Claude's first API version in three years turns sampling parameters into 400s, and a poisoned LiteLLM release stole credentials from machines that merely had it installed. Meanwhile Agent Skills moved to a neutral foundation and OpenAI added a 12-hour cache and a second cloud. I'd spend this week pinning SDKs, auditing the gateway and dependency path, and keeping a provider abstraction so the next channel or version change stays out of business code.
Why it made the cutA genuinely breaking change: new SDKs default to it, so an unpinned dependency bump can take production down.
Read the analysis →Why it made the cutThe model gateway holds the most keys in the stack, and here just having it installed was enough to leak them.
Read the analysis →Why it made the cutWith MCP and Skills both in AAIF, writing prompts and runbooks as SKILL.md is now a portable, low-regret bet.
Read the analysis →Why it made the cutCache structure is becoming the main cost lever; this one pays off only if you do the reuse math first.
Read the analysis →Why it made the cutRead next to LiteLLM: coding agents install packages where secrets live, and this puts a checkpoint on that path.
Read the analysis →Why it made the cutOpenAI can now be bought on two major clouds, so channel choice becomes a procurement decision, not a code change.
Read the analysis →