{"schemaVersion":1,"editionId":"2026-09-29","title":"ElevenLabs releases v4 voices; NVIDIA adds agent controls and Meta targets enterprises","dek":"New speech tools and agent infrastructure, plus dated catch-ups on NaiveAI's open model and game-based evaluation.","coverageNote":"Partial research for September 29, Europe/Berlin. With no completed prior cutoff, we assumed September 28 at 12:45 CEST to September 29 at 12:45 CEST, plus a partial September 15–29 gap scan. Date-only announcements cannot be placed precisely within that window. Catch-ups retain their original dates. X originals and several safety, model and creative leads remain unverified; no complete research cutoff advances.","summaryIntro":"Today's releases raise practical questions: how to direct a synthetic voice, where an agent's permissions are enforced, and what is available versus planned. Two catch-ups examine the hardware behind open weights and why a model's game ranking depends on the game.","publishedAt":"2026-09-29T10:55:10.056Z","updatedAt":"2026-09-29T10:55:10.056Z","version":1,"items":[{"id":"eleven-v4-speech","kind":"news","thumbnail":{"src":"/api/media/thumbnails/2026-09-29/eleven-v4-speech/80c1f343abe1a99d78ad9568a4a33d10ff6a2308486f321bb5877106b89dc5b9","alt":"Conceptual editorial paper collage of a microphone and layered expressive audio waveforms.","credit":"AI Daywire · AI-assisted editorial illustration","kind":"editorial-illustration"},"title":"ElevenLabs releases v4 speech models with expressive direction and a Turbo option","category":"Creativity & Media","eventDate":"2026-09-28","status":["available","company-claim"],"sourceIds":["S1"],"short":"ElevenLabs launched Eleven v4 and v4 Turbo on September 28, with natural-language voice direction and inline audio tags. The company says both support more than 90 languages and are available through ElevenCreative, ElevenAgents and ElevenAPI.","claims":[{"text":"ElevenLabs reports roughly 150 milliseconds to first audible speech for Turbo, excluding network latency.","sourceIds":["S1"]}],"long":[{"text":"Creators can describe delivery and add tags for emotions or sounds. ElevenLabs presents v4 as its expressive speech model and Turbo as the low-latency option for conversational agents.","sourceIds":["S1"]},{"text":"The advertised roughly 150-millisecond median measures request-to-audible-speech over WebSocket streaming with network delay removed; it is not a guaranteed end-to-end response time. We have not tested either model. The announcement was updated September 29.","sourceIds":["S1"]}]},{"id":"nvidia-open-agent-safety","kind":"news","thumbnail":{"src":"/api/media/thumbnails/2026-09-29/nvidia-open-agent-safety/28f3705f939e6558da4ae42cd5e9615224d7f8ecd075bf56f7b7420a9b349e73","alt":"Conceptual editorial paper collage showing a protective boundary around agent task cards.","credit":"AI Daywire · AI-assisted editorial illustration","kind":"editorial-illustration"},"title":"NVIDIA announces agent controls spanning OpenShell software and optional Sentry hardware","category":"Safety & Society","eventDate":"2026-09-28","status":["announced","company-claim"],"sourceIds":["S2","S3"],"short":"NVIDIA's September 28 Open Agent Safety Platform combines OpenShell, which the company says is broadly available, with a Sentry reference design for separate hardware monitoring. The distinction matters: deploying the software does not itself supply the hardware layer.","claims":[{"text":"NVIDIA describes Sentry as an optional enforcement layer running on BlueField-4 DPUs, outside the agent's host environment.","sourceIds":["S2","S3"]}],"long":[{"text":"OpenShell lets operators define which files, networks, tools, processes and credentials an agent may access, then enforces those policies during execution. NVIDIA's technical explanation places the controls outside the model itself.","sourceIds":["S3"]},{"text":"Sentry adds monitoring on BlueField-4 hardware. NVIDIA claims it can quarantine an agent in milliseconds; we have not independently tested that claim or the platform's resistance to escape