I’ve been saying for the last year we’re awash in cheap AI tools and the free lunch is not going to last. They’re losing money on every unit and trying to make it up on volume. The math doesn’t math.

https://arstechnica.com/ai/2026/06/anthropic-pauses-token-based-billing-for-its-claude-agent-sdk/

Anthropic "pauses" token-based billing for its Claude Agent SDK

Move originally planned for Monday would have heavily increased power users' costs.

Ars Technica

Fwiw if you’re willing to use open tools and providers, you can avoid vendor lock-in and save a ton of money. The integration isn’t as smooth, but it’s more resilient in the long(er) term.

No idea where this all goes 6-12 months from now but I’m certain tying tooling & process to a single vendor’s service and technology stack will be the most pain in the end.

I consumed around 459M tokens in the month of May. Total AI spend: US$10.

I use #OpenCode + whatever models & providers are cheap at the moment. A side effect is all of my tooling is much more resilient to switching up providers & models at any moment, even mid-task.

One last thing. Whatever models “are cheap at the moment” doesn’t mean crappy. I use a ton of deepseek v4 pro & glm 5/5.1. All three of those are powerhouses and rival opus at a fraction of the cost. Deepseek v4 flash is roughly sonnet.
@zcutlip Yeah only reason I’m not doing a lot more of that is that I’m not paying for it. But by the end of the year I’m sure the subscriptions will be dead or nerfed to hell so for any hobbyist use it’ll be impossible not to get good at using cheaper models.

@g Man, I use Claude Code + opus/sonnet at work, and it’s so velvety smooth. I have basically no guardrails set up and it just does what I say. The trade-offs are real.

Plus Claude Code and the Anthropic models are well tuned together. You can ask Claude to configure itself or write a skill and it just nails it.

So no judgement for anyone using Claude.