Ember-1: A/B Test Its 40% Token Savings in Python

Fireworks AI shipped Ember-1 this week and it went straight to the top of Hacker News (578 points). The pitch is narrow and useful: it is built on Kimi K3, it keeps K3's quality on coding and agent tasks, and it burns roughly 40% fewer tokens to get there. Fireworks reports about 35% fewer tokens per task in customer A/B tests.

Vendor numbers are a starting point, not a decision. Your prompts, your reasoning effort and your prompt-cache hit rate decide whether you actually save money. So instead of a review, this guide gives you a small Python harness that runs the same tasks against Ember-1 and a Kimi K3 baseline, counts billed tokens split into uncached input, cached input and output, prices them, and tells you the real percentage difference on your workload.