One for all or all for one?

TL;DR: Most developers drifted into depending on one AI platform without making an active choice. Three incidents from the past year show what that costs when something changes. Model quality at the top has converged. Pricing has not. Worth thinking about before you have to.

Here is a fun little thing that happened to us.

We looked at our setup one day and realized Claude was running basically everything. Not because we planned it that way. Not because we sat down and said "yes, one vendor, let's go." It just... accumulated. CLAUDE.md in every project, prompts tuned to one model, muscle memory doing the rest. The stack had opinions and we had stopped having them.

Creative people building on autopilot. Not our best look.

The industry has had a rough year that makes this worth talking about.

Anthropic cut off Claude subscriptions in third-party tools like OpenCode earlier this year. Most developers had no fallback ready. Not because building one is hard. Because nobody had thought about it.

Last summer Cursor swapped 500 requests a month for $20 of usage. Same $20, roughly half the Claude requests. Badly explained. The CEO apologized and offered refunds. People left.

Lovable, $6.6 billion valuation, employees at Nvidia and Microsoft reportedly among its users, left other people's project source code readable by any logged-in account, along with any keys hardcoded into it. A researcher reported it. It sat for 48 days. Bug reports filed. Bug reports closed. When it came out their first response was that the exposure was intentional behavior.

Reader, it was not intentional behavior.

If you had keys sitting in one of those projects, the sane move was to rotate them and assume they had been seen.

These are not uniquely bad companies. They are fast-moving ones. And fast-moving platforms make fast-moving decisions that your workflow did not account for.

We use Claude daily, Codex for heavier agentic work (AGENTS.md pulling the same weight as CLAUDE.md on the Anthropic side), Gemini when the task calls for it. OpenCode as the connective tissue since it talks to 75+ providers and lets you swap models mid-session without losing your thread. Messy to set up once. Extremely not messy when something changes.

The whole point of creative thinking is that you stay in the driver's seat. That goes for AI tooling too.

What does your stack look like right now? And be honest, was it a choice or did it just happen to you?