STEVEN STEINBECK
← ALL DISPATCHESDISPATCH 65

WEEK OF MAR 21 – 27, 2026

Gemini goes realtime, open-weight voice lands, and the Pentagon fight reaches a courtroom

Realtime voice and vision agents got a production-tier model, speech models went open-weight, and a judge told the government its Anthropic ban looked like punishment.

Google shipped Gemini 3.1 Flash Live for realtime voice and vision agents, and Mistral and Cohere both open-sourced speech models on the same day — the audio stack stopped being a paid API dependency.

Meanwhile two industry storylines hardened: OpenAI killed Sora, and Anthropic confirmed a leaked frontier model it describes as a cybersecurity step change.

THU · Mar 26, 2026modelsrealtimevoice-agentsgoogleapi

Google ships Gemini 3.1 Flash Live, a realtime model built for voice and vision agents

Google released Gemini 3.1 Flash Live, its new realtime audio model, alongside a developer rollout of the Gemini Live API in Google AI Studio.

Google frames it as its highest-quality audio and speech model yet, with lower latency, better function calling, 2x longer conversation memory in Gemini Live, and SynthID watermarking of generated audio. The launch spans Gemini Live, Search Live, AI Studio preview and enterprise CX surfaces, with 70 languages and a 128k context window. Third-party benchmarking from Artificial Analysis highlighted the reasoning-against-latency tradeoff: 95.9% on Big Bench Audio at high thinking level with 2.98s time-to-first-audio, versus 70.5% at minimal with 0.96s TTFA.

Google ships Gemini 3.1 Flash Live, a realtime model built for voice and vision agents
storage.googleapis.com

WHY IT MATTERS

Directly relevant to anyone wiring voice or vision into an agent loop — this is the serving tier where latency and function-calling reliability matter more than raw benchmark scores.

Google — Build real-time conversational agents with Gemini 3.1 Flash LiveGoogle — Making audio AI more natural and reliable
THU · Mar 26, 2026open-weightsspeechinferenceservingvllm

Two open-weight speech models in one day: Mistral Voxtral TTS and Cohere Transcribe

Mistral released Voxtral TTS, a roughly 4B-parameter open-weight text-to-speech model aimed at production voice agents, with 9-language support, about 90ms time-to-first-audio, and human-preference results compared favorably to ElevenLabs.

Cohere launched Cohere Transcribe, its first audio model, under Apache 2.0, claiming the top English result on the Hugging Face Open ASR leaderboard at 5.42 WER across 14 languages. Cohere also upstreamed encoder-decoder serving work to vLLM — variable-length encoder batching and packed decoder attention — reported to give up to 2x throughput on speech workloads.

Two open-weight speech models in one day: Mistral Voxtral TTS and Cohere Transcribe
Mistral AI

WHY IT MATTERS

Frontier-quality speech is now something you can self-host, and the accompanying vLLM serving changes are the kind of inference-efficiency detail that actually shows up in your bill.

Mistral — Voxtral TTSCohere — Introducing Cohere Transcribe
THU · Mar 26, 2026anthropicfrontier-modelssecurityindustry

Anthropic confirms leaked 'Claude Mythos' model after a CMS misconfiguration exposed it

A configuration error in Anthropic's content management system let Fortune discover that the company is testing a new model, Claude Mythos; Anthropic confirmed the project, describing it internally as a step change in capabilities.

Reporting also surfaced a related security-focused model effort given early access to organizations to harden codebases against AI-driven exploits, news of which knocked cybersecurity stocks including CrowdStrike and Palo Alto Networks down more than 5%. The leak landed days after reports that OpenAI finished pretraining its next frontier model, codenamed Spud.

WHY IT MATTERS

Signals where frontier capability and the security narrative are heading — plus the reminder that a config error, not a hack, leaked a company's biggest roadmap item.

Fortune — Anthropic confirms Mythos after data leakSiliconANGLE
THU · Mar 26, 2026policyregulationlitigationanthropicgovernment

Federal judge pauses the Pentagon's 'supply chain risk' designation for Anthropic

A federal judge temporarily blocked the government from labeling Anthropic a supply-chain risk, an order set to take effect in seven days, in Anthropic's suit over the Pentagon's designation.

At the earlier injunction hearing the judge said the government's ban looked like punishment. The case is a direct test of how far national-security procurement authority can be used against an AI vendor whose terms of use constrain military applications.

WHY IT MATTERS

Policy risk is now a real line item for anyone building on a frontier API — procurement blacklists can change a vendor's viability faster than a model update.

NPR — Judge temporarily blocks Trump administration's Anthropic banNPR — Judge says government's Anthropic ban looks like punishment
TUE · Mar 24, 2026openaiproductindustryvideo

OpenAI shuts down Sora, its AI video generator and standalone app

OpenAI abruptly announced it was saying goodbye to Sora, the AI video generator it made publicly available in 2024 — roughly six months after launching the standalone app.

The shutdown was read as a strategic retreat from side projects in favor of core productivity and enterprise work, with reporting the same week noting OpenAI had also deprioritized other non-core efforts while it prepares the Spud frontier model.

WHY IT MATTERS

A useful data point on where a frontier lab is willing to spend compute and headcount — and on how fast a shipped consumer product can be killed when it isn't the core line.

The Guardian — OpenAI shutters AI video generator SoraTechCrunch
MON · Mar 23, 2026agentsharnesstoolingopen-sourcegit

Cline Kanban: an open-source board for running parallel coding agents in git worktrees

Cline launched Kanban, a free open-source local web app that orchestrates multiple CLI coding agents in parallel across isolated git worktrees, supporting Claude Code, Codex and Cline.

The board lets you chain task dependencies, review diffs and manage branches in one place, and the reception from builders was strong — several calling it a likely default multi-agent interface because it attacks the two practical bottlenecks of coding-agent workflows: inference-bound waiting and merge-conflict-heavy parallelism.

WHY IT MATTERS

Agent harnesses are becoming the differentiator, and this is a concrete, self-hostable answer to running many agents without hand-managing branches.

GitHub — cline/kanbanCline docs — Kanban overview