AI news: OpenAI custom voices, ChatGPT Slides and Meta Muse Spark Open — 17 October 2026

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.

Key product updates at a glance

OpenAI’s voice creation and slide workflows, Meta’s downloadable model, and Google’s study-video tools lead this edition. Developer updates also include Claude Code routine fixes and GitHub security, model and collaboration changes.

The scan covers 14 October at 12:00 through 17 October at 11:59 Europe/Istanbul (UTC+3). Items below retain their actual announcement dates; a 17 October edition is not a claim that every feature launched today. Performance statements are attributed to providers, not MUVO tests.

OpenAI: prompt-generated voices are available for GPT-Live

OpenAI’s 16 October release notes announce general availability of prompt-to-voice for API customers. A text description creates a saved voice that can be reused in GPT-Live sessions. Creating it is free; speech generation, including listening in the Playground, uses normal session billing. Dated release notes, custom-voice documentation.

The compatibility boundary matters: prompt-generated voices work with GPT-Live, not the separate Text to Speech, Realtime or Chat Completions APIs. Voice creation from a recording remains a different, eligibility-limited process with consent requirements. Developers should check that distinction before planning a narration or voice-agent workflow.

OpenAI also lists nine new GPT-Live voices across German, French, Hindi, Japanese and Korean. Its 15 October notes describe native Decisions API integration, allowing GPT-Live to invoke decisions directly or as transcript deltas arrive. The same day’s consumer update begins rolling out GPT-Live 1.1 and 1.1 Quip in ChatGPT Voice; voice choices can differ by account and region. These are separate consumer and API changes. OpenAI release notes.

ChatGPT: Slides rolls out to eligible workspaces

The 16 October ChatGPT notes introduce editable slide creation from conversations, uploaded files, outlines or reference presentations. Slides support text, images and charts, with collaboration and comments. Web exports include PowerPoint, PDF and Google Slides; desktop users can open presentations in installed PowerPoint or Keynote. ChatGPT release notes.

The rollout names eligible Pro, Business and Enterprise accounts with Space access on web and desktop. This does not establish availability for every ChatGPT account. For creators, the practical use is turning research or a campaign outline into a presentation that remains editable after export.

The same dated entry describes GPT-6 Pro in the chat tab, with provider-claimed response and tool/plugin improvements for eligible Pro users and workspaces. It is not an announcement that every API model has changed. Separately, the 15 October product notes remove the fixed number-of-Sites cap for Plus and Pro while retaining usage and storage limits. Product release notes.

Meta: Muse Spark Open weights and hosting options arrive

Meta announced Muse Spark Open on 15 October, with downloadable weights and partner-hosted access. Meta describes a multimodal model with 816 billion total parameters, 42 billion active parameters and a one-million-token context window. Those are provider specifications, not our benchmark results. Meta’s announcement.

The announcement names hosted options including Fireworks, Groq, Together and other partners, plus distribution through Hugging Face, Amazon SageMaker AI, Google Model Garden and Microsoft Foundry. Availability, billing and deployment terms depend on the chosen provider.

Meta demonstrates creative work including DaVinci Resolve workflows. Its local-computing example uses two M5 Ultra Mac Studios with 512 GB each, which should not be read as a claim that an ordinary laptop can run the same setup. The weights carry the Muse Model License, including commercial-use conditions; downloadable does not mean unrestricted use.

Analysis: For teams considering local creative automation, model access, usable hardware and license permission are separate checks. Hosted access may avoid local hardware requirements but does not automatically reproduce Meta’s demonstration or its performance claims.

Google: Gemini Notebook highlights short study videos

Google’s 16 October education guide highlights 60-second Short Video Overviews in Gemini Notebook. It also describes Interactive Learning Overviews that combine material with slides, quizzes and flashcards, and additional quiz formats with feedback. Google’s original guide.

Google describes these as recently launched features; the article’s publication date does not establish the precise release date of each format. The guide does not provide a complete plan or regional availability matrix. Check the tools exposed in your account before building a workflow around them.

For educators and creators, this is a documented study-content workflow. It does not establish a new standalone Veo release or a change to all Google video-generation products.

