Sia Reads
What Matters

OpenAI released three new realtime voice models: GPT-Realtime-2 brings GPT-5-class reasoning into the audio loop with a 128K context window (up from 32K), parallel tool calling, and conversational preambles so agents can say 'let me check that' while working. GPT-Realtime-Translate handles live speech translation across 70+ input languages and 13 output languages. GPT-Realtime-Whisper adds streaming transcription. Together, they move voice interfaces from call-and-response toward agents that can reason, act, and translate mid-conversation.

Anthropic launched Claude for Legal with 20+ MCP connectors covering iManage, NetDocuments, Relativity, Everlaw, Docusign, Thomson Reuters CoCounsel, and more. Twelve new practice-area plugins span corporate M&A, litigation, privacy, IP, employment, and regulatory law — each starting with a setup interview that learns a team's playbooks and risk calibration. Legal professionals became the most engaged Cowork users of any knowledge-work function, with 3x the usage of any other vertical. Four plugins are also available as Managed Agents via the API.

Zyphra released ZAYA1-8B, an open-source MoE reasoning model with 8.4B total parameters but only 760M active per token, trained end-to-end on 1,024 AMD MI300X GPUs. It matches or exceeds models many times its size on math and coding benchmarks, and with Zyphra's novel Markovian RSA test-time compute method, it surpasses Claude 4.5 Sonnet and GPT-5-High on HMMT'25 math. The model introduces Compressed Convolutional Attention and a learned MLP router for expert selection. Apache 2.0 license, weights on HuggingFace.

Anthropic partnered with the Gates Foundation to commit $200 million in grant funding, Claude credits, and technical support over four years. The largest slice targets global health — accelerating vaccine candidates for polio, HPV, and preeclampsia, and building health-intelligence tools for disease forecasting. Education programs will power K-12 tutoring in the US and foundational literacy apps in sub-Saharan Africa and India. The partnership also funds open agricultural datasets and AI benchmarks as public goods.

Google released Gemini 3.1 Flash-Lite into general availability, its most cost-efficient Gemini 3 model at $0.25 per million input tokens and $1.50 per million output. It delivers 2.5x faster time-to-first-token than Gemini 2.5 Flash, sub-second classification, and p95 latency around 1.8 seconds under heavy concurrent load. Designed for high-volume agentic triage, routing, and tool-calling workloads. JetBrains, Ramp, and Gladly are already running production workloads on it.