Kimi K3 on its cheapest host undercuts Fireworks' Ember-1 despite a 23% cut in reasoning tokens
Fireworks' Ember-1 used 23% fewer reasoning tokens than Kimi K3 in The New Stack's tests, yet Kimi on the cheapest host would cost $1.96 to Ember's $2.48. Ember beats Fireworks' own Kimi rate and loses at the cheapest, so buyers have to price the host before the model.
The Engineer · Build desk
What happened
- Fireworks launched Ember-1 on September 23 as a research preview built on Moonshot's open-weight Kimi K3.
- Fireworks bills Ember-1 at $3 per million input tokens and $15 per million output, the rate it charges for Kimi K3, while other hosts sell Kimi for as little as $1 and $9.
- Across three tests of rising difficulty, run five times each, Kimi K3 was right on all 15 runs and Ember-1 on 14.
- With both models served by Fireworks, Ember-1 finished each set of tests 3.4 times faster than Kimi K3.
Why it matters
- What's now possible capability Agent pipelines that wait on wall-clock time can get a large latency cut from Ember on Fireworks now, an advantage not yet tested against the cheaper Kimi hosts.
- Who's exposed exposure A buyer who adopts Ember on Fireworks' benchmark and customer figures carries the risk that its own tasks resemble neither, since the evaluations are vendor-run and the customers unnamed.
- What it sets up precedent Fireworks calls Ember-1 the first of a series of specialized models from its Serverless Training service, so more house-trained derivatives will likely arrive at its own rates for the base model.
Next story · Product · 8h ago
China's draft export list would cover the weights of its biggest AI models while leaving API access openConfirmed8 publishers
Reasoning tokens bill as output [4]. In The New Stack's runs, output was nearly the whole bill [28]. Kimi K3 cost $3.26 for the full suite on Fireworks and would have cost $1.96 at the cheapest listed price [19][20]. The cheaper bill is 0.60 of the Fireworks one [27]. The output prices, $9 and $15, sit in the same ratio, while the input prices differ by a factor of three [28]. Moving Kimi to the cheapest host saves $1.30 on the suite. Moving to Ember on Fireworks saves $0.78 [30].
A bill made of output tokens has a clean break-even. At $15 against $9, Ember-1 has to emit 40% fewer output tokens than Kimi just to tie the cheaper host [29]. Fireworks advertises roughly 40% fewer tokens as Ember's headline result [2], a coincidence I would not put in a budget. The measured reasoning-token cuts were 18% on the logic puzzles, 16% on deploy scheduling and 35.5% on the probability questions [10][11][31].
The tester built the suite on a stated hypothesis: if Ember cut reasoning it needed, accuracy would slip first on the hardest problems [16]. The probability test produced both Ember's largest cut and its only error [31][12]. On the fifth run it answered the first question 0.94619 instead of 0.94629, and the slip carried into two other answers [12]. The tester wrote that the miss "doesn't seem too significant to me since it was one out of 15" [15].
According to runtimewire, Fireworks found that turning down Kimi K3's reasoning-effort setting weakened its results [23]. Its researchers instead post-trained a K3-based model to shorten reasoning while keeping the parts that help it recover from mistakes and adapt to feedback [23]. Training the shorter trace into the model, once the setting had failed, is the sound choice. The work ran to more than 50 training experiments and over 200 evaluations [23]. Fireworks said the model "learned to cut unnecessary reasoning while keeping the thinking that matters" [3].
Fireworks' own evaluations put Ember-1 at 92.2% on SWE-bench Verified against 93.2% for K3 at maximum effort, and at 82.0% against 80.9% on Terminal-Bench 2.1 [24]. The benchmark sets run from 50 to 500 examples [24]. In a live test with one of two customers, Fireworks reported 39% fewer total tokens and 71.3% fewer reasoning tokens, at a task score of 0.753 against 0.751 [25]. If that customer's other tokens held flat, reasoning was about 55% of its baseline [33]. A team whose traffic is less reasoning-heavy would get a smaller total cut from the same model [33].
