--- title: >- Voice models stopped waiting for tools, and the architecture decision moved to concurrency, metering, and trust boundaries publication: The Model Pulse slug: 2026-W38 issueNumber: 22 isoYear: 2026 isoWeek: 38 cadence: weekly publishedAt: '2026-09-19' periodLabel: Week 38 of 2026 canonicalUrl: https://brianletort.ai/industry/models/2026-W38 pdfUrl: https://brianletort.ai/downloads/model-pulse-2026-W38.pdf schemaVersion: 2026.05.02 treeDelta: added: - gemini-3-8-live - gemini-3-8-live-extended-thinking updated: [] frontierMovements: - modelId: gemini-3-8-live name: Gemini 3.8 Live vendor: Google DeepMind releaseDate: '2026-09-15' tier: frontier architecture: multimodal - modelId: gemini-3-8-live-extended-thinking name: Gemini 3.8 Live Extended Thinking vendor: Google DeepMind releaseDate: '2026-09-15' tier: reasoning architecture: reasoning - modelId: null name: Koa vendor: Salesforce and NVIDIA releaseDate: '2026-09-15' tier: specialist architecture: reasoning openWeights: - modelId: null name: Atria Dawn Preview vendor: Shanghai AI Laboratory releaseDate: '2026-09-14' tier: open_frontier architecture: moe architecturePatterns: - Conversation and execution become parallel planes - Domain post-training moves behind the system-of-record boundary - Compatibility becomes an open-model distribution feature benchmarks: - Artificial Analysis Speech-to-Speech leaderboard - Salesforce CRM Benchmark scorecardAsOf: '2026-09-19' vendorSignals: - vendor: Google date: '2026-09-15' signal: Live audio priced as separate input and output minute meters - vendor: Salesforce date: '2026-09-15' signal: Domain model offered in select pilots with US general availability expected in winter 2026 - vendor: Anthropic date: '2026-09-19' signal: Possible model release remains a reported deliberation rather than a product event --- # Voice models stopped waiting for tools, and the architecture decision moved to concurrency, metering, and trust boundaries *The Model Pulse · Issue 22 · Week 38 of 2026 · Published 2026-09-19* ## Big Read Google's Gemini 3.8 Live and Live Extended Thinking are the only new closed-frontier model releases verified in the September 14-19 window. The architectural change is not simply better speech: the live model can keep listening and responding while tools run in the background, while Extended Thinking can reason and narrate progress at the same time. Google prices audio input at $0.005 per minute and output at $0.018 per minute, so an architect can now model the conversational meter separately from the cost and latency of the tools behind it. Salesforce supplied the other consequential model event by post-training NVIDIA Nemotron 3 Super into Koa, a CRM reasoning model whose weights and inference remain inside Salesforce's trust boundary. That is a domain model in pilot, not a new foundation model, and its three-times-fewer-errors result is vendor-reported on Salesforce's own CRM benchmark. The open-weight side was quieter inside the window: Shanghai AI Laboratory published the technical report for Atria Dawn, a 744B Mixture-of-Experts agentic derivative whose weights appeared immediately before the window. The procurement read is to evaluate concurrency, meter composition, and deployment boundary alongside benchmark rank; none of those properties is captured by a single model score. ## Tree delta Two rows added for Google's live model pair. They share a voice-first lineage but separate low-latency scale from deeper concurrent reasoning. **Added (2):** `gemini-3-8-live`, `gemini-3-8-live-extended-thinking`. **Updated:** none. *Koa is covered as a domain post-training event but is not added to the general LLM tree this week; Salesforce has not published model parameters, a model card, or weights. Atria Dawn's paper landed in-window, while its weights appeared before the Monday cutoff, so it is covered without being counted as a W38 tree addition.