Last verified
BUILT ON KIMI K3SAME RATE AS K3SHORTER REASONING1M CONTEXTFIREWORKS ONLY

Ember-1 API Pricing

Ember-1 is Fireworks' own model: Kimi K3 retrained by Fireworks Research to spend fewer tokens thinking. It is priced exactly like Kimi K3 - $3.00/M input, $0.30/M cached, $15.00/M output - so the saving, if there is one, is in the token count rather than the rate. Fireworks reports 35-50% shorter reasoning at matching accuracy on its own tests. It runs only on Fireworks, with no open weights.

Input - per 1M tokens
$3.00/M
Kimi K3 $3.00/M at Moonshot = Kimi K3
Output - per 1M tokens
$15.00/M
Kimi K3 $15.00/M at Moonshot reasoning billed here
Cached input - per 1M tokens
$0.30/M
No cache-write rate published -90%
Effective - agentic blend
$1.92/M
92/8 split - 82% cache hit rate
§ 01 / TERMINAL

Run the numbers.

Calculator pre-loaded with Ember-1's rates, which are Kimi K3's rates. It prices tokens, not tasks: the case for Ember-1 is that the same job produces fewer output tokens, so compare the two coding scenarios below rather than the per-million figures.

$ /mo
Workload split
Prompt cache hit rate
Tokens you can process
—
Words equivalent (English)
—
Effective rate
—
Open full calculator (all models · share URL · CSV) →
§ 02 / SCENARIOS

Real-world presets.

CODING AGENT

Bug fix at Kimi K3's token count

$0.244/task
120,000 in - 10,000 out~409 units/$100
CODING AGENT

Same fix, 39% fewer output tokens

$0.186/task
120,000 in - 6,100 out~538 units/$100
LONG DOC

Reading 100-page contracts

$0.261/doc
180,000 in - 8,000 out~382 units/$100
ASSISTANT

Research planning turn

$0.044/turn
18,000 in - 2,000 out~2,265 units/$100
§ 03 / TOKENIZER

Paste text. See tokens. See cost.

What's a token →
Estimate · moonshot-tokenizer-estimate · ≈3.85 chars/token Auto-counts as you type

This is a chars-per-token approximation, not a real tokenizer. Actual tokens vary by language, code density, and tool-call overhead — counts are typically ±10–20% off for English prose, more for code or non-Latin scripts. For exact billing, use the vendor's official tokenizer.

Characters 338
Words 58
Tokens (estimated) 88 tokens
Cost as input · uncached $0.00026 USD
Cost as output · uncached $0.00132 USD
Cost as cached input $0.00003 USD
§ 04 / SHELF

Up against the shelf.

All models →
Model Input /M Output /M Effective blended Context Best for
Ember-1 Current $3.00 cache $0.30 $15.00 $1.92 agentic 92/8 1M Coding agents with shorter reasoning traces
Kimi K3 $3.00 cache $0.30 $15.00 $1.92 same 1M The base model, direct from Moonshot
Claude Sonnet 5.5 $2.00 cache $0.10 $10.00 $1.29 −33% 1M Anthropic's mid tier for coding agents
GPT-6.1 Sol $2.00 cache $0.10 $10.00 $1.29 −33% 1.05M OpenAI's mid tier with reasoning
Grok 4.7 $2.00 cache $0.50 $6.00 $1.19 −38% 500K A lower output rate for long answers
GLM-5.3 $1.40 cache $0.26 $4.40 $0.78 −2.5× 1M Frontier-class text model at a lower rate
DeepSeek V4.1 Flash $0.30 cache $0.006 $1.20 $0.15 −13× 1M Budget reasoning and coding

Frequently asked.

What Ember-1 costs, whether it is actually cheaper than Kimi K3, and what Fireworks has not priced.

Q · 01 How much does Ember-1 cost? +
$3.00/M input, $0.30/M cached input and $15.00/M output on Fireworks' serverless API, or $1.92/M on our 92/8 agentic blend with 82% cache hits. Reasoning tokens bill as output.
Q · 02 What is Ember-1? +
Fireworks' own model, and the first of a series Fireworks Research says it will build: Kimi K3 post-trained to cut unnecessary reasoning while keeping the self-correction that matters. The model page lists a 2.78-trillion-parameter mixture-of-experts, a context of 1,040K tokens, image input and function calling. It was announced on September 23, 2026.
Q · 03 Is it cheaper than Kimi K3? +
Not per token - the card is identical to Moonshot's K3 list and to Fireworks' own K3 serverless row. It can be cheaper per task, because Fireworks reports that it thinks less: reasoning 35-50% shorter across seven benchmarks at matching accuracy, about 35% fewer tokens per task in live A/B tests with two customers, and in one worked example 29.9K output tokens where K3 used 49.3K. Those are Fireworks' measurements, not reproduced here. The two coding scenarios above show what a 39% cut in output tokens is worth at these rates.
Q · 04 Where can I run it? +
Only on Fireworks - per token on the serverless API, or on dedicated on-demand GPUs billed per GPU-second, which is not a token rate. There are no open weights, and the model page lists fine-tuning as not supported.
Q · 05 Is there a batch or priority rate? +
Not one published for Ember-1. Fireworks bills its Batch API at 50% of serverless prices for batch-compatible models - any model that supports on-demand deployment, which Ember-1's page lists - but it warns newly added models can lag, and prints no Ember-1 batch line. Ember-1 is also absent from the Priority and Fast pricing table, which Fireworks calls the source of truth for Priority availability. This page stores only the serverless rate.
Q · 06 How does caching compare with Moonshot's? +
Cached input is $0.30/M on both, 90% off input. Moonshot now prints a separate cache-write line for K3 - $3.00/M for a 5-minute TTL, equal to its input rate, or $6.00 for an hour - while Fireworks publishes no write rate for Ember-1, so a missed token costs the input rate. At the default TTL the two work out the same.
Y
Reviewed by Yaroslav Vikhariev Founder · AI//COST · Prices checked by hand against fireworks.ai · Last verified Sep 24, 2026
Methodology Report a correction More by Y.V.