Fireworks Ember-1: Pricing, Benchmarks and Setup

Fireworks Ember-1 is a research model announced 23 September 2026, built on Kimi K3. Fireworks says it delivers Kimi K3 quality with about 40% fewer tokens by cutting reasoning that the task does not need. The announcement links fireworks.ai/models/fireworks/ember-1. It does not print a separate list price or a RouterPlex model id.

Ember-1 is not on RouterPlex yet. Register for the current 60+ models. Kimi K3 is already live as kimi-k3 at $3 input and $15 output per 1M tokens. Use that route today. Use Ember-1 here once the marketplace lists it, at vendor list price, 0% markup, prepaid, with a hard spend limit.

There is no separate Ember price on the announcement #

Fireworks computed per-benchmark cost with public Kimi K3 pricing: uncached input $3 per 1M, cached input $0.30 per 1M, output $15 per 1M. RouterPlex sells kimi-k3 at $3 / $15 and does not publish that $0.30 cache-read rate. Do not treat $0.30 as a RouterPlex bill.

A 40,000-in / 2,000-out turn with no cache is $0.12 + $0.03 = $0.15 on the K3 sticker, whether you are on Fireworks' public rates or on this catalog. Ember-1's savings, if the vendor's token claim holds on your trace, show up as fewer tokens, not as a lower sticker this post can cite.

Fireworks describes a research preview on serverless: two weeks of access, then a decision to keep the model if demand is there. That is their rollout note. It is not a RouterPlex commitment.

Benchmarks versus Kimi K3 #

Fireworks' table. Columns are K3 at low, high, and max effort, then Ember-1. The last column is printed as Ember versus K3 max: a token or cost change, not a score delta. N is the sample count they printed.

Read the score and the cost column separately. On SWE-bench Verified, Ember-1 is 1.0 point under K3 max and Fireworks still prints a lower bill. On DeepSWE, Ember-1 is the higher score. Fireworks says that across benchmarks with more than 50 samples, Ember-1 sits on or near the Pareto frontier versus those K3 effort settings.

Training notes from the same post: 50+ training experiments, 200+ evals, Fireworks' own data, no customer data. Live A/B tests with two customers, in Fireworks' prose, showed about 35% fewer tokens per task at comparable quality, and one customer moved Ember-1 to production. See the Fireworks post for the per-customer breakdown.

Bedside Bench chart #

The chart above is Fireworks' Bedside Bench figure, Doximity, 500 clinical cases in the post. Axes are score (avg@3) and cost per task in USD. Labeled points include Ember-1, Kimi K3, GLM 5.3, Gemini 3.8 Flash, GPT-5.6 Sol, GPT-6 Astra, and Claude Opus 5. The blog's claim is a new Pareto frontier against GPT-5.6 Sol, GPT-6 Astra, and Claude Opus 5 on cost per task. That is a vendor claim. Fireworks does not print exact values for each point, so read the chart for position, not precise scores.

Call Kimi K3 today #

kimi-k3 is the live base model on RouterPlex.

curl https://api.routerplex.com/v1/chat/completions \
  -H "Authorization: Bearer $ROUTERPLEX_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "kimi-k3",
    "messages": [
      {"role": "user", "content": "Summarize the tradeoff between a shorter reasoning trace and a wrong patch."}
    ]
  }'

That request bills $3 / $15 per 1M on this catalog. It is Kimi K3, not Ember-1. The catalog and the OpenAI-compatible API cover the models that are actually listed.

On RouterPlex, once Ember-1 is listed #

We will name the route when the row exists. We will not guess a Fireworks account id in a curl example. Price will be the vendor list price, 0% markup, prepaid, hard per-key budget, same as the rest of the marketplace. Create a RouterPlex account for the 60+ models live now, and for Ember-1 when it is one of them.