The AI Wire ●

SpaceXAI (xAI) released Grok Voice Transcribe 2.0 on September 18, a speech-to-text model it calls the most accurate available — claiming double the accuracy of version 1.0 at unchanged pricing: $0.10 per audio hour for batch transcription and $0.20 for real-time streaming, with speaker diarization, word-level timestamps, and key-term biasing included. It is built on the audio foundation behind Grok Voice, trained on live, noisy, multilingual audio and refined with post-training.

The company says word error rate fell from 20.6% to 6.8% on short multilingual voice commands, from 10.6% to 7.1% on telephony, and from 8.7% to 3.3% on conversational audio — and that it ranks first among 32 streaming models on the public Artificial Analysis leaderboard. Those are vendor-reported numbers; independent confirmation matters more than usual here, since xAI's Terminal-Bench claims on Grok 4.7 recently showed a 12-point gap to independent runs.

Distribution is the interesting part: Atlassian's Loom now uses it to transcribe every video, per RuntimeWire. The model is live through the Speech-to-Text API as grok-voice-transcribe-2.0, with 1.0 headed for deprecation in the coming weeks — developers who need the old model should pin grok-voice-transcribe-1.0 explicitly.

Qwen's first omni-modal model accepts text, images, audio, and video with 1M context and is built for agentic tool-calling workflows. API is $0.15 in / $0.47 out per million tokens, with audio input pricing cut roughly 98%. Live on QwenCloud, Alibaba Cloud Model Studio, and Qwen Studio; no open weights announced.

A speed-focused iteration of GLM-5.3-Flash (320B total / 18B active) claiming 200 tokens/second — 5x the prior Flash — optimized by a self-hosted 'Infra Agent'; inference reportedly runs wholly on domestically produced accelerators. Priced at 2.5x the original Flash rate; Zhipu's stock gained roughly 7% on the news. Throughput and hardware claims are lab-reported and not yet independently reproduced.

Grok models are now selectable inside Microsoft Copilot, confirmed September 12 — the xAI/Microsoft partnership moving from press release to product. Users get a non-Microsoft model choice inside the same Copilot surface.

Separately, SpaceXAI was added as a Copilot subprocessor with an opt-in admin setting that enables Grok inside Word, Excel, and PowerPoint for Frontier customers (excluding EU/EFTA/UK and government tenants), effective September 18. This is the real distribution play: Grok inside Office documents.

Watch the opt-in numbers — model choice inside enterprise Copilot is a genuine shift in how frontier models reach office workers.

The UAE Cyber Security Council and Open Innovation AI announced a specialized AI model for cybersecurity, built in-house in the UAE for vulnerability analysis, secure code analysis and remediation, incident investigation, and security operations support. The launch (announced September 17) is framed as sovereign, locally driven AI capability; no model name, specs, or availability details were published.

Caltech-born compression startup PrismML released Bonsai 2 27B, a derivative of Alibaba's Qwen3.8 27B compressed to 5.9 GB — a 9–10x memory reduction that can fit on a PC and possibly a high-end smartphone. Backed by Khosla Ventures and advised by Databricks co-founder Ion Stoica; a rumored Apple partnership is unconfirmed (CEO declined to comment).

Meta released a dedicated Mac desktop app for its Muse personal AI agent, extending the assistant from mobile and web to the desktop. On the Mac, Muse can interact with files, messages, calendar, notes, and mail inside their native applications. Alexandr Wang announced it on X, and Mark Zuckerberg punctuated the launch by saying "the team is shipping fast."

Meta framed the desktop agent around opt-in permissions: users control what Muse can access on their hard drives, and the app asks for approval before sensitive actions. The permissions model later became controversial after complaints that Muse accessed private Messages data, prompting Apple to announce tighter Full Disk Access controls on October 2.

The Mac launch sits inside a crowded consumer-agent race. Meta has been expanding Muse fast — early access programs, smart-glasses support, open-source hardware gadgets — while competitors push their own agentic assistants, all of them chasing the same thesis that agents, not chatbots, are the future interface for computing.

OpenAI’s release notes (dated September 17) extend multi-account plugin connections beyond Gmail, Google Calendar, and Google Contacts — and lift the plan restriction entirely. The feature is now “available on all ChatGPT plans across web, mobile, and desktop,” so Free and Go users are included this time.

The point is practical: connect a personal and a work account to the same plugin and ChatGPT can pull from both in one conversation. For anyone whose calendar is split between an employer account and a private one, checking both calendars before accepting an invitation finally works without switching sign-ins.

