Download on theApp Store
Menu

AI audio, creative workflows & developer tools · Product news & updates

AI news: Eleven Realtime, OpenAI caching and OneDrive Copilot — 21 October 2026

ElevenLabs launches Realtime voice models; OpenAI adds 12-hour caching. OneDrive Copilot, Google previews, Claude Code fixes and Perplexity updates explained.

By · Published · Updated

Publisher disclosure: MUVO offers competing music, remix, and video tools. This report summarizes public sources and provider documentation; it is not an independent product test or a claim that MUVO is free of similar issues. Read our reporting method.

What changed in this edition

ElevenLabs introduced Realtime voice models, OpenAI added a longer cache option, and Microsoft announced broader Copilot workflows in OneDrive. Other consequential changes include Google government and private-infrastructure previews, Claude Code and Copilot CLI fixes, and Perplexity’s model and access-control updates.

The research window is 18 October at 11:52 through 21 October at 12:01 Europe/Istanbul (UTC+3). Dates below distinguish announcements from availability. Vendor performance and fix claims have not been independently tested by MUVO.

ElevenLabs launches Eleven Realtime for voice agents

Eleven Realtime v1 Mini and v1 Max combine the voice-agent pipeline into speech-to-speech models. The 20 October announcement positions Mini for conversational assistants, tutoring and characters, and Max for complex workflows and tool use. ElevenLabs says both support more than 90 languages and its voice library. Original launch.

The standalone Realtime API follows OpenAI’s Realtime specifications over WebSocket. Existing WebRTC integrations need a transport change. Channels, knowledge bases and other platform features belong to ElevenAgents rather than automatically accompanying the standalone API.

Workspace eligibility and regional limits

ElevenAgents access is rolling out gradually by workspace. The documentation says an absent voice-mode selector means the workspace is not enabled; API requests do not bypass that eligibility. Mini supports US, EU, India and Singapore data-residency environments; Max is unavailable in India and Singapore. Mini is not available for Zero Retention Mode or HIPAA-configured workspaces. Availability and configuration.

These models manage the speech and language stack, limiting component customization. ElevenAgents billing combines per-minute usage with model tokens; Max costs more. Practical implication: confirm workspace access, residency and channel requirements before planning a migration.

OpenAI adds 12-hour prompt caching, with different API write costs

Developers can select a 12-hour cache lifetime or retain the 30-minute default for supported newer models. The 20 October announcement links to the updated API documentation; a 21 October community update says 12-hour caching is enabled by default in Codex and ChatGPT Work. Dated announcement, default-setting update.

For GPT-6.1 in the Responses API, prompt_cache_options.ttl accepts 12h or 30m; individual breakpoints can override the request setting. Cache writes cost 2.5 times the uncached input rate for 12 hours, versus 1.25 times for 30 minutes. These are total write rates, not added surcharges. Cache-read pricing stays the same. Older models retain their documented behavior. API caching guide.

Analysis: stable instructions and shared context reused hours apart may benefit, but higher write costs mean savings depend on actual reuse. A cache lifetime is not a data-deletion deadline. Codex credit billing has no separate cache-write charge; API-key usage follows API pricing. Credit-billing distinction.

Microsoft expands OneDrive Copilot and previews sovereign AI choices

OneDrive: file analysis, reusable skills and creative outputs

Microsoft’s 20 October OneDrive announcement describes Copilot analyzing multiple files, organizing folders, creating interactive HTML reports and saving repeatable processes as Markdown skills. It identifies Copilot in OneDrive and SharePoint as generally available, while the wider announcement mixes available features with coming-month rollouts. It does not provide a complete plan-and-region matrix. Original OneDrive announcement.

HTML and Markdown editing and collaboration are part of the update. For personal photos, Microsoft shows natural-language discovery, slideshow creation and AI editing. Availability caution: the future team-media workflow—searching, reviewing and approving photos and videos together—is a roadmap item, not a fully released capability. Creators should check the features exposed in their own app and account before relying on the demonstration.

