TL;DR

Claude Max is an ~9× subsidized token buffet that is already tightening via daily routine caps, peak-hour burn, and uptime degradation.

  • Hard caps at 2:36 AM — Claude Max now pauses scheduled routines after 15 daily runs.
  • The 9× subsidy math — A 5-city tabiji content batch cost $137.39 on the raw API (~$5,500 for 224 pages), versus ~$600/month across three Max plans.
  • Dual-stack hedge — Recurring crons run on OpenClaw routing across 21 models (Z.AI/GLM, MiniMax, Codex, Gemini); Claude Code stays for interactive one-offs.

The 7 AM routine pause

I checked my phone over coffee and found a Claude notification from 2:36 AM:

Email from no-reply-claude at 2:36 AM titled 'Your routines are paused — daily limit reached'. Body reads: You've hit today's limit of 15 routine runs, so scam-narrative-rewrite-batch and any other routines are paused for now. An admin...
The Max plan stops at 15 routine runs per day. We hit it during sleep.

I pay for three Claude Max plans, one Codex Pro plan, plus MiniMax and Z.AI/GLM. And the job that got killed — scam-narrative-rewrite-batch — was doing the exact overnight, non-interactive batch work that a flat-rate Max plan is supposed to unlock.

Hitting a hard 15-run daily ceiling at 2:36 AM was the clearest signal yet that the flat-rate subsidy is closing.

The math the buffet hides

Last week I ran a five-city batch through the tabiji.ai content pipeline directly against the metered Anthropic API:

Cost breakdown table for 5 cities (Abu Dhabi, Acapulco, Accra, Addis Ababa, Agadir) showing per-PR cost ranging from $20.25 to $37.51, total $137.39, total API time 2 hours 14 minutes wall. Steady-state cost lands around $20-25 per city once cache warms up. Full 224-page drain at that rate equals approximately $5,500 versus my early $7,400 estimate.
Five cities, $137 in API tokens, ~2.25 hours of compute. The full 224-page run extrapolates to ~$5,500.

$137.39 for five cities — roughly $25 per city once the prompt cache warms up, or ~$5,500 for the full 224-page queue.

Compare that to my flat-rate bill: ~$600/month for three Claude Max plans plus ~$200 for Codex Pro. For ~$800/month, I routinely burn more tokens than that $5,500 API batch. Anthropic is subsidizing my workload by roughly 9×. They lose money on every heavy operator. The only way that unit math closes is if the median subscriber barely touches the terminal — or if Anthropic reprices, throttles, and caps.

OpenAI faces the same physics on Codex Pro. Inference economics are unforgiving: tokens cost GPU-hours, and a $200 subscription buys a finite slice of compute no matter what the pricing page promises. The buffet is real, and it is temporary.

And it’s already shrinking

You don’t have to guess when the squeeze starts. Look at Anthropic’s 90-day status page:

Anthropic 90-day status page showing All Systems Operational header. Below: claude.ai 98.75% uptime, Claude Console (platform.claude.com) 99.09%, Claude API (api.anthropic.com) 99.05%, Claude Code 99.22%. Each row shows daily uptime bars over 90 days with significant numbers of yellow and red incident days.
Anthropic’s own 90-day status page. The headline numbers sound great until you do the arithmetic on them.

The banner says “All Systems Operational,” but the bars underneath tell the real story. Claude Code at 99.22% uptime over 90 days equals 16 hours of degraded or downed service per quarter. Claude.ai at 98.75% equals 27 hours. Every yellow or red bar is a window where unattended crons failed, hit rate-limit walls, or timed out while still burning daily routine quota.

When demand outstrips GPU supply, a provider only has three levers: price, quota caps (the 15-run wall), or service degradation. All three are moving at once.

Then there’s the undocumented fourth lever: peak-hour quota burn. The exact same batch job that runs cheaply at 2 AM consumes noticeably more of your daily allowance during US business hours. There’s no surge-pricing warning on the dashboard, but if you log per-task quota drain on heavy crons, you see your meter burn ~30% faster at noon than at midnight.

Layer on model-deprecation risk — where a provider can ban a harness or retire a model weight on 60 days’ notice — and single-vendor cron infrastructure becomes a liability.

21 models, 5 providers, one Mac mini

Here is the alternative. We run tabiji’s recurring infrastructure on OpenClaw, an open-source agentic harness that routes across whatever model providers we wire into it. Here is the live model config our agent Psy posted to Slack:

Slack message from Psy listing the OpenClaw model configuration: 21 models across 5 providers. Defaults: primary z.ai/glm-5.1, fallback 1 minimax/MiniMax-M2.7, fallback 2 openai-codex/gpt-5.4. Sub-agent model glm-5.1, image gen google/gemini-3-pro-image-preview, video gen google/veo-3.1-lite-generate-preview. Z.AI 11 models including glm-5.1, glm-5.1-turbo variants and vision models. MiniMax 2 models. Xiaomi 3 models including mimo-v2-flash 1M context, mimo-v2-pro, all free. Google 3 models all 1M context vision. Anthropic 2 models: claude-opus-4-6 and claude-sonnet-4-6, plus codex/gpt-5.4 referenced as fallback.
Psy’s OpenClaw config: 21 models across Z.AI, MiniMax, Xiaomi, Google, and Anthropic. Defaults and fallbacks named explicitly.

Our routing table: Z.AI’s GLM-5.1 runs as primary, MiniMax-M2.7 sits as Fallback 1, and GPT-5.4 (via Codex) sits as Fallback 2. Xiaomi’s mimo-v2-flash and mimo-v2-pro handle 1M-context sweeps for free. Gemini 3 Pro and Veo 3.1 handle image and video generation. Anthropic’s claude-opus-4-6 and claude-sonnet-4-6 stay in the pool for tasks that need them — as a choice, not a dependency.

When GLM rate-limits, OpenClaw fails over to MiniMax. When MiniMax hiccups, it drops to GPT-5.4. The overnight crons finish without sending a 2:36 AM pause email.

Apple Mac mini sitting on a wooden desk with a power cable plugged in. Compact, silver, about the size of a paperback.
The actual hardware. ~$600 of Mac mini, sitting on a desk.

All of it runs on a $600 Mac mini sitting on my desk — a few YAML files and a router script dispatching jobs to whichever provider has clean capacity.

My dual-stack reality

I still pay for three Claude Code Max plans and use Claude Code every day. For interactive one-offs — architecture exploration, live debugging at 11 PM, audits, drafting posts — nothing beats it. The harness is fast and the tool use is sharp (even if you have to watch its shortcuts at scale).

Everything recurring lives on OpenClaw: scheduled content pipelines, nightly repo audits, multi-page sweeps. Any job that runs unattended at 3 AM cannot depend on a single vendor with a 15-run daily cap.

That’s the split: Claude Code for the 9 AM interactive loop, OpenClaw for the 3 AM cron.

The hedge thesis

Running OpenClaw isn’t a protest against Anthropic; their models are still best-in-class. It’s refusal to bet production infrastructure on a subsidized pricing tier holding forever.

A Max plan gives you one vendor and a hard ceiling. OpenClaw gives you 21 models and an automatic circuit breaker. When a provider raises prices, tightens caps, or deprecates a model, my migration cost is editing one line of YAML. Keep your recurring pipelines portable, use Claude Code while the subsidy lasts, and be ready to route around the buffet the day it closes.