If your cost tool can't tell you its tokenizer, it's guessing. This week: 1. Log token fields from API responses, not estimates 2. Ask your cost tool which tokenizer it uses 3. Distrust estimates for…

The moat: proprietary data — your rants, your numbers, your benchmarks. You can't scrape experiments back out of someone. We measured ours: • two tokenizers, same text: 20% apart • DeepSeek cache:…

The tools you use for this are 6 months behind the models. Most "cost dashboards" estimate instead of reading the API response. Some don't even show cache hit rate. You're flying on guesses.

That post got SPECIFIC questions, not "how much". Pet hair. Carpet. Consumable costs. Map algorithms. They answered with solutions only. Admitted what they couldn't do. Users asked: "which one did…

Why it compounds: the 30x cache lever works on cache hit tokens. Measure the input side wrong by 20%, your cache hit rate is wrong by 20%, and every optimization inherits the error. You can't manage…

What I changed after 30 days: 1. Frozen system prompt 2. Cron jobs off-peak (peak is 2x) 3. Retries capped at 2 with backoff 4. Stopped estimating tokens - real counts are in API responses, free to…

What actually cuts the bill: 1. Keep prompts stable, cache survives 2. Batch jobs off-peak, prices double in peak 3. Cap retries, storms triple the bill 4. Watch output tokens, the most expensive row