AI News, 7 October: Decisions API, ElevenAgents Architect, Mistral Large 4 and new creative models

01Decisions API turns a fixed question into a typed answer

OpenAI opened the Decisions API in public beta on October 6, following a limited preview described in September. A developer sends text or inline images to POST /v1/decisions and asks named questions. With gpt-6-luna, the endpoint can return a probability for a predicate, a choice from supplied options with confidence, or a score against ordered rubric levels.

Consider an intake queue: classify a customer request into a fixed department, flag a product photo for visible damage, and send uncertain results to a person. The endpoint produces decisions, not a custom JSON object or a tool action; OpenAI directs those broader jobs to Responses. The beta currently supports Luna only, and its reference restricts input to user text and inline images, with possible refusals. Files and audio are outside that input scope.

OpenAI lists $0.10 per million input tokens for Luna on this endpoint, with no cache or output token charge. Regional and long-context premiums can apply.

02ElevenAgents Architect drafts repairs to voice agents

ElevenLabs announced ElevenAgents Architect on October 6 in alpha. Through voice or chat, it can inspect an agent's configuration, investigate conversation failures, propose changes to prompts, procedures, tools, knowledge bases or guardrails, and write and run simulations. Proposed configuration changes become versioned drafts for a person to publish.

A support team could examine failed refund calls, revise the procedure, simulate the changed path, and decide whether to release it. Architect cannot publish those drafts or approve its own proposals, and it does not yet run scheduled background work.

Some alpha entry points are rolling out gradually. ElevenLabs says Architect use is free through October 2026 and its conversations do not count toward agent minutes or LLM cost during alpha; later terms need checking. The announcement names external assistant surfaces, but current authentication documentation says the dashboard is the only working surface today. Architect chats are also outside Zero Retention Mode, so sensitive customer details need care.

03Mistral Large 4 opens a multimodal API preview

Mistral announced Large 4 on October 6 as a public preview through Mistral Studio. The company describes a mixture-of-experts model built for coding, agent work, visual understanding and enterprise tasks. Its model documentation lists mistral-large-4 with a one-million-token context window, structured outputs, function calling, document question answering, batching and support for its Agents and Conversations APIs.

A team could use the preview to ask questions across a long technical document and related images, then call a tool to prepare a reviewable summary. A trial should check source grounding and tool calls on real documents.

Weights are promised for the end of October; this preview is an API route, not a released local checkpoint. Mistral lists a sale API price, checked October 7, of $0.68 input, $0.07 cached input and $2.09 output per million tokens. Its pricing table gives no sale end date. Check current access and rates before a production migration.

04Kandinsky 6 publishes weights for short video and sound

Kandinsky Lab's October 6 open-source update makes Kandinsky 6.0 Video available with published weights, after an October 4 paper. Lite 3B and Pro 29B generate five-second video with synchronized audio from text or a starting image. Separate super-resolution weights target 1080p. The project provides setup instructions, a command-line route and ComfyUI extensions.

A filmmaker could draft a five-second effects shot, then check continuity and lip sync before editing. XANTRANS has not run the model or judged its results.

The repository and model card state an MIT license. Local setup calls for an NVIDIA GPU, recent Python and CUDA tooling; consumer GPU installation can require compiling SageAttention. The repository's own timing examples show that a five-second Pro clip can take many minutes even on a high-end card, excluding load and encoding. This is a demanding workstation workflow, not instant browser generation.

05Nano Banana 2.1 reaches the Gemini API

Google's Gemini API changelog dates general availability of gemini-nano-banana-2.1 to October 6. The hosted image model supports iterative generation and editing. Google highlights improved text and infographic layout, wider panoramic formats and reference-based consistency. Its image guide describes use of up to 14 reference images and optional image-search grounding.

A designer could assemble several approved product references, ask for a poster, then compare likeness to the real product and proofread every line of copy. Google's model card itself flags small text, long paragraphs, character consistency, spatial errors and occasional timeouts as limitations. Image-search grounding also cannot use photos of real people returned from web search.

This is API access to a hosted model, not open weights. Google's model card and endpoint page show conflicting context and output token limits; implementation should follow the endpoint documentation. Check the current API price before budgeting production runs.

06Research catch-up: Prism's preview weights are visible

An October 4 Prism paper from Tencent and Fudan describes sparse attention for joint video and audio work. Preview weights are now visible in the authors' model repository, but its release log leaves the exact code and weight release date as a placeholder. Treat this as a paper catch-up with accessible preview resources, not an October 7 launch or the unreleased Prism-pro model.

The repository says 720p inference needs an 80 GB GPU with CPU offload and recommends at least four GPUs for 1080p or 2K. Researchers could evaluate a short shot, but this is a demanding local experiment.