* ## Frontier movements ### Gemini 3.8 Live `gemini-3-8-live` **Google DeepMind** · 2026-09-15 · frontier · multimodal _Native speech-to-speech model keeps dialogue active while visual grounding and tools run in the background_ The procurement change is concurrency: the model can acknowledge a request and continue the conversation while an API call finishes, rather than forcing a silent turn boundary. Google publishes separate audio input and output rates, making the conversational layer measurable, but buyers still need to add backend tools and any downstream model charges to get the workflow cost. Source: [Google](https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-3-8-live-gemini-3-8-live-extended-thinking/). ### Gemini 3.8 Live Extended Thinking `gemini-3-8-live-extended-thinking` **Google DeepMind** · 2026-09-15 · reasoning · reasoning _Concurrent deeper reasoning narrates progress without ending the live conversation_ Google separates a higher-complexity live model that reasons and speaks simultaneously, including progress narration during multistep work. That is operationally useful for long tool chains, but enterprises receive it only in private preview initially, so production plans should distinguish developer API availability from enterprise support and governance. Source: [Google](https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-3-8-live-gemini-3-8-live-extended-thinking/). ### Koa **Salesforce and NVIDIA** · 2026-09-15 · specialist · reasoning _CRM reasoning model post-trained from Nemotron 3 Super stays inside Salesforce's trust boundary_ Koa is a domain-specialized derivative moving into select pilots, not a frontier foundation release. Salesforce controls the weights and performs post-training and inference within its own infrastructure, which is the procurement advantage; its three-times-fewer-errors benchmark claim remains vendor-run and should be reproduced on the buyer's own CRM actions. Source: [Salesforce](https://www.salesforce.com/news/press-releases/2026/09/15/koa-reasoning-model/). ## Open weights ### Atria Dawn Preview **Shanghai AI Laboratory** · 2026-09-14 · open_frontier · moe _744B Mixture-of-Experts agentic derivative publishes its technical report under an MIT-weight release_ The report date falls inside the window, while the repositories appeared immediately before it, so this is a documentation event rather than a clean W38 weight launch. The permissive license and OpenAI-, Anthropic-, and Responses-compatible hosted interfaces lower switching friction, but a 744B system still carries a substantial serving floor and requires independent evaluation. Source: [Shanghai AI Laboratory technical report and model repositories](https://github.com/InternLM/Awesome-Atria-Dawn). ## Architecture watch ### Conversation and execution become parallel planes _Examples:_ Gemini 3.8 Live, Gemini 3.8 Live Extended Thinking. The live model no longer has to stop speaking while a tool or deeper reasoning path runs. That improves perceived latency but creates concurrent failure and cost domains, so architecture reviews need explicit behavior for tool timeout, stale narration, cancellation, and a user interrupt that arrives while background work is still active. Source: [Google](https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-3-8-live-gemini-3-8-live-extended-thinking/). ### Domain post-training moves behind the system-of-record boundary _Examples:_ Salesforce Koa, NVIDIA Nemotron 3 Super. Salesforce post-trained an open NVIDIA base on synthetic CRM-shaped data, retained control of the resulting weights, and runs inference in its own trust boundary. That gives an incumbent a middle route between a general frontier API and a self-hosted model, but it also concentrates evaluation, benchmark design, and deployment control with one vendor. Source: [Salesforce](https://www.salesforce.com/news/stories/why-we-post-trained-our-own-reasoning-model/). ### Compatibility becomes an open-model distribution feature _Examples:_ Atria Dawn Preview. Atria Dawn pairs downloadable weights with hosted interfaces compatible with several dominant API shapes. The practical switching cost therefore moves away from request syntax and toward evaluation, tool semantics, safety policy, and the infrastructure needed to serve a 744B Mixture-of-Experts model. Source: [Shanghai AI Laboratory repositories](https://github.com/InternLM/Awesome-Atria-Dawn). ## Benchmark moves ### Artificial Analysis Speech-to-Speech leaderboard Google reports Gemini 3.8 Live Extended Thinking at number one on the leaderboard at launch - Gemini 3.8 Live Extended Thinking: #1 - Gemini 3.8 Live: Released