# AI News, 13 October: Muse Wonder, Codex code review and expressive speech

Meta launches Muse Wonder for longer agent tasks. Also in this edition: a practical Codex review workflow and an older voice tool worth rediscovering, with clear access and language limits.

Published: 2026-10-13. Updated: 2026-10-13. Author: XANTRANS editorial.

Muse Wonder adds another option for sustained agent work. Codex brings review evidence into one place, while DramaBox offers directable English speech for creative prototypes. Access, data settings and output review still matter.

## A new model, a review workflow and a voice discovery

Meta announced Muse Wonder and Muse Wonder Full on October 12, aiming at tasks that take sustained work rather than one quick reply. That is the fresh release in today’s brief. Alongside it are two practical catches: Codex’s existing code-review experience, documented by OpenAI, and DramaBox, a May 13 voice-model release worth revisiting for expressive dialogue. These are different kinds of developments, not three products launched today.

## Muse Wonder: memory and persistence for longer jobs

Meta describes two central capabilities. Active Memory retrieves relevant details across conversations. Active Focus is designed to keep a task moving, recover from failed steps, try another approach and check important actions. The intended use extends from document-heavy work to coding. A useful example is a software change that must carry a brief through implementation, tests and revisions, where losing an earlier requirement can undermine the result.

Wonder is available through Meta Model API and Muse Code. Full is a separate variant available through Muse Code’s Power Usage and Maximum Usage subscriptions. Setting the API model’s reasoning to its maximum does not select Full. Meta says integration into the consumer Muse agent will follow later, first for subscribers, so the announcement should not be read as a feature already present in every Muse conversation.

The developer interface accepts text, images, video, audio and PDFs and returns text. It lists a 1,048,576-token context window, tool calling, structured output and web grounding. Those interfaces make it possible to build a workflow around mixed source material; they do not show how accurately a particular project will be completed. For a team evaluating it, a bounded existing task with known requirements and tests is more informative than a headline benchmark ranking.

Published API prices are $4 per million input tokens, $0.20 for cached input and $20 for output, with web search billed separately. Check data settings as well as price: Muse Code subscriptions default to Contributor mode, which allows submitted content to train future Meta models in exchange for more usage. It can be switched off. Wonder on Model API is Standard mode only, where content is not used for training. That distinction matters before uploading client code or confidential documents.

## Codex: bring the review and the conversation together

OpenAI’s code-review guidance describes a useful way to handle the work after code has been written. Code Review gathers pull requests, diffs, findings, comments and checks in one place. Each pull request has a private conversation where a reviewer can ask about the change, bring in team guidelines and prepare a response. This is coverage of an existing workflow, not a claim of a new model or a newly launched review service.

For example, a reviewer could ask how a sign-in change handles a failed database request, inspect the relevant lines, and request a narrowly scoped fix with a regression test. The point is to keep the question, evidence and proposed repair attached to the same change. Generated findings still need to be checked against the code and test results before comments are submitted or changes merged.

The plugin supports desktop and web, subject to repository permissions and workspace availability. GitLab merge-request review is in preview. That preview does not include automatic GitLab cloud code reviews, and connecting a repository does not grant broader access than the connected account already has.

## Older discovery: DramaBox directs a performance, not just a voice

Resemble AI released DramaBox on May 13. It takes performance direction in ordinary language: put the spoken dialogue inside quotation marks and describe delivery, pauses or emotional changes outside them. An optional short voice reference sets the timbre; without one, the model can generate a voice from the scene description. For a game or film prototype, this offers a way to explore how the same line changes when delivered hesitantly, warmly or with frustration.

Access is through the official model, code and hosted demo links. The repository lists roughly 24 GB peak GPU memory for its documented setup, so downloadable weights do not mean every laptop can run it comfortably. The model uses the LTX-2 Community License, whose conditions should be checked for the intended use. Its optional reference-denoising dependency has a separate non-commercial license.

The original release is explicitly English-only. Creator experiments in other languages are not evidence of official support, especially for a Latvian production. Use voices you have permission to use, retain the default watermark and check the generated performance before delivery. XANTRANS has not run an independent audio-quality comparison.

## Sources

- [Meta: Introducing Muse Wonder models, 12 October 2026](https://about.fb.com/news/2026/10/introducing-muse-wonder-models/) (checked 2026-10-13)
- [Meta Model API: models and availability](https://dev.meta.ai/docs/models) (checked 2026-10-13)
- [Meta Model API: pricing and Contributor mode](https://dev.meta.ai/docs/pricing-rate-limits) (checked 2026-10-13)
- [OpenAI: Review pull requests with Codex](https://help.openai.com/en/articles/20001552-review-pull-requests-with-codex) (checked 2026-10-13)
- [Resemble AI: DramaBox launch, 13 May 2026](https://www.resemble.ai/resources/dramabox-tts-saving-drama-for-the-performance-not-the-security-review) (checked 2026-10-13)
- [Resemble AI: DramaBox repository, setup and licensing notes](https://github.com/resemble-ai/DramaBox) (checked 2026-10-13)

Prepared with AI assistance from official sources. Practical examples are editorial suggestions; XANTRANS has not independently tested these models or workflows.


Original page: https://xantrans.com/blog/ai-news-2026-10-13-muse-wonder-codex-dramabox/