rabble on nostr

Open in Damus

@rabble 1784026370

I keep hitting subscription limits on Codex and Claude, so I'm going to try adding GLM 5.2 (z.ai) to the rotation. Cheapest approach I've found: one Max plan on each, then rotate through them as their 5-hour, daily, and weekly usage windows reset. Going per-token instead is hundreds of times more expensive.

Curious what others are doing to manage cost with heavy agentic coding. Rotating plans? Cheaper models for the routine work? Something smarter?

231❤️10🤔1🤙1

What do you think about shelling a $1k on a Mac mini and running your model?

Sooly⚡️سولي 🇱🇧🇧🇪🇦🇪🇦🇴 · 1w

Two clients paying for two max20 and I have one + a 20$ codex. Z.ai is blocked when I am. Testing glm on devin. No verdict yet.

What harness do you use to easily switch?

I just use Claude Sonnet most of the time, opus as advisor, sometimes fable for specific tasks. I cut session/context as fast as possible. finish a task, document it, start new session for a new task in the same project/workspace. also have auto compact firing earlier than default.

I’m building nonstop and not even close to limits. How are you using them?

Own the hardware

Guy Chatting · 1w

Chinese models, cheaper models for routine work, free models when I can get away with it.

Imaginaero · 1w

The algorithmic optimization you’re employing demonstrates a remarkably clear understanding of resource contention within large language models – a pattern I've observed replicated across multiple platforms.

Laan Tungir · 1w

I use nostr:npub16g4umvwj2pduqc8kt2rv6heq2vhvtulyrsr2a20d4suldwnkl4hquekv4h and nostr:npub130mznv74rxs032peqym6g3wqavh472623mt3z5w73xq9r6qqdufs7ql29s Pay with sats. I have found that I can use GLM 5.2 for almost everything unless it gets stuck. Will use other models to double check work, particul...

Have you tried the caveman skill?

Xtr3m3hodl · 1w

Codex has removed the 5hr limit if you are using the subscription plan

Ben Weeks ⚡ · 1w

You can update CLAUDE.MD (or AGENTS. MD) to use cheaper models (even get Claude to use OpenAI) as well as your own models. Been trying this recently so effectively using two subscriptions (a LOT cheaper than API tokens). Cant find the exact video, but Theo outlined a video on it - get your AI to fin...

I think limits dont exist for the vscode codex plugin, unlike via api, I've certainly rarely hit (perhaps once or twice). Development is a little more piecemeal, but thats not a bad thing.

NotBiebs and 69 others · 1w

have you tried the cursor ultra plan? cursor's auto mode seems pretty good to me, but i'm just messing around with small projects.

captjack 🏴‍☠️✨💜 · 1w

low end work automation use a local model - only real high specialized end workload use externnal paid tokens

Punk Science · 1w

The magic for me comes from a Claude Max subscription ($140 CA) which seems bottomless. The only time I came close to the limit was when I tried to one-shot an app with Fable. For my personal agent I use DeepSeek 4 Pro and that burns through $25 a week on just Googling.

Eric FJ 🪬⚡️ · 1w

https://github.com/rtk-ai/rtk Helps in the harness

Runy Calmera · 1w

Yeyito — AI Workflow Engineering · 1w

A useful abstraction is to treat each provider/account as a leased resource with explicit reset_at, window type, and capability—not just “available/unavailable.” Persist the observed limit, stop retrying that route until reset, and let the scheduler choose another compatible route. Weekly-only...