AI news: Decisions API, ElevenLabs translation and Kling 4.0 in Luma — 23 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 Decisions API is generally available, ElevenLabs has opened an enterprise alpha for live translation, and Luma now offers Kling 4.0 inside its creation workflow. HeyGen’s dated announcements add a free voice-API preview and a general-purpose video model. Anthropic, Codex and Google have separate developer and advertising updates.

The research window is 20 October at 11:58 through 23 October at 11:57 Europe/Istanbul (UTC+3). Several announcements below were published on 21–22 October and are newly verified in this edition. Vendor quality and speed claims are attributed, not MUVO test results.

OpenAI: Decisions API exits beta with voice integration and caching

OpenAI made the Decisions API generally available on 22 October using GPT-6 Luna. Developers supply bounded questions and receive a choice, probability or score. Text, images and native JSON are supported, with up to 400 questions per request. New capabilities include prompt caching and native GPT-Live integration for decisions during live conversations. Launch announcement.

Published rates are $0.10 per million uncached input tokens and $0.01 per million cached input tokens, with no output-token charge. Cache writes for 30-minute and 12-hour retention have separate pricing. US and EU inference residency is available. Official pricing summary and documentation links.

OpenAI claims up to ten times the speed of GPT-6 Luna through Responses; that is a vendor comparison. Practical impact: developers can use the API for routing or scoring within voice and media workflows, while testing whether its constrained answers suit the decision they need. It is not a replacement for every free-form generation task.

ElevenLabs: live translation is an enterprise alpha

Eleven Realtime Translation launched in alpha for ElevenAPI enterprise customers on 21 October. It translates speech during a conversation so participants can speak and listen in their own languages. ElevenLabs describes uses in phone networks, contact centers and live events. Original announcement.

The service uses Eleven v4 Turbo by default. ElevenLabs lists more than 90 languages, automatic language detection, library voices and supported voice clones, with controls that trade translation speed against quality. Access is through enterprise onboarding; the announcement does not publish a self-service price or regional rollout table.

This is a distinct translation product from the Eleven Realtime voice-agent models covered on 21 October. Organizations should verify language pairs and latency against their own conversations before treating the alpha as production-ready.

Luma: Kling 4.0 joins the project board

Luma says Kling 4.0 is now available to every Luma customer, according to its 23 October announcement. Users select the model within a project board for text-to-video or image-to-video work, alongside other creation tools. Listed aspect ratios are 16:9, 9:16, 1:1 and 21:9. Luma’s integration announcement.

Luma positions the integration for fast action and complex camera movement. Those quality claims are its own. The post does not specify generation prices or establish that every capability in Kling’s standalone service is exposed through Luma. The news is the Luma integration, not a new launch date for the underlying Kling 4.0 model.

HeyGen: free voice preview and a general-purpose video model

Voice Turbo’s free API has capacity limits

HeyGen’s 21 October announcement offers Voice Turbo free through 30 November 2026. An account and API key are required; no credit card is required for the free model. Public voices and instant clones are supported, while professional clones require heygen-voice-1. Original Voice Turbo announcement.

The free tier is best effort, with no uptime or latency guarantee; capacity exhaustion can return 429. That makes the offer useful for evaluation, but it should not be confused with a guaranteed production service. HeyGen’s leaderboard claims are not MUVO comparisons.

HeyGen Video creates picture and sound together

HeyGen also lists a 21 October launch announcement for its general-purpose video model. The current product page describes text, image and reference input, native scene audio, clips of 5–15 seconds and output up to 2K. It supports up to 12 combined references, subject to per-media limits. Dated product listing, current HeyGen Video page.

The original article now redirects to that product page. Its broader account-plan table should not be read as a precise entitlement or credit quote for every model operation. Check the selected model’s in-product or API pricing before generation; we did not run a paid test.

Anthropic: a new managed route through Google Cloud

Claude Models on Google Cloud (Anthropic managed) launched on 22 October, including fast mode for Claude Opus 5.5. Anthropic operates the API, Google IAM controls access and Google Cloud Marketplace handles billing. This is distinct from the Google-managed offering. Dated release notes.

Setup creates a separate Anthropic organization for the Google Cloud project. Existing discounts and Zero Data Retention arrangements do not automatically transfer. The documented location is global; that identifier is separate from inference geography. Organizations should check model access, feature support and rate limits before moving traffic. Setup and migration documentation.

This released service is also separate from OpenAI’s forthcoming Google Cloud access covered yesterday.

Claude Code and Codex CLI: permission and workflow fixes

Claude Code 2.1.309 was published at 01:18 Istanbul on 23 October (22 October, 22:18 UTC). Its release notes describe fixes to sandbox startup-file handling, symlink-related write restrictions, MCP configuration protection, disabled-server reconnection and permission inheritance for agent teammates. Additional fixes address session recovery, hooks and editor behavior. Official 2.1.309 release.

These are maintainer-reported fixes. Teams using affected sandbox or automation workflows should test the updated behavior; the notes do not establish exploitation of any particular customer system.

Codex CLI 0.164.0 was released on 23 October. It adds ten reusable /presets slots, inline model and permission arguments, image previews in compatible terminals, broader enterprise MCP authentication and credential masking in supported remote sessions. Fixes cover restored session settings, compressed histories, compaction, Windows proxy authentication and macOS sandbox checks. Official 0.164.0 release.

The release also enables the Amazon Bedrock setup wizard by default. This is a CLI release; it does not establish that every desktop distribution updated simultaneously.

Google: Gemini features reach advertising workflows

Google’s 22 October Marketing Platform update describes Gemini-curated media bundles in Display & Video 360 Marketplace and a new Search Ads 360 Home experience. The latter brings together insights and suggestions; Marketplace bundles support one-click activation around live sports and major events. Google also names LinkedIn connected-TV inventory as a DV360 addition. Original Google update.

These changes concern the named advertising products, not the Gemini consumer app or Veo generation. The post does not provide a complete region or account-eligibility matrix. For creative teams, the practical change is in media planning and distribution after assets are produced; no return-on-investment improvement is guaranteed.

Hot topic: Agent Skills and Open Responses move to neutral governance

The Agentic AI Foundation announced four new projects on 22 October: Anthropic’s Agent Skills, OpenAI’s Open Responses, Hugging Face’s OpenEnv and Alibaba’s OpenSandbox. They cover reusable agent knowledge, model-interface interoperability, training environments and execution infrastructure respectively. Foundation announcement.

Anthropic says the Agent Skills donation includes its name, logo, GitHub organization and domain, and that participation does not require foundation membership. Versioned releases, a proposal process and security guidance are proposed next steps. Anthropic’s announcement.

Analysis: shared formats can reduce integration work for creative applications that use several providers. A governance transfer does not guarantee identical feature support, safe execution or permission to access a provider; each integration still needs its own checks.

Sources

• OpenAI: Decisions API launch, API pricing summary, Codex CLI 0.164.0.
• Creative tools: Eleven Realtime Translation, Kling 4.0 in Luma, HeyGen Voice Turbo, HeyGen product announcements.
• Anthropic: platform release notes, Google Cloud documentation, Claude Code 2.1.309.
• Google: Marketing Platform October update.
• Open standards: AAIF project additions, Agent Skills donation.

We checked the wider music, video, remix and major-provider watchlist. Undated pages, repeated announcements and inaccessible originals do not establish new releases. Service disruptions and recoveries appear in the separate incident edition. See our reporting method.

Update timeline

• 23 October 2026, 11:57 Istanbul / 08:57 UTC: Research cutoff. Added verified product changes, release notes and the open-standards development, distinguishing alpha, preview and available services.