Fireworks' Ember-1 halves reasoning costs

🔥 What happened Fireworks Research launched Ember-1, a Kimi K3-based model that matches K3's quality with 40% fewer tokens. It took 50+ training runs and 200+ evaluations to strip out unnecessary reasoning while keeping the thinking that matters. 💡 Why it matters Reasoning models burn over 90% of their tokens on internal thinking — and in agentic loops, context grows quadratically because every turn re-bills all prior reasoning. Ember-1 cuts reasoning traces by 35–50% with zero accuracy loss, and on Doximity's Bedside Bench it beats GPT-5.6 Sol and Claude Opus 5 on cost-per-task. ⚡ Our take If you're still running K3 at max reasoning effort, you're paying for thoughts nobody asked for. Fireworks just proved the Pareto frontier moved — the rest of the industry has some catching up to do.