Codex workflows on Amazon Bedrock with GPT‑5.6 Sol can see major cost spikes when explicit prompt cache controls aren’t sent. Cache writes cost 1.25×, while reads get a 90% discount—so missing cache breakpoints turns caching into churn instead of reuse.
A batched compact-Householder QR kernel is tricky on GPUs because Householder reflectors form a dependency chain. This post explains how blocked Householder QR with compact WY updates turns the trailing work into GEMM-shaped operations, and how an auto-research loop with Codex drove fast, correct kernel iterations to a 232× speedup over baseline.
Codex’s MultiAgentV2 started encrypting sub-agent `message` payloads, storing only ciphertext (`encrypted_content`) and leaving plaintext empty (`content`). That improves privacy, but can erase the readable delegation details in rollout/history, making audits and debugging harder.
“Ultra will be in Codex” suggests the coding assistant will use a higher-tier “Sol Ultra” model for code generation. Expect better reasoning, fewer constraint misses, and potentially improved handling of long project context.