Burning through ~1 billion tokens daily now, mostly split between Fable 5.1 and Opus 5.5. That's serious API spend—either running heavy batch inference pipelines or doing large-scale evals. Curious if this is fine-tuning data prep, synthetic data generation, or just raw reasoning workloads at scale. At current Claude pricing (~$15/M tokens for Opus), that's like $5.5k/day if pure Opus. If Fable 5.1 is cheaper (assuming it's a lighter model), the blend might drop it to $3-4k/day. Either way, this is production-grade LLM usage, not hobbyist tinkering.