attempts. The announcement distinguishes available software from a reference system design and does not establish universal hardware availability.","sourceIds":["S2","S3"]}]},{"id":"naive-n05-flash-catchup","kind":"quick-update","thumbnail":{"src":"/api/media/thumbnails/2026-09-29/naive-n05-flash-catchup/165bf34b9c4cd5fe847fdeee2f19c6fa64a5f545093dbfc95025f60cc6e515e2","alt":"Conceptual editorial paper collage of research notes and a sparse geometric grid for AI-assisted research.","credit":"AI Daywire · AI-assisted editorial illustration","kind":"editorial-illustration"},"title":"Catch-up: NaiveAI releases a large open-weight model for coding and AI research","category":"Models & Agents","eventDate":"2026-09-27","status":["available","company-claim"],"sourceIds":["S4","S5"],"short":"NaiveAI's September 27 Naive-N0.5-Flash announcement describes a 309-billion-parameter model with 15.5 billion active parameters and a one-million-token context. Its authors say weights and inference code use the MIT license; their release text still describes API access as forthcoming.","claims":[{"text":"The model card says FP8 deployment requires compatible NVIDIA GPUs and approximately 315 GB for weights, plus inference memory.","sourceIds":["S5"]}],"long":[{"text":"The model builds on MiMo-V2.5 and replaces full-attention layers with a combination of local and sparse attention. The weights are openly released, but the documented memory requirement makes this a substantial deployment rather than a typical laptop model.","sourceIds":["S4","S5"]},{"text":"NaiveAI reports peak single-stream decoding of 2,122 tokens per second on eight GPUs using its specialized NaiveRT runtime. That is a vendor result, not expected speed on arbitrary hardware. The team describes human researchers guiding AI-assisted development; we have not reproduced its benchmarks or verified live API access.","sourceIds":["S4"]}]},{"id":"meta-enterprise-platform","kind":"news","thumbnail":{"src":"/api/media/thumbnails/2026-09-29/meta-enterprise-platform/87ecd6c1cb9a6f08b0fd9d8db0749fd68c059b44315613d7219bdd0cc5893afe","alt":"Conceptual editorial paper collage of an office building linked to task cards.","credit":"AI Daywire · AI-assisted editorial illustration","kind":"editorial-illustration"},"title":"Meta announces an enterprise AI effort led by CJ Desai","category":"Business & Infrastructure","eventDate":"2026-09-28","status":["announced","company-claim"],"sourceIds":["S6"],"short":"Meta announced Meta Enterprise Platform on September 28, with an initial focus on bringing Muse, Meta Business Agent, Muse API and Muse Code to businesses and developers. Chirantan 'CJ' Desai will join as Chief Enterprise Platform Officer, reporting to Mark Zuckerberg.","claims":[{"text":"Meta frames Enterprise Platform as a new business effort focused on converting its AI stack into deployable enterprise products and services.","sourceIds":["S6"]}],"long":[{"text":"This is a new organizational and product-direction announcement, distinct from the consumer Muse updates covered earlier. Meta says Desai joins from MongoDB, where he was CEO and president.","sourceIds":["S6"]},{"text":"The statement outlines what the enterprise effort will focus on; it does not provide a unified launch price, regional rollout or general-availability date for the whole suite. Those details should not be inferred from the platform's name.","sourceIds":["S6"]}]},{"id":"game-arena-report-catchup","kind":"news","thumbnail":{"src":"/api/media/thumbnails/2026-09-29/game-arena-report-catchup/7d0bb9a2ab7ea71f9ca127480140b22a7ee9d093615b21af1ccad2d457e2a4d0","alt":"Conceptual editorial paper collage of a chess knight, playing cards and role tokens for game-based model evaluation.","credit":"AI Daywire · AI-assisted editorial illustration","kind":"editorial-illustration"},"title":"Catch-up: Game Arena report shows why one game cannot rank every AI skill","category":"Science & Research","eventDate":"2026-09-25","status":["research-preprint"],"sourceIds":["S7","S8"],"short":"A September 25 technical preprint