Token savings do not account for Ember's speed [18][17]. On the logic puzzles, Ember produced about 60 reasoning tokens per second of wall time and Kimi about 22 [32]. Fireworks was generating Ember's tokens faster [32]. Dmytro Dzhulgakov, a Fireworks co-founder, wrote in a September 27 post on X that Ember-1 was "hot on HN" and said it was 40% faster and cheaper, according to runtimewire [21]. The outlet reports that the launch materials back the token claim but do not establish a general 40% speed gain [22].
The cost case is strongest for multi-turn agents, according to runtimewire. Fireworks says those agents may send earlier reasoning back to the model on each turn, so a long trace is processed repeatedly [5]. Its listed rate for cached input is $0.30 per million tokens [8]. In my view the comparison that holds is cost per correct task on a team's hardest workload, priced at every host it would buy from. On The New Stack's suite, Ember cost about 18 cents per correct run and Kimi on the cheapest host about 13 [34].
What to watch
- Speed and cost runs of Kimi K3 on the $1/$9 hosts, to show whether Ember's lead on time holds outside Fireworks' serving stack.
- Independent multi-turn agent tests that resend reasoning each turn, the setting where Fireworks says shorter traces cut cumulative token use.
- Any change to Ember-1's $3/$15 pricing now that its model page lists it as ready for serverless use after the research-preview window.
Clarity's read
What the record supports and how the coverage leans. The claims behind it follow.
Reality
- Evidence55
- Adoption30
- Hype gap+25
- Incentives70
- Confidence58
Perspective Coverage
3 publishers- Builder
- Builder 52%
- Operator
- Operator 30%
- Investor
- Investor 18%
Claim ledger
Ranked by verification strength, evidence, and original report placement.
- [1]
Fireworks Research launched Ember-1 on September 23 as a research preview, built on Moonshot's open-weight model Kimi K3.
- [2]
Fireworks claims Ember-1 matches Kimi K3's quality using roughly 40% fewer tokens.
- [3]
Fireworks says Ember-1 "learned to cut unnecessary reasoning while keeping the thinking that matters."
- [4]
Reasoning tokens bill as output.
- [5]
Fireworks says multi-turn agents may send earlier reasoning back to the model at each turn, so lengthy traces can be processed repeatedly as a task continues.
- [6]
Launch materials described Ember-1 as a research preview with two-week access for research models, continued availability tied to developer demand; the model page now lists it as ready for serverless use.
- [7]
On OpenRouter, Ember-1 costs $3 per million input tokens and $15 per million output tokens; Fireworks charges the same for Kimi K3, but other providers sell Kimi K3 for as little as $1 per million input tokens and $9 per million output tokens.
- [8]
Fireworks' serverless model page lists Ember-1 at $3 per million input tokens, $0.30 per million cached-input tokens and $15 per million output tokens.
- [9]
The tester called both models through OpenRouter with identical prompts and default reasoning settings, routed both to Fireworks, and ran each of three tests (logic puzzles, deploy scheduling, probability) five times per model, logging reasoning tokens separately.
- [10]
On the logic puzzles, Ember-1 averaged 13,630 reasoning tokens, 3 minutes 46 seconds and $0.27 per run; Kimi K3 averaged 16,679 reasoning tokens, 12 minutes 26 seconds and $0.34, making Ember 18% lower on reasoning tokens.
- [11]
On deploy scheduling, Ember-1 averaged 6,543 reasoning tokens, 1 minute 29 seconds and $0.10; Kimi K3 averaged 7,792 reasoning tokens, 4 minutes 46 seconds and $0.13, 16% fewer reasoning tokens for Ember.
- [12]
On the probability test, Kimi K3 answered all five questions correctly on every run; Ember-1 got them all right four times, and on the fifth made an arithmetic slip on the first question, 0.94619 instead of 0.94629, which carried into two other answers.
- [13]
On the probability test, Ember-1 averaged 6,242 reasoning tokens, 1 minute 47 seconds and $0.13; Kimi K3 averaged 9,682 reasoning tokens, 6 minutes 48 seconds and $0.19.