One gap worth noting: OpenAI names only the three Google plugins it started with and doesn’t say which other plugins now accept a second account, or how many accounts each can hold. Which plugins support it shows up in each plugin’s own ChatGPT settings instead — and workspace controls still sit on top of availability.

ChatGPT’s Word add-in is now generally available on every plan. A ChatGPT sidebar inside Word drafts from notes, summarizes, revises text, and adjusts headings and formatting — joining Excel and PowerPoint behind the same add-in.

For Business, Enterprise, and Education workspaces the add-in switches to default-on starting October 1, 2026. If your org wants it off, that’s an admin setting to change before the deadline.

This is the practical, user-relevant tier of ChatGPT news: writing workflows inside the apps people already open all day.

Grok Build — xAI’s coding agent — now keeps persistent markdown notes on project conventions and decisions, and reads them back in later sessions. The agent stops re-learning your codebase’s habits every time you open a new chat.

It’s live for new sessions. For anyone using Grok Build on real projects, this is the difference between a clever autocomplete and a collaborator with a memory.

Memory is becoming table stakes for coding agents — Claude Code, Codex, and Copilot all have some form of it. xAI is just catching up, in public, on a live product.

TypeSafe AI, a startup with $40 million in funding, released Jev — a model that returns typed probabilistic decisions (e.g. {"billing": 0.08, "technical": 0.85}) instead of natural language, aimed at machine-to-machine use like sorting support tickets and agent workflows. The Register demoed it playing Doom off structured game-state data; the pitch is cutting the parse-and-validate overhead of LLM text output.

Google DeepMind released Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking on September 15, its most advanced live dialogue models to date. Both are native speech-to-speech models for real-time voice agents, live in the Gemini Live API and Google AI Studio. The headline trick: they reason and call tools in the background while the conversation continues, instead of going silent while the agent works.

The pair splits by job. 3.8 Live is the scale tier for low-latency conversation; Live Extended Thinking is built for multi-step work like technical diagnosis or tutoring, and it tops Artificial Analysis' Speech-to-Speech Quality Index at 82.6. Google reports 97 language support, near-real-time visual understanding, and audio watermarked with SynthID.

Early enterprise partners include Salesforce, Genspark and Lumeris. The release extends the Gemini push into voice, a front where OpenAI's Realtime API and newer agentic voice startups have been competing — and where the ability to act mid-conversation is becoming the key differentiator.

UPDATE 2026-09-23: the rollout keeps widening. Google’s conversational Search Live now runs on 3.8 Live for more natural voice search, and the Gemini Omni Flash Preview endpoint is deprecated September 30 — developers still on the preview must migrate to 3.8 Live by then. Low-latency audio-to-audio voice models are now Google’s default voice stack.

UPDATE 2026-09-23: migration deadline — Google's earlier Gemini Omni Flash Preview endpoint is scheduled for deprecation on September 30, 2026. Developers still on the preview must move production voice workloads to Gemini 3.8 Live or 3.8 Live Extended Thinking before the cutoff. (ts2.tech, citing Google's blog and Gemini API release notes)

Gensyn launched open-1b, which the company calls the industry's first auditable AI model — researchers can inspect and independently verify the training process rather than simply trusting the released weights, code, and model card. It builds on Gensyn's work on verifiable distributed compute.

Sakana AI released Fugu Max v1.0 and Fugu Ultra v2.0 on September 11 — not monolithic models but learned orchestration engines: a language model trained to route a task across a pool of open-weight and specialized models (including NVIDIA Nemotron) and recursively call instances of itself, via an OpenAI-compatible endpoint. Fugu Max is the cost tier at $2 per million input and $6 per million output tokens; Fugu Ultra v2 is the quality tier at $5/$30 ($10/$45 above 272K context), with a 1M-token context window.

Ultra v2 ships without any proprietary frontier model in its pool — its 48.3 on Chartography reportedly beats Opus 5's 27.3 without using Opus. API-only; no downloadable weights; not available in the EU/EEA. (Filed September 21 as backfill — a significant release the sweep missed on September 11.)

DeepSeek launched V4.1 Flash, a 552B mixture-of-experts model (only ~8B parameters active while reading, ~16B while writing) with aggressive KV-cache compression that pushes API prices far below Western frontier models. DeepSeek is routing flagship V4 Pro traffic to it. Independent coding benchmarks put it roughly on par with GPT-5.6 Sol — Matthew Berman covered the launch the same week.