Chimera brings a voiced character into the video frame

ElevenLabs introduced Chimera on 21 October as a template inside Eleven Creative that pairs Eleven v4 voice generation with Seedance 2.5 video generation. The practical change is a single character workflow: supply an image, add a spoken performance, and generate a short clip whose visible character is built around that voice. For a filmmaker, the obvious use is a brief character line or a social-format talking shot that starts with a chosen face and an intentional read rather than a silent video awaiting a separate voice pass.

The current Chimera screen asks for a character image and 4 to 30 seconds of speech. A motion prompt is optional; its example describes the character laughing and leaning toward camera during a slow push-in. The screen offers 9:16, 1:1 and 16:9 frames, plus 480p, 720p and 1080p choices. That gives a creator room to frame a vertical character beat or a square post before generating. The interface says it will estimate cost after speech is added, so there is no verified fixed price to quote from an empty form. The announcement invites users to try the template in Eleven Creative; plan entitlements, API access, render time and output consistency have not been established by these sources.

Claude Duet starts a task before you reopen the chat

Anthropic opened a Claude Duet research preview on 21 October. Duet is a separate app, available on iOS, Android and claude.ai, that can continue assigned work in its cloud computer and surface items it thinks need attention. Its morning brief can look across connected email and calendar, then prepare first steps such as a reply draft. Existing Claude memory, skills and connected apps carry over, so a user does not have to teach a second assistant the same preferences from scratch. Anthropic says Duet asks before actions such as sending a message or making a purchase; that approval design is part of the product, not evidence that every suggestion or action will be correct.

For a small studio, Duet could flag a time-sensitive client thread and draft a response for review while the owner is away. It would be a poor choice to assume it saw every file or every urgent message: Anthropic says account checks happen at intervals, Duet cannot read all local files or projects, and users cannot yet watch or take over its browser. It can misjudge urgency. Access starts through a waitlist for US Pro and Max subscribers only, with invitations in waves; Team and Enterprise are excluded for now. Duet shares the user's Claude plan limits, and background checks consume that allowance. A Latvian reader should treat this as a product preview to follow, not a service they can necessarily activate today.

OpenAI models are headed into Google Cloud's Model Garden

OpenAI and Google Cloud announced on 21 October that OpenAI frontier models and Codex agent workflows are coming to Google Cloud's Model Garden. The appeal for an enterprise developer is access to those tools in the cloud environment where its data permissions, governance and infrastructure already live. A team building an internal coding agent could evaluate the model alongside its existing Google Cloud services rather than moving the whole workflow to another platform. OpenAI also says its models will come to Google Antigravity and the Gemini Agent, giving those tools more model choice.

The verb is 'coming.' OpenAI says general availability is planned in the coming weeks, while the current preview is invite-only for selected development and internal evaluation. The access form explicitly does not describe production use for that preview. Model Garden availability, Antigravity access, pricing through Cloud Marketplace and any production integration should be checked when the rollout actually reaches the intended organization. This is an infrastructure and distribution announcement, not a claim that an OpenAI model is already selectable in every Google product.

A 20 October cache option can keep an agent's shared context warm longer

One useful catch-up from 20 October: OpenAI added an optional 12-hour minimum prompt-cache lifetime to the Responses API for supported newer models. A developer can keep stable instructions, tool definitions or reference material eligible for cached-input reads across longer gaps, while leaving changing conversation content on the default 30-minute lifetime. That can help an agent resume work later in the day without paying the full input cost for the same long prefix each time.

The longer lifetime is a pricing choice, not free permanent memory. OpenAI lists 12-hour cache writes at 2.5 times the ordinary input-token rate, versus 1.25 times for the 30-minute default; reads have their own lower rate. It pays off only when enough later requests reuse the same eligible prefix. Entries are not stored indefinitely, and the chosen 12 hours is a minimum lifetime rather than an expiry promise. Check the supported model and measure actual cache hits and total charges before switching a production agent.

Sources & verification

Prepared with AI assistance from the linked primary sources. Vendor demonstrations and proposed checks are not independent XANTRANS tests.