GPT-6.1 Terra arrived at half the price of Sol and was on Bedrock, Copilot and Codex within days, while Copilot Fusion showed that routing per task can roughly halve the cost of each task. The quieter items matter as much: Cloudflare found GLM-5.3 responses carrying other customers' data, Gemini API now stores requests by default for new paid projects, and OpenAI's Model Spec spelled out how far an agent may act on its own. This week I'd add per-task routing, set store=false explicitly, check KV pool sizes on self-hosted engines, and clear Opus 4.8 fast mode out of configs before November 30.
Why it made the cutIt reached Bedrock, Copilot and Codex within four days, so the cheaper tier is already within most teams' reach, sometimes unasked.
Read the analysis →Why it made the cutNo error, just data quietly crossing tenants; anyone self-hosting open models should check KV pool size against the advertised context.
Read the analysis →Why it made the cutScoring 60 at about half the cost per task makes the case that routing per task beats fixing one model.
Read the analysis →Why it made the cutA silent compliance change: new paid projects keep requests by default, so turn it off in code, not just the console.
Read the analysis →Why it made the cutIts agent sections read as a design checklist, and the same week Copilot and Amazon Quick shipped model-screened approvals along those lines.
Read the analysis →Why it made the cutA hard deadline with no fallback: after November 30, Opus 4.8 with fast mode just errors, so search configs now.
Read the analysis →