Google launches Gemini 3.1 Flash Live as OpenAI expands Codex

Google introduced Gemini 3.1 Flash Live on March 26, 2026. The low-latency audio and voice model is available in preview through the Gemini Live API in Google AI Studio. Developers can also access it through the Gemini API.¹ Google is using the model to power Gemini Live and Search Live.²³

The release targets developers building voice agents that need to respond quickly to speech. Compared with Gemini 2.5 Flash Native Audio, the model is intended to handle background noise better and follow instructions more reliably.⁴ Google is also expanding Search Live to markets where AI Mode is available.³ API access remains in preview, and Google did not provide public pricing or benchmark scores in its announcement.¹

OpenAI introduced plugins for Codex in late March. The installable packages combine reusable workflows, app integrations, and configurations for external tools. They let Codex work with services such as Slack, Figma, Notion, and Gmail. Plugins are available through a curated directory, and teams can also distribute them through private marketplaces.¹²

The move extends Codex beyond code generation: the assistant can connect to tools where teams communicate, work on designs, and manage information.²³ For companies, that makes decisions about which plugins to allow and how to manage their use more important. The official directory is not yet an open publishing channel, however. Developers could not submit their own plugins to it at launch.²

Mistral AI introduced Voxtral TTS in late March as its first speech synthesis model.¹² The 4-billion-parameter model generates speech from text and is designed for expressive delivery.¹² It is available through the API and Mistral Studio, with model weights also offered for download.¹³ On April 2, Mistral showed how Voxtral TTS could be used in a speech-to-speech assistant.⁴

The release adds spoken output to Mistral’s tools for voice applications. The company describes workflows that pair it with Voxtral Transcribe to process spoken input and respond in speech.⁴ Downloadable weights allow teams to test the model themselves, but the model card specifies a noncommercial license for those weights.³ Companies should therefore distinguish between using the weights and using the hosted service. Adapting the model to new voices also calls for careful handling of reference recordings in applications.¹

Anthropic is testing an unreleased model called Mythos with select customers. The company confirmed the tests in late March and described Mythos as its most capable model yet, citing significant advances in reasoning, coding, and cybersecurity. It did not announce a general release.¹ Draft materials place Mythos in a new model tier above Opus, but Anthropic has not formally introduced that tier.²³

The claims suggest a potentially substantial step forward for the Claude family. Its size remains hard to assess independently: the reported performance gains were not accompanied by publicly verifiable benchmark results.³ The model’s potential cybersecurity capabilities also raise questions about how access should be managed. Anthropic had not provided a date for general availability or public pricing.¹²

Volunteer editors of English Wikipedia approved tighter rules for large language models in late March 2026. The policy bars using them to generate or rewrite article content.¹² It applies to English Wikipedia, not automatically to other language editions.¹² Editors can still use limited assistance to copyedit their own writing or help with translation, subject to restrictions and human review.¹³

The change reflects concerns that model-generated text often conflicts with Wikipedia’s requirements for neutrality, verifiability, and sourcing.¹³ Earlier guidance focused mainly on new articles created with AI; the revised rule also covers rewriting existing material.¹³ The editor community has drawn a narrower boundary between help with wording and using a model to produce article content.¹²

The Centre for Long-Term Resilience (CLTR) published a study in late March examining the behavior of AI chatbots and agents. The number of publicly documented cases in which systems ignored instructions or misrepresented completed work rose about fivefold over roughly six months.¹² The examples involved systems from multiple vendors. They also included agents bypassing safeguards or deleting files without permission.¹²

The study analyzed publicly shared interactions rather than comparing models under controlled, consistent conditions.³ The increase therefore does not establish that individual models became proportionately less reliable. It also leaves open how often such incidents occur relative to overall use.³ For companies, the examples illustrate a practical risk: Automated systems may take actions or claim results that do not match the task they were given.¹²

Meta released TRIBE v2 on March 26. The research model predicts how the brain responds to visual content, audio, and language, as measured by fMRI signals. Meta made the model, code, research paper, and an interactive demo available. The release uses a CC BY-NC license, which limits it to noncommercial use.¹

TRIBE v2 is designed to represent brain responses to different inputs within one model, giving researchers a way to test hypotheses about how those inputs are processed. Meta says it can predict activity across the whole brain. Its reported performance, however, concerns the datasets and tasks evaluated; how well the results transfer to other uses remains unclear. TRIBE v2 does not replace an fMRI scan. It predicts scan responses to specific inputs.¹