News · 2026-09-28

Fireworks ships Ember-1, an API model built to spend fewer reasoning tokens

Fireworks has launched Ember-1, a public API model designed to cut the hidden reasoning-token overhead of agents while retaining planning and tool-use quality. The concrete product is live; the larger claim—that it achieves comparable quality with 35–50% shorter reasoning—is Fireworks' own measurement rather than an independent benchmark result.

Key facts

• Product: Ember-1, a specialized model Fireworks says was built from Kimi K3.
• Claim: 35–50% shorter reasoning traces at comparable quality across the company's evaluations.
• Availability: the Ember-1 model page offers serverless REST, Python and OpenAI-compatible inference.
• Price: $3 per million input tokens, $0.30 per million cached-input tokens and $15 per million output tokens.

The business problem is not simply that large models are expensive. In agent workflows, a model's earlier reasoning often comes back as context for the next turn. A needlessly long internal trace can therefore be paid for twice: once when it is generated and again when it is reread. Fireworks' launch post says Ember-1 was trained to keep useful self-reflection, environmental feedback and planning while avoiding unnecessary internal narration.

Think of an experienced mechanic writing a repair log. The important notes are the diagnosis, the test and the next action; a transcript of every glance at every bolt makes the notebook harder and costlier to reuse. Ember-1's intended advantage is to produce the compact, decision-relevant version. Fireworks says it conducted more than 50 training experiments, more than 200 evaluations and two production coding A/B tests. Those are meaningful release details, but they are still the vendor's evidence.

The most useful hard number for a buyer is the price list, because it makes the savings claim legible. A 35–50% reduction in reasoning tokens can be material when an agent keeps revisiting context over long tool loops. The model is not released as downloadable weights; it is a hosted API offering, so readers should not infer a disk download or VRAM requirement. A model page that lists an endpoint is not the same thing as an open checkpoint.

Fireworks also describes an availability wrinkle. Its announcement says training support is being rolled out, while the current model page says fine-tuning is not supported. The reliable statement is that inference is usable now. It would be premature to promise customer fine-tuning until the product documentation changes.

The strongest counterargument is that “comparable quality” is the hardest part of the claim and no independent evaluator has yet established the tradeoff. Short traces can save money by omitting useful checking, and benchmark equivalence may not survive a different agent harness. This is especially relevant after current work on inference cost and token economics: output and repeated context have a different cost profile than a single prompt, so the evaluation needs to capture an entire workflow.

Still, Ember-1 is more than a paper claim. It has a named public endpoint, a published rate card and ordinary developer access. That makes it a shipping product in a part of the market where many announcements are not. The early story is therefore a narrow but relevant one: Fireworks is trying to compete on useful reasoning per billed token rather than maximum visible deliberation. Whether that frontier holds up under independent, long-running agent tasks is the next test.

Primary source, verified: read the paper →

Ground Truth. (2026, September 28). Fireworks ships Ember-1, an API model built to spend fewer reasoning tokens. Ground Truth. https://groundtruth.day/news/fireworks-ember-1-reasoning-token-api.html

@misc{groundtruth:fireworks-ember-1-reasoning-token-api,
  title  = {Fireworks ships Ember-1, an API model built to spend fewer reasoning tokens},
  author = {{Ground Truth}},
  year   = {2026},
  month  = {sep},
  url    = {https://groundtruth.day/news/fireworks-ember-1-reasoning-token-api.html}
}

Topics: product-launch · models · agents · inference-cost · api