OpenAI's new /v1/decisions endpoint, in beta on gpt-6-luna, returns probabilities, choices and scores instead of text. It charges $0.10 per million input tokens and nothing for output.
RTX Spark laptop preorders opened October 7 with two N1X configurations, 20-core/6,144-core and 18-core/5,120-core, both rated 45-80 W. NVIDIA publishes the CPU-GPU link speed but not the memory bandwidth.
Microsoft Execution Containers shipped SDK v1.0.0 for Rust, .NET and Node on October 7. A day earlier, the README lost its warning that MXC policies could be overly permissive.
Cloudflare's open-weight Clef and Clef-flash decision models beat Typesafe's Jev on most benchmarks and run up to 13x faster. The exception is When2Call, the benchmark closest to the 'should I call this tool?' gate Cloudflare is selling them for.
Fable 5.1 held its price, cut cache reads to $0.25, and broke three requests that worked on Fable 5.
Anthropic's J-space paper, and the ablation that separates what Claude can report from what Claude can do.
Anthropic's red team on how agent swarms fail — conformity, collusion, sabotage, and the coordination that intelligence doesn't buy.
