Salesforce launches Koa, a business reasoning model built on Nvidia Nemotron 3 Super for long-running AI agents, alongside AIforce, Claudeforce’s 37 pre-built sales skills, and Headless 360 API access at Dreamforce 2026.
CADDi raises a $114 million Series D at a $1.2 billion valuation to accelerate its North American push and build manufacturing-specific AI models, with Salesforce Ventures, Woven Capital and Moore Strategic Ventures joining the round.
Google launches Gemini 3.8 Live and Extended Thinking, voice AI models that switch automatically among 97 languages and can run background API calls during conversations, with rollout spanning the Gemini API, Search Live and Workspace apps including Gmail and Docs.
Salesforce launches Koa, its first CRM reasoning model for Agentforce, built on NVIDIA Nemotron and slated for U.S. general availability in Winter 2026 after training on proprietary synthetic CRM scenarios across more than 14 industries.
Ex-OpenAI researcher Diogo Almeida unveils TypeSafe’s Jev, an AI for pre-defined software decisions that claims zero hallucinations by selecting only from preset options and returning confidence scores.
Apple’s iOS 27 and macOS Golden Gate code indicates Siri can hand requests to third-party AI models such as Claude and potentially use GPT-5.6 for its server-side intelligence, though Apple currently exposes the “Ask...” interface only to ChatGPT.
ElevenMusic launches Music v2.5 as its default AI music model, promising richer melodies and more natural instruments while giving free users five lossless downloads daily and Pro users 400 monthly.
Colibrì now runs MoE models as large as Kimi K3’s 2.8 trillion parameters on machines with just 25 GB of memory by streaming routed experts from disk, though cold GLM-5.2 inference falls to 0.05–0.1 tokens per second.
AWS says GPT-5.6 Luna now delivers Bedrock’s lowest observed cost per successful outcome after an 80% July 30 price cut, costing $0.0021 per correct AIME answer versus GPT-5.4-mini’s $0.0139.
OmniVoice now generates 24 kHz speech across 600-plus languages from 3–10-second voice samples, though cross-lingual clones retain the accent of the reference speaker’s language.