describes Kaggle Game Arena evaluations in chess, poker and Werewolf. The authors find that model strength and cost-performance ordering vary across games, complicating the idea of a single best strategic model.","claims":[{"text":"The report's limitations include inefficient fixed match counts and difficulties comparing results as model versions enter or leave the evaluation.","sourceIds":["S8"]}],"long":[{"text":"The three pilot games probe different conditions: visible information, hidden information and multiplayer interaction. The authors say they release an evaluation harness and full game trajectories to support further research. September 25 is the report's submission date, not a newly established launch date for Game Arena.","sourceIds":["S7","S8"]},{"text":"Their conclusion is useful when reading leaderboards: rankings from one setting do not automatically transfer to another. The paper reports game-specific measures and leaves a consolidated cross-game rating for future work. These are preprint results on the evaluated model pool, not evidence of general workplace competence or our own replicated tests.","sourceIds":["S8"]}]}],"sources":[{"id":"S1","title":"Introducing Eleven v4, our most emotive model","publisher":"ElevenLabs / Mati Staniszewski and Piotr Dabkowski","url":"https://elevenlabs.io/blog/eleven-v4","publishedDate":"2026-09-28","kind":"vendor-announcement"},{"id":"S2","title":"NVIDIA Launches Open Agent Safety Platform to Secure Agents From Testing to Deployment","publisher":"NVIDIA","url":"https://investor.nvidia.com/news/press-release-details/2026/NVIDIA-Launches-Open-Agent-Safety-Platform-to-Secure-Agents-From-Testing-to-Deployment/default.aspx","publishedDate":"2026-09-28","kind":"vendor-announcement"},{"id":"S3","title":"NVIDIA Open Agent Safety Platform: A Reference for Continuous In-Silicon Agent Monitoring","publisher":"NVIDIA","url":"https://developer.nvidia.com/blog/nvidia-open-agent-safety-platform-a-reference-for-continuous-in-silicon-agent-monitoring/","publishedDate":"2026-09-28","kind":"vendor-technical-explanation"},{"id":"S4","title":"Naive-N0.5-Flash: Building Frontier AI with AI","publisher":"NaiveAI Team","url":"https://naive.ai/en/research/","publishedDate":"2026-09-27","kind":"vendor-announcement"},{"id":"S5","title":"Naive-N0.5-Flash model card","publisher":"NaiveAI / Hugging Face","url":"https://huggingface.co/NaiveAI/Naive-N0.5-Flash","publishedDate":null,"kind":"official-model-card"},{"id":"S6","title":"Launching Meta Enterprise Platform","publisher":"Meta","url":"https://about.fb.com/news/2026/09/launching-meta-enterprise-platform/","publishedDate":"2026-09-28","kind":"company-statement"},{"id":"S7","title":"Game Arena: Strategic LLM Evaluation in Competitive Environments","publisher":"Bovard Doerschuk-Tiberi et al. / arXiv","url":"https://arxiv.org/abs/2609.31473","publishedDate":"2026-09-25","kind":"research-preprint"},{"id":"S8","title":"Game Arena — full technical report","publisher":"Bovard Doerschuk-Tiberi et al. / arXiv","url":"https://arxiv.org/html/2609.31473v1","publishedDate":"2026-09-25","kind":"research-preprint"}],"review":{"passed":true,"reviewedAt":"2026-09-29T10:53:53.399Z","method":"source-check"},"corrections":[],"canonicalUrl":"https://aidaywire.com/briefings/2026-09-29","language":"en","aiGenerated":true,"summary":"Today's releases raise practical questions: how to direct a synthetic voice, where an agent's permissions are enforced, and what is available versus planned. Two catch-ups examine the hardware behind open weights and why a model's game ranking depends on the game.\n\nElevenLabs launched Eleven v4 and v4 Turbo on September 28, with natural-language voice direction and inline audio tags. The company says both support more than 90 languages and are available through ElevenCreative, ElevenAgents and ElevenAPI.\n\nNVIDIA's September 28 Open Agent Safety Platform combines OpenShell, which the company says is broadly available, with a Sentry reference design for separate hardware monitoring. The distinction matters: deploying the software does not itself supply the hardware layer.