Fabian de la Fuente

Weekly · Vancouver, BC

Two Days Without Credits Cost Me USD 132.55: What I Learned About Tokens, Plans, and Overage

When the weekly credit pool ran out, two days of reduced work cost me USD 132.55 on-demand. What went wrong, what we changed, and why I design for an empty tank.

Fabián de la Fuente ·

Norman Rockwell–style illustration: a founder in a cardigan at a wooden desk stares at a glowing billing meter and an empty glass jar labeled EMPTY, while three boxy robot assistants in neckties pause beside an hourglass in warm morning light.

On a Thursday at 5:30 in the morning, Vancouver time, the weekly credit pool resets. For the rest of the week, if it runs out, you pay list price. I knew that. I still designed the team as if the gauge never hit zero.

That week, it hit zero. With tasks already dialed back, two days without included credits cost me USD 132.55 in on-demand usage: about CAD 186 at the October 22 exchange rate (USD/CAD 1.4048). It wasn't a bug. It was an invoice.

Operational Friction

Running bots doesn't feel like a fixed expense. It feels like an office where the lights never turn off. Every routine, every email review, every browser capture burns tokens. As long as the weekly pool holds out, the pain is invisible. The day it runs out, every token comes at list price, and the month carries a cap for extra usage.

On October 27 at 4:31 AM PT, the Usage & Billing screenshot said it plain: included weekly limit at 69%; remaining additional credits USD 27.47; month's additional usage USD 132.55 out of a USD 200 monthly cap (USD 67.45 margin); automatic USD 40 top-up when the balance drops below USD 8.

That was the gauge. Not a metaphor.

The Hard Decision

The plans are clear. According to the public pricing table: Premium USD 30, Super USD 100, and Ultra USD 200 per month (4× / 15× / 50× the Plus plan; Plus costs USD 8 without Grok Bot). There is a weekly pool that resets Thursday ~5:30 AM PT. When it runs out, usage continues at list price, with a monthly cap on additional usage.

The decision wasn't "buy more and forget about it." It was to treat the included quota as a hard capacity, just like server bandwidth or disk space. If the tank can hit empty, the system has to keep running on empty—or shut down non-critical operations before getting there.

What Broke

The illusion that "there's always a little bit more" broke.

We had written nice rules and left routines running as if the Thursday reset were just an administrative detail. It isn't. It's the edge of the tank. When you cross that edge, the cost stops being abstract: every review, every open browser, every long translation turns into a line item on the billing report.

The idea that the most expensive model has to do everything broke, too. Translating a long blog post from Spanish to English in a Gemini test cost USD 0.0051 and took 4.2 seconds. Today we use Gemini only on the free tier: if the free quota runs out (exit code 2), the work waits until tomorrow. There is no backup paid switch. A cheap—or free—intern doing the heavy lifting on text is worth more than an expensive executive doing the same work out of inertia.

What We Actually Changed

Three measures, all real:

  1. We lowered routine frequencies. Immediate alerts are handled by "timbres" (watchdogs that don't burn the large model). WhatsApp and COO inbox reviews were backed off to every 3 hours on business days, once at night during the week, and every 4 hours on weekends. Other signatures and backups were spaced out, too. Fewer trips to the gauge.

  2. We paused or deferred heavy lifting when credits got tight. If the weekly pool is running low, we don't invent work "just because the bot can." We wait for the reset or send heavy text to a cheap secondary AI.

  3. We prefer CLIs, scripts, and connectors over opening Chrome. The browser burns more quota. If a command or an API can fetch the same data, we don't open a tab "just in case."

None of those three are glamorous. All three show up on the next month's report.

The Principle: Design for the Empty Tank

If I had to leave behind a single rule, it would be this: treat the included quota as a hard limit and design for the empty tank.

Don't expect the gauge never to hit zero. Set up timbres for what's urgent, CLIs for the routine, a cheap secondary AI for heavy text, and an explicit pause for anything that can wait until Thursday. The speed of an AI team isn't measured solely in tokens per minute. It's measured by how much it can get done when the counter reads zero... without turning you into the involuntary sponsor of your own inertia.

The USD 132.55 didn't teach me to hate tokens. It taught me to respect the gauge.

← Blog