Models
Frontier and open models: what shipped, what it can do, and how the claims hold up.
-
South-East Asia builds sovereign AI on NVIDIA’s open Nemotron models
Singapore, Malaysia, Vietnam and Thailand are fine-tuning NVIDIA’s open weights for local languages rather than training from scratch.
-
Anthropic prices Opus 5.5 down, and prints the benchmarks it loses
Opus 5.5 costs 40% less to run than Opus 5, and Anthropic’s own table shows GPT-6 Astra ahead on two of its…
-
Google ships Gemini 3.8 Live for real-time voice agents
Google’s new Gemini 3.8 Live audio models complete 35.1% of Sierra’s τ-Voice-banking tasks, the figure that says most about where voice agents…
-
Moonshot’s Kimi K3 arrives on Amazon Bedrock
The open-weight 2.8-trillion-parameter model is now available behind AWS procurement terms, and is the first open model on Bedrock with explicit prompt…
-
xAI claims halved transcription errors at the same price
Grok Voice Transcribe 2.0 keeps version 1.0 pricing and includes speaker diarisation, though most of the accuracy evidence is xAI’s own.
-
OpenAI launches Astra for Law, with its own benchmark numbers
OpenAI’s new legal configuration of GPT-6 Astra passed 54.0% of a research benchmark it ran itself, against 38.7% for the same model…
-
A Google network learns to spot cats unsupervised
A Google Brain network trained on ten million YouTube stills learns a cat detector without labels.
-
LeCun trains a network to read zip codes
Yann LeCun applies backpropagation to a convolutional network reading US postal codes.
-
MYCIN diagnoses infections from rules
Shortliffe’s MYCIN uses hundreds of rules with confidence scores to recommend antibiotics.
-
The Neocognitron introduces convolution
Kunihiko Fukushima publishes a layered network using convolution and pooling for pattern recognition.