- [14]
Across the three tests, Kimi K3 had 15 perfect runs out of 15, and Ember-1 had 14.
- [15]
The tester wrote of Ember-1's miss: "it doesn't seem too significant to me since it was one out of 15."
- [16]
The tester reasoned that if Ember cut reasoning it actually needed, accuracy should slip first on the hardest problems, and built progressively harder tests.
- [17]
Ember-1 finished each set of tests 3.4 times faster than Kimi K3, with both routed to Fireworks.
- [18]
Overall, Ember-1 used 23% fewer reasoning tokens than Kimi K3.
- [19]
In total, Ember-1 cost $2.48 compared with Kimi K3's $3.26 on Fireworks, making it 24% cheaper.
- [20]
At Kimi K3's lowest listed price of $1 per million input tokens and $9 per million output tokens, the same runs would have cost $1.96, less than Ember-1.
- [21]
Fireworks co-founder Dmytro Dzhulgakov, in a September 27 post on X, described Ember-1 as "hot on HN" and said it was 40% faster and cheaper.
- [22]
Fireworks' launch materials substantiate the token-reduction claim with benchmark and customer-test results but do not establish a general 40% improvement in response speed.
- [23]
Fireworks says lowering Kimi K3's reasoning-effort setting weakened its results, so its researchers post-trained a K3-based model to shorten reasoning while retaining the parts that help it recover from mistakes and adapt to feedback, across more than 50 training experiments and over 200 evaluations.
- [24]
Fireworks reports Ember-1 at 92.2% on SWE-bench Verified versus 93.2% for Kimi K3 at maximum reasoning effort, and 82.0% versus 80.9% on Terminal-Bench 2.1; benchmarks range from 50 to 500 examples and are Fireworks' own evaluations.
- [25]
In one customer test, Fireworks reports a 39% reduction in total tokens and a 71.3% reduction in reasoning tokens, with task score 0.753 for Ember-1 against 0.751 for K3; it ran live tests with two customers, one of which put Ember-1 into production, and did not identify them.
- [26]
Fireworks says Ember-1 was developed using its Serverless Training service and is the first in a series of specialized models.
- [27]
Kimi K3's cheapest-host bill is about 0.60 of its Fireworks bill for the same runs.
- [28]
The output price ratio between the cheapest Kimi host and Fireworks is 0.60 and the input ratio is one third; since the overall bill ratio is 0.60, input tokens were a negligible share of the bill in these runs.
- [29]
At $15 versus $9 per million output tokens, Ember-1 must emit 40% fewer output tokens than Kimi K3 to tie the cheapest host on output cost.
- [30]
Moving Kimi K3 from Fireworks to the cheapest host saves $1.30 on the suite; switching to Ember-1 on Fireworks saves $0.78.
- [31]
On the probability test, Ember-1 used about 35.5% fewer reasoning tokens than Kimi K3, its largest cut of the three tests.
- [32]
On the logic puzzles Ember-1 produced about 60 reasoning tokens per second of wall time versus about 22 for Kimi K3, so a 23% token cut cannot alone explain the 3.4x speed gap; Fireworks was generating Ember's tokens faster.
- [33]
If the customer's non-reasoning tokens were unchanged, reasoning was about 55% of that customer's baseline tokens.
- [34]
Per correct run, Ember-1 on Fireworks cost about 18 cents and Kimi K3 at the cheapest host about 13 cents.
Sources
3 independent publishers whose own reporting we read for this story.
- blog.vercel.comEmber-1 from Fireworks now available on AI Gateway
1 article · September 26, 2026
- runtimewire.comFireworks says its Kimi K3 variant cuts reasoning tokens by 40%
1 article · September 27, 2026
- thenewstack.ioEmber-1 vs. Kimi K3: Nearly identical results at 3.4 times the speed
1 article · September 29, 2026
Topics and entities
Topics
- LLM Inference Pricing
- Open-Weight Model Publishing
- Reasoning token efficiency