serve, or hosted by Mistral with agents deploy. The point is the infrastructure you no longer build yourself: deployed agents run in the background for hours or days, resume from the last finished step after a crash, can pause a turn for a person's approval, and get per-session sandboxes, connectors, MCP servers, skills, secrets and hooks, with every model and tool call traced in Mistral Studio without setup. Mistral stores the session, so you host no database; by default a session ends after 24 hours, or after 4 hours idle, and you can raise both limits. The same release adds a mistral apps CLI (mistral apps init, dev and deploy) that scaffolds an app from a capability registry, runs it locally and deploys it to Studio in one command. Both are in public preview, and Mistral warns that the API, behaviour and limits can change in ways that need code changes. If you have been hand-rolling a queue, a state store and an approval step around a long-running agent, this is worth a test; agents init scaffolds a working project to start from.

    Source: Mistral docs: release notes, Nov 5, 2026 ↗Mistral docs: Managed Agents overview ↗@MistralAI on X: Studio managed agents and apps ↗

    Read the full write-up about Mistral opens Managed Agents and Studio apps in public preview

  • OpenAI releases GPT-6 Cyber for vetted security teams (opens in new tab)

    Why it matters: OpenAI has released GPT-6 Cyber, a cyber-specialised variant of GPT-6 Astra trained to find, validate and patch exploitable vulnerabilities while staying inside the scope a researcher sets. Most builders can't call it yet: it is available through the API and Codex only to a select group of approved Daybreak Red customers, with access expanding to all Daybreak Red customers in the coming weeks. Daybreak is OpenAI's trusted-access cyber program, and the system card says Daybreak Red access needs identity verification, single sign-on, multi-factor authentication and API IP allowlisting; API customers on Zero Data Retention generally also need Private Safety Processing, a new option OpenAI announced the same day that runs an automated safety review on records kept in your own environment. OpenAI is also making GPT-5.6 Cyber available to all Daybreak Red customers and widening Daybreak Blue alpha access to GPT-6 Astra for more organizations and individual security researchers. OpenAI rates the model Critical in cybersecurity under its Preparedness Framework, the same as GPT-6 Astra. The figures are OpenAI's own: 48.5% on ExploitGym versus 32.1% for GPT-5.6 Cyber, staying in scope in 99.93% of an alignment test without its deployment prompt and 100% with it, and 81 previously unknown high-severity Firefox bugs found, which Mozilla has started fixing. If you run a security team or build a security product, apply for Daybreak; if you maintain open-source code, expect more A.I.-found bug reports with candidate patches to land in your queue.

    Source: OpenAI: Introducing GPT-6 Cyber ↗GPT-6 Cyber system card ↗@OpenAIDevs on X: Private Safety Processing ↗@OpenAI on X: GPT-6 Cyber ↗

    Read the full write-up about OpenAI releases GPT-6 Cyber for vetted security teams

  • Moonshot releases open weights for Kimi K3.1 Pro (opens in new tab)

    Why it matters: Moonshot has published the full weights of Kimi K3.1 Pro on Hugging Face, keeping the Nov 4 date it promised at launch on October 23. You can now download and self-host the same multimodal agentic model that has been on the Kimi API, under a permissive custom licence with conditions for very large companies. The model is a 2.8T-parameter mixture-of-experts model with 104B active parameters, a 1M-token context and text plus image input, shipped as MXFP4 weights; Moonshot recommends serving it with vLLM, SGLang or TokenSpeed, so plan for multi-GPU server hardware, not a workstation. Set thinking with the reasoning_effort field (low, high or max, which is the default), and in multi-turn or tool-calling code pass back the full assistant message, including reasoning_content and tool_calls, or the model loses its own reasoning. The Kimi K3.1 Pro License lets you use, modify, fine-tune and sell the model, but a Model as a Service business with over $20M revenue in 12 months needs a separate agreement with Moonshot, and products with over 100M monthly users or $20M monthly revenue must show "Kimi K3.1-Pro" in the interface; internal use and use through Moonshot or its certified partners are exempt. That is not an OSI open-source licence. The benchmark table on the model card (for example 65.7 on Terminal-Bench 4.0, run in the Claude Code harness) is Moonshot's own; Artificial Analysis's independent listing is linked below. If you were waiting on the weights to test Kimi K3.1 Pro in your own stack, you can start now; if you only need the API, nothing changed.

    Source: Hugging Face: moonshotai/Kimi-K3.1-Pro model card ↗@Kimi_Moonshot on X: weights release ↗Kimi K3.1 Pro License ↗Artificial Analysis: Kimi K3.1 Pro ↗

    Read the full write-up about Moonshot releases open weights for Kimi K3.1 Pro

  • OpenAI releases Images 2.6 with GPT-Image-2.6 Flare and Sunburst in the API (opens in new tab)

    Why it matters: OpenAI has released Images 2.6 in ChatGPT, Codex and the API, with two new API models: GPT-Image-2.6 Flare for fast iteration and GPT-Image-2.6 Sunburst for detailed work and consistent edits. It adds animated GIF and SVG output, markup-guided edits and a new complexity setting, and OpenAI says generation is 50% faster than before. Both models work on POST /v1/images/generations and /v1/images/edits (not the Responses, Chat Completions or Batch endpoints, per the model pages) and are priced the same: $5 input and $1.25 cached input per million text tokens, and $8 input, $2 cached input and $30 output per million image tokens. The new complexity parameter (low, medium or high, and high when you leave it out) controls how much the model plans before rendering, separately from quality (auto up to xhigh and max); OpenAI recommends low complexity for quick edits, and says Flare at low quality can return an image in as little as six seconds. Markup lets you upload a transparent PNG overlay the same size as the input image to point at exactly what to change, and it also works on GPT Image 2.5 models. Generations can now return svg or gif, which matters if you ship logos, icons or short product loops straight into a site or app. If you already call GPT Image 2 or 2.5, OpenAI's prompting guide says to test Flare first when speed matters and Sunburst when Flare falls short on detail or edit accuracy, using your existing prompts. The speed and quality claims are OpenAI's own.

    Source: OpenAI: Introducing Images 2.6 ↗OpenAI API changelog: Nov 4, 2026 ↗OpenAI image prompting guide: model parameters ↗GPT-Image-2.6 Sunburst model page and pricing ↗@OpenAIDevs on X: Image-2.6 in Codex and the API ↗

    Read the full write-up about OpenAI releases Images 2.6 with GPT-Image-2.6 Flare and Sunburst in the API

Newsletter

Get the best new A.I. picks in your inbox

The best new agent workflows and tools we’ve vetted, with a clear verdict on each. Short, occasional, useful.

Free. No spam. Unsubscribe anytime.