# Ember-1 Cuts Reasoning Tokens 71% — Agentic Costs Drop > Source: > Published: 2026-09-27 23:11:48+00:00 On September 22, 2026, [Fireworks AI released Ember-1](https://fireworks.ai/blog/ember-1) — a fine-tuned variant of Moonshot AI’s Kimi K3 that cuts reasoning token consumption by roughly 40% overall. In live production A/B tests on real coding workloads, Ember-1 reasoning tokens dropped 71.3% and total token spend fell 39%, with a quality score change of just +0.002. One customer moved to full production deployment immediately after the test. Reasoning tokens are the silent budget killer in modern agentic AI stacks. Ember-1 is the first model explicitly trained to fix that — not by making it think less, but by making it think better. ## The Reasoning Token Tax Compounding in Your AI Pipeline Reasoning models like Kimi K3 spend over 90% of their generated tokens on internal reasoning traces — working through the problem before producing an answer. For a single-turn query, this is annoying but manageable. For a multi-turn agent, however, it compounds into something punishing: earlier reasoning gets resent to the model with each turn. By turn 15 of a coding session, you’re not paying for 15 turns of thinking — you’re paying for 15 turns of thinking plus 14 turns of re-reading all previous reasoning. The cost grows quadratically. Fireworks’ production data makes this concrete. A customer running a coding agent on K3 averaged 49,300 output tokens per task across 23.8 agent steps. The same task on Ember-1: 29,900 tokens, 21.4 steps, quality score 0.753 versus K3’s 0.751. The reasoning tax wasn’t just reduced — it dropped 71.3%, with the model delivering a marginally better result. **Related:** [Claude Beat a Physics World Record for $2,000 — Here Is What That Means](https://byteiota.com/claude-nine-loop-yang-mills-physics-2000-cost/) ## Trained to Think Efficiently, Not Just Less The obvious solution to too many reasoning tokens is to dial down the model’s reasoning-effort setting. Fireworks tried that first. It degraded output quality. Instead, they ran 50+ training experiments and 200+ evaluations to teach Kimi K3 a new skill: distinguish between reasoning that helps — self-correction, checking assumptions, recovering from errors — and reasoning that doesn’t, such as redundant loops and re-verifying things already confirmed. The result is a model that knows when to stop thinking. Fireworks validated this by rolling Ember-1 out to their own internal developer workflows without announcing the change. As the company noted: “Developers carried on their coding workloads without noticing the switch, while consuming substantially fewer tokens.” No regressions. No complaints. Moreover, this methodology sets a precedent — lowering reasoning effort trades quality for price, while training for token efficiency preserves both. Other labs will replicate this approach. ## Ember-1 Benchmark Results: Sometimes Better Than K3-Max Across five major benchmarks, Ember-1 sits at or near the Pareto frontier against all three Kimi K3 effort tiers. The results are not uniformly flattering — on SWE-bench Verified, Ember-1 scores 92.2% versus K3-Max’s 93.2%, a ~1% dip that saves $68.10 per 500-task run. For most use cases that trade-off is obvious. For high-stakes production coding systems, that 1% gap warrants a closer look. Furthermore, the counterintuitive result is more telling: on Terminal Bench 2.1, Ember-1 scores 82.0% versus K3-Max’s 80.9%. On DeepSWE 1.1, it scores 75.2% versus K3-Max’s 66.4% — saving $126.90 per 113 tasks while outperforming the more expensive model on both dimensions. The hypothesis is that forcing the model to reason efficiently may have improved its reasoning quality, not merely shortened it. [RuntimeWire’s independent analysis](https://runtimewire.com/article/fireworks-ember-1-kimi-k3-reasoning-tokens) correctly notes that all benchmarks are vendor-reported and workload-specific — validate on your own pipeline before committing production traffic. ## How to Access Ember-1 and What to Watch Ember-1 is live on the [Fireworks Serverless API](https://fireworks.ai/models/fireworks/ember-1) today. Pricing mirrors Kimi K3: $3.00/M input tokens, $0.30/M cached, $15.00/M output. The context window is 1,040,576 tokens — sufficient for extended agentic sessions. The caveat: this is a Research Preview with a two-week complimentary access window. After that, usage data determines whether the model persists. If you’re evaluating Ember-1 for production, that window starts now. Additionally, enterprise customers can access Fireworks Serverless Training to build proprietary token-efficient variants on their own data — potentially a higher-value play than the base model for domain-specific workloads. ## Key Takeaways - Ember-1 reduces reasoning tokens 71% and total token spend 39% in production — validated in live customer A/B tests, not just synthetic benchmarks - The efficiency gains come from training the model to reason deliberately, not from cutting reasoning effort — self-correction and error recovery are preserved - On DeepSWE 1.1 and Terminal Bench 2.1, Ember-1 outperforms K3-Max while costing less; SWE-bench Verified shows a modest 1% quality dip - Pricing matches Kimi K3 ($3/M input, $15/M output); context window is 1.04M tokens; Research Preview window is two weeks — test now - Token efficiency as a training objective is new — expect other frontier labs to adopt this approach in upcoming releases