@outsource_ I gave my agent your prompt info to replicate your claims on my 4090. My claude and GLM both say this is clickbait and there is no way you’re fitting qwen3.6-27b 4km and a smaller matching model for spec dev and have room for context
@realsigridjin Opus 4.7 uses a new tokenizer. 1M tokens ≈ 555k words, vs Opus 4.6's 1M tokens ≈ 750k words.
That means 4.7 consumes ~35% more tokens per word of text than 4.6. Same $5/$25 per MTok pricing, but
you'll hit your rate limits faster on 4.7 than on 4.6