Foundry Local: announced general availability at Ignite

Microsoft says Foundry Local will become generally available at Ignite, supporting connected and disconnected Azure Local deployments with a wider model catalog and local agentic retrieval. GitHub Enterprise Local and locally hosted Copilot inference are also announced for Ignite. These are future availability statements from 20 October, not evidence they are generally available today. Sovereignty announcement.

The connected-to-disconnected transition is entering public preview. The practical audience is organizations choosing where AI inference and developer infrastructure run; the announcement does not change every consumer Copilot account.

Gemini for Government Services

Google announced a preview for US federal, state and local agencies on 20 October. It combines government-oriented skills, connectors and partner agents, with source grounding and human approval for consequential steps. This is an agency offering rather than a new consumer Gemini plan. Official product explanation, preview scope.

A separate 20 October preview brings managed retrieval and search inside Google Distributed Cloud’s air-gapped environment. It targets NVIDIA Blackwell Ultra/HGX B300 infrastructure and supports mixed source material, including documents, images, WAV audio and MP4 video, with citations. Original technical announcement.

This matters to teams working with restricted multimedia collections on their own infrastructure. It is not a Veo, Flow or MusicFX consumer rollout; hardware, deployment and preview access remain separate requirements.

Claude Code and GitHub Copilot CLI fix developer workflows

Claude Code 2.1.305 and 2.1.306

Anthropic’s 2.1.305 release fixes a 2.1.304 regression in which rejected Tool Search requests could break calls through Vertex, Bedrock and other gateways. It retries without Tool Search. The release also adds resumable workflow journals and corrects review behavior around newer commits and unresolved blocking findings. It was published 20 October at 23:46 Istanbul. Release notes.

Two configuration changes merit attention: Linux nested-user-namespace support is opt-in and weakens filesystem sandbox protection; managed MCP configurations containing only environment-variable placeholders can now load where they were previously filtered out. Administrators who need exclusion should review explicit deny rules.

Version 2.1.306, published 21 October at 06:16 Istanbul, fixes a cloud headless-login error and misleading prompt-length errors when many MCP tools are present. Release notes. These are vendor-documented fixes, not a declared consumer Claude outage.

Copilot CLI stable release and separate prereleases

GitHub’s stable Copilot CLI 1.0.101, published 20 October, improves dynamic-workflow subagents and structured output on supported models, alongside MCP authentication, model-switching and sandbox fixes. Stable release.

The 21 October 1.0.102 prereleases separately address long-running commands, provider-overload retries and managed Auto behavior with headless or piped input. Teams should distinguish the prerelease channel from stable before upgrading. 1.0.102-0, 1.0.102-1.

Perplexity documents faster Computer models and more access controls

Perplexity’s 19 October changelog adds an Ultrafast option for GPT-6 Astra and GPT-6.1 Sol in Computer on the web for Max subscribers. It consumes more credits than Normal or Fast. Original changelog.

The same update documents:

A separate 20 October community announcement highlights Decider v1.2. Its benchmark-ranking claim is the provider’s, not MUVO’s evaluation. Model announcement.

Other developments affecting deployment and agent operations

DeepSeek and US civilian-agency procurement: an official notice published 20 October announces exclusion and removal requirements covering DeepSeek products, including third-party implementations, in federal civilian-agency procurement and IT. This is a defined government scope, not a public API shutdown or a ban on all private users. The notice’s publication date is not automatically the underlying order’s issue date. Federal Register notice.

Grok Bot support workflows: SpaceXSI’s 20 October case study describes IncidentWatch flagging ticket clusters for human investigation and IncidentSweep preparing a fixed batch of replies for approval before sending. It is a vendor-described internal use case, not a new public outage or a promise that every customer has those bots. Analysis: its useful lesson is the separation between detecting a pattern, confirming an incident and authorizing customer communication. Original case study.

Sources

Original announcements, documentation and release notes support the dates and availability labels above. We have not run paid generations, reproduced vendor benchmarks or tested these updates in private customer accounts. Sparse or inaccessible specialist-tool sources are recorded as research gaps; current feature pages alone do not establish new launches.

See the separate 21 October incident report, 20 October product news and our reporting method.

Update timeline