OpenAI cut GPT-6 Sol and Luna API prices 50% versus GPT-5.6, while Anthropic lists Opus 5.5 at $4/$20 per million input/output tokens, 20% below Opus 5. The releases advance cost efficiency, but their benchmark claims use different test designs and are not directly comparable.
vLLM explains how its Mooncake connector streams hidden states between disaggregated inference and training across a 12-node GB300 NVL72 setup for Kimi K3. The full note supplies the topology, configuration, and scaling constraints needed to distinguish its experiment from a general deployment claim.
OpenAI’s implementation note explains the 30-minute cache-eligibility window, diagnostics for misses, explicit breakpoints, stable tool definitions, and prewarming. Those operational choices retain value beyond the headline price cut, although the reported cache-hit gains are workload-specific.
Artificial Analysis places Step 5 Preview at 44 on its Intelligence Index against a 24 median, with 92.1 output tokens per second and a $0.72 index-task cost. Its 160 million output tokens across the evaluation are an operational measurement, not evidence of real-world utility.
64 source items · 76% editorial confidence
BLS projects total employment up 5.9 million, or 3.5%, between 2025 and 2035. It expects computer and mathematical work to grow 7.3%, while office and administrative support falls 4.0%, or 752,100 jobs; these are projections rather than causal estimates.
319 source items · 64% editorial confidence
loading form …