for lower-cost scale Source: [Google citing Artificial Analysis](https://blog.google/innovation-and-ai/technology/developers-tools/build-real-time-voice-applications-gemini-audio/). ### Salesforce CRM Benchmark Koa is reported to match or exceed leading models on CRM actions with three times fewer errors - Koa: 3x fewer errors, vendor-reported - Leading general models: Comparison set not published Source: [Salesforce](https://www.salesforce.com/news/press-releases/2026/09/15/koa-reasoning-model/). ## Tier scorecard _As of 2026-09-19._ | Tier | Leader | Challenger | Note | |---|---|---|---| | Closed frontier | Claude Fable 5.1 | GPT-6 Astra | No general frontier model displaced the W37 ordering; Reuters reported competitive pressure but no Anthropic release. | | Open frontier | DeepSeek-V4.1-Flash | Atria Dawn Preview | DeepSeek remains the more documented serving proposition; Atria Dawn brings permissive agentic weights at a much larger serving floor. | | Reasoning | Claude Fable 5.1 | Gemini 3.8 Live Extended Thinking | The challenger leads a speech-specific leaderboard but has not displaced the standing general reasoning leader. | | Coding | GPT-6 Astra | Claude Fable 5.1 | No coding benchmark moved comparably inside the window, so W37's order is carried forward. | | Multimodal | Gemini 3.8 Live Extended Thinking | Gemini 3.8 Live | The live pair owns this week's multimodal movement through native audio, visual grounding, and concurrent tools. | | Edge / small | Nex-N2.5 mini | Desert Ant on-device model set | No in-window edge release supplied stronger verified evidence, so the prior ordering remains. | ## Vendor signals - **2026-09-15 · Google — Live audio priced as separate input and output minute meters** - At $0.005 per input minute and $0.018 per output minute, the voice layer becomes independently budgetable. Buyers must still add tool, reasoning, media infrastructure, and idle-conversation costs before comparing it with a cascaded architecture. - _Source:_ [Google](https://blog.google/innovation-and-ai/technology/developers-tools/build-real-time-voice-applications-gemini-audio/) - **2026-09-15 · Salesforce — Domain model offered in select pilots with US general availability expected in winter 2026** - Koa is not generally available and its benchmark is vendor-controlled. Procurement should use the pilot to measure error rates on the organization's own CRM actions and preserve a general-model fallback. - _Source:_ [Salesforce](https://www.salesforce.com/news/press-releases/2026/09/15/koa-reasoning-model/) - **2026-09-19 · Anthropic — Possible model release remains a reported deliberation rather than a product event** - Reuters reported that Anthropic was considering a release amid investor scrutiny, but no model, date, price, or benchmark was announced. Treat it as competitive context and keep it out of production roadmaps. - _Source:_ [Reuters](https://www.reuters.com/business/anthropic-considers-releasing-new-ai-model-ahead-ipo-sources-say-2026-09-19/) ## Watchlist - **Sep 21-30 — Independent Gemini Live speech and interruption measurements.** The published rate card needs latency, interruption, and tool-timeout measurements before it supports architecture selection. - **Oct 2026 — Koa pilot benchmark disclosure.** A public task set, comparison roster, and absolute error rates would determine whether the reported multiple is reproducible. - **Winter 2026 — Koa US general availability.** General availability, regional scope, and model-catalog controls decide whether the specialist model moves beyond evaluation. - **Next 30 days — Anthropic model decision.** A release, delay, or explicit cancellation would turn this week's reported deliberation into a procurement-relevant event. ## Changelog - Evidence cutoff September 19, 2026; primary vendor sources used for releases, rates, availability, and benchmark labels. - Two Gemini Live rows added to the LLM Evolutionary Tree. Koa remains a Pulse-only specialist event pending a public model card or fuller architecture disclosure. - Atria Dawn is treated as an in-window technical-report event, not as a W38 weight launch, because the repositories appeared before September 14. --- Source of truth: `src/data/industry/models/2026-W38.ts`. Canonical HTML: . PDF: . Tree: .