Daily digest — 23 Sept 2026

8 verified AI/ML items, ranked by impact. Five minutes, every weekday. Email, WhatsApp and Telegram delivery start in Phase 0; RSS works today.

1. Claude Opus 5.5 ships: Opus-tier quality at a 20% lower price
   Teams on Claude Opus 5 get a cheaper successor with the same 1M context and 128K output. Four breaking API changes (thinking can't be disabled, lower default effort, no forced tool_choice, preserved thinking) mean this is a code change, not a model-string swap.
2. Claude Opus 5.5 reaches AWS Bedrock in Mumbai (ap-south-1), preview
   Indian teams with data-residency requirements can trial Opus 5.5 in-region the day after launch. Preview means no SLA and no batch inference yet, so keep production on Opus 5 in ap-south-1 for now.
3. Colibrì runs the 744B GLM-5.2 on 16–25 GB of RAM with no GPU, slowly
   A tiny open-source C engine streams a frontier-size MoE model's experts from disk, so it runs offline with no GPU and no per-token bill. Needs ~372 GB of NVMe and runs at about 1 token every 10–20 s on a 25 GB machine: great for private batch jobs, not for chat.
4. Claude API: per-message effort control (beta) without a prompt-cache reset
   Agent loops can drop to low effort for routine turns and raise it for hard ones mid-conversation while keeping the cached prefix. Useful cost lever for long-running agents.
5. LangGraph (demo release) renames checkpointer interfaces — breaking for custom savers
   If you wrote a custom checkpointer (for example on pgvector/Postgres), it will not load after upgrading. Pin the previous minor version until you port the interface.
6. Llama open-weight 2 70B (demo) lands on Bedrock in preview
   A cheap open-weight option for high-volume classification on Bedrock. No structured-output support yet, so extraction pipelines need a validator.
7. Paper with code (demo): late-interaction reranking cuts RAG hallucinations on long PDFs
   A runnable reranker that drops into existing retrieval pipelines. Worth a 30-minute test if your RAG answers cite the wrong page; our RAG suite run is scheduled.
8. Rumor: OpenAI Frontier 3 (demo) in limited testing Rumor
   Nothing to do yet. We track it so you are not surprised; no facts will be added until a primary source confirms.