Labs squeeze third parties while absorbing their features

Anthropic stopped letting Claude Code subscriptions power OpenClaw, citing "outsized strain" on its systems. It offered a one-time credit and discounts. Commenters called this a money grab: you pay for a token allowance, then get told using it is too much. One person says they hit rate limits within 1:45 on afternoons and can't justify variable-cost extra usage. Separately, OpenAI's Decisions API hit public beta, and the thread read like "we told you they'd build this," with someone asking how pricing compares to a startup called Jev.

The pattern here: the model labs control both the supply of tokens and the harness around them. They will close one or absorb the other when it suits them. One analysis of Claude Code's suggested-message feature argued the real customer is the model itself, because predictions of what you'd type train it. Even small UI features are read as strategy.

Builders are responding with hedges. Treg pitches itself as "OpenRouter for tools," and commenters immediately flagged that the underlying providers' terms don't allow that kind of resale. People expect agent-agnostic tools that run on any CLI to multiply.