\n\nNaiveAI's September 27 Naive-N0.5-Flash announcement describes a 309-billion-parameter model with 15.5 billion active parameters and a one-million-token context. Its authors say weights and inference code use the MIT license; their release text still describes API access as forthcoming.\n\nMeta announced Meta Enterprise Platform on September 28, with an initial focus on bringing Muse, Meta Business Agent, Muse API and Muse Code to businesses and developers. Chirantan 'CJ' Desai will join as Chief Enterprise Platform Officer, reporting to Mark Zuckerberg.\n\nA September 25 technical preprint describes Kaggle Game Arena evaluations in chess, poker and Werewolf. The authors find that model strength and cost-performance ordering vary across games, complicating the idea of a single best strategic model.","fullText":"ElevenLabs releases v4 speech models with expressive direction and a Turbo option\n\nCreators can describe delivery and add tags for emotions or sounds. ElevenLabs presents v4 as its expressive speech model and Turbo as the low-latency option for conversational agents.\n\nThe advertised roughly 150-millisecond median measures request-to-audible-speech over WebSocket streaming with network delay removed; it is not a guaranteed end-to-end response time. We have not tested either model. The announcement was updated September 29.\n\nNVIDIA announces agent controls spanning OpenShell software and optional Sentry hardware\n\nOpenShell lets operators define which files, networks, tools, processes and credentials an agent may access, then enforces those policies during execution. NVIDIA's technical explanation places the controls outside the model itself.\n\nSentry adds monitoring on BlueField-4 hardware. NVIDIA claims it can quarantine an agent in milliseconds; we have not independently tested that claim or the platform's resistance to escape attempts. The announcement distinguishes available software from a reference system design and does not establish universal hardware availability.\n\nCatch-up: NaiveAI releases a large open-weight model for coding and AI research\n\nThe model builds on MiMo-V2.5 and replaces full-attention layers with a combination of local and sparse attention. The weights are openly released, but the documented memory requirement makes this a substantial deployment rather than a typical laptop model.\n\nNaiveAI reports peak single-stream decoding of 2,122 tokens per second on eight GPUs using its specialized NaiveRT runtime. That is a vendor result, not expected speed on arbitrary hardware. The team describes human researchers guiding AI-assisted development; we have not reproduced its benchmarks or verified live API access.\n\nMeta announces an enterprise AI effort led by CJ Desai\n\nThis is a new organizational and product-direction announcement, distinct from the consumer Muse updates covered earlier. Meta says Desai joins from MongoDB, where he was CEO and president.\n\nThe statement outlines what the enterprise effort will focus on; it does not provide a unified launch price, regional rollout or general-availability date for the whole suite. Those details should not be inferred from the platform's name.\n\nCatch-up: Game Arena report shows why one game cannot rank every AI skill\n\nThe three pilot games probe different conditions: visible information, hidden information and multiplayer interaction. The authors say they release an evaluation harness and full game trajectories to support further research. September 25 is the report's submission date, not a newly established launch date for Game Arena.\n\nTheir conclusion is useful when reading leaderboards: rankings from one setting do not automatically transfer to another. The paper reports game-specific measures and leaves a consolidated cross-game rating for future work. These are preprint results on the evaluated model pool, not evidence of general workplace competence or our own replicated tests.","checksum":"94bd30d3f9c31cfa194572e9fff6daa8ae00a483d8c258a070fe56adb2f38602"}