Claude Code 2.1.302: scheduled-routine and cost-display fixes

Anthropic published Claude Code 2.1.302 on 16 October. Its release notes say failed scheduled routines could previously restart from the beginning an hour or more later, potentially duplicating work; the release fixes that behavior. They also correct routines incorrectly appearing completed or cancelled when they had failed. Original versioned release.

Other fixes cover cloud-session usage resetting to zero after a process restart, prompt suggestions consuming credits after credit-billed replies, and tools failing with servers using the 2026-07-28 MCP protocol. Anthropic also documents a fix for errors after switching between Vertex and Bedrock during a session.

These are provider-described software fixes, not an independently measured outage or a guarantee that earlier duplicate work has been undone. Teams using scheduled routines can check their installed version and the affected run history. The notes also add an in-product notice that the standalone Claude Design app closes by 14 December 2026; this is a future product notice, not a shutdown today.

GitHub: security tools, HydraFusion Max and model retirements

AI Scan and agentic autofix become generally available

GitHub’s 15 October AI Scan announcement expands pull-request security analysis beyond CodeQL’s coverage. Findings are advisory: they do not fail the code-scanning check or block merges. Fork and Dependabot pull requests are excluded. Public repositories are supported; private/internal organization repositories require Code Security or Advanced Security. A Copilot license is not required, but usage consumes Code Security AI Credits. AI Scan announcement.

The separate agentic autofix release can investigate default-branch alerts, propose a fix, rerun CodeQL and open a draft pull request. Enterprise Cloud customers need the relevant security product, administrator enablement and a Copilot billing method; the workflow consumes security AI Credits and Actions minutes. Human review remains part of the documented flow. Agentic autofix announcement.

HydraFusion Max remains a research preview

The 14 October HydraFusion update adds Max and revises its model pool and orchestration. Both options work across VS Code, the GitHub Copilot app and CLI, on Pro, Pro+, Business and Enterprise plans with preview features enabled, subject to organization policies. GitHub describes higher reasoning effort for demanding Max tasks; no independent performance comparison is established here. HydraFusion update.

Selected Copilot models retire on 16 November

GitHub announced a 16 November retirement date for Claude Opus 4.8, Claude Opus 5 and GPT-5.6 Terra across Copilot experiences. Individual annual subscribers retain an explicit Opus 4.8 exception. Suggested replacements are Opus 5.5 and GPT-6.1 Sol or GPT-6 Luna; administrators should check their model policies before the date. These are GitHub availability changes, not evidence that Anthropic or OpenAI is retiring the same models everywhere. Deprecation notice.

GitHub’s 15 October Slack/Teams update also enables work across multiple repositories in one conversation and better use of shared Slack context. The announcement does not supply a complete plan-eligibility matrix. Multi-repository collaboration update.

ElevenLabs: new enterprise assurance claims have specific scope

ElevenLabs’ 16 October announcement lists certification renewals and new assurance milestones. In particular, it identifies SOC 2 Type I for ElevenProductions and FedRAMP 20x Class A. These are the provider’s published claims; we have not independently audited them or extended their scope to every ElevenLabs product. Procurement teams should inspect the underlying documents for the service they intend to use. ElevenLabs announcement.

Sources

• OpenAI: product release notes for 15–16 October; custom voices; ChatGPT release notes, 16 October.
• Meta: Muse Spark Open, 15 October.
• Google: Gemini Notebook study formats, 16 October.
• Anthropic: Claude Code 2.1.302, 16 October, 20:53 UTC / 23:53 Istanbul.
• GitHub: AI Scan, agentic autofix, model retirements and Slack/Teams, 15 October; HydraFusion Max, 14 October.
• ElevenLabs: certification renewals and additions, 16 October.

Music, video and remix/stem specialists were included in discovery. Previously covered releases, including Suno collaboration and ElevenLabs Instant Soundtracks, remain in the 16 October edition. Some sources were inaccessible or lacked a dated material update; no entry here means no qualifying new finding, not proof a product is unchanged. This scan is not continuous monitoring. Outages and customer problems appear in the separate 17 October incident edition.

Update timeline

• 17 October 2026, 11:59 Istanbul: Research cutoff. Consolidated original announcements and release notes, separating consumer features, APIs, versioned fixes, previews and future retirements. Provider claims have not been independently benchmarked.