Fireworks priced Ember-1 identically to Kimi K3 — the discount is in the token count, and it expires in two weeks unless enough people use it

A price cut with no price change

Every cost reduction this desk has covered this year showed up as a smaller number on a rate card. Ember-1 does not.

Fireworks published Ember-1 on 23 September 2026, describing it as a reasoning model from Fireworks Research built on Kimi K3, designed to “make every token go further” by producing shorter reasoning traces — roughly 40% fewer tokens than the base model across its evaluations, and about 35% fewer tokens per task in live customer A/B tests at comparable quality.

The rate card reads: $3.00 per million input tokens, $0.30 cached, $15.00 output, on a context window of roughly 1.04 million tokens.

Those are exactly the numbers Fireworks charges for base Kimi K3 on serverless — $3.00, $0.30 and $15.00 — and exactly what Moonshot charges on its own API. Not approximately. Identically.

This is an unusual and, on reflection, rather clean way to sell a model. There is no premium to justify, no discount to explain, and no pricing-page comparison to run. The proposition is simply: same rates, same quality, fewer tokens. Which means the only instrument that can tell you whether it worked is your invoice.

Where the saving goes, and where it evaporates

Reasoning tokens bill as output tokens. That single fact determines the entire economics here, because it means Ember-1 shrinks exactly one line of your bill and leaves the other untouched.

Take a task with 10,000 input tokens and 4,000 output tokens on Kimi K3:

• Input: 10,000 ÷ 1M × $3.00 = $0.030
• Output: 4,000 ÷ 1M × $15.00 = $0.060
• Total: $0.090

Now cut output by 35%, to 2,600 tokens, at identical rates:

• Input: $0.030 (unchanged)
• Output: 2,600 ÷ 1M × $15.00 = $0.039
• Total: $0.069

That is a 23% saving, not 35%. The headline number applies to a line item, and the line item is a fraction of the bill.

Now change one thing — make the prompt long, which is the shape most retrieval-augmented and long-context agentic work actually takes. 100,000 input tokens, same 4,000 output:

• Kimi K3: $0.300 + $0.060 = $0.360
• Ember-1 at 35% fewer output tokens: $0.300 + $0.039 = $0.339

A saving of under 6%. Same model, same claim, same arithmetic — and the benefit fell by a factor of four because of prompt shape alone.

This is the practical headline for buyers. Ember-1’s value is inversely proportional to how input-heavy your workload is. Short prompts with long deliberation — hard reasoning, multi-step agent loops — approach the advertised saving. Long prompts with short answers get close to nothing. No benchmark table can tell you which you are; a single query against your own logs can.

It is worth noting the one case where the logic inverts pleasantly: shorter outputs also mean lower latency and fewer tokens for downstream steps to read, so an agent chain compounds the benefit across hops in a way a single-call cost model understates.

The benchmarks are mixed, and self-reported

Fireworks published a comparison against K3 max that does not flatter itself uniformly:

Two up, two down. The declines are small, but Fireworks printed them rather than selecting the favourable pair, and the pattern is coherent with the training objective: compressing reasoning helps where over-deliberation wastes tokens and costs a little where the extra thinking was earning its keep.

The caveat that matters more than any individual row is that all of these are Fireworks measuring Fireworks, including the ~35% token reduction from customer A/B tests. No independent replication existed at the time of writing. A serving provider has a fair claim to knowing its own token counts better than anyone — but a token-reduction claim published by the company that bills per token is precisely the figure to reproduce on your own traffic before planning around it. The same scepticism this desk applies to vendor-published launch benchmarks generally applies here without modification.

The two-week clock

Ember-1 ships as a research preview. Fireworks describes a new programme in which research releases get two weeks of serverless access and become permanent based on community demand. Published 23 September, that window closes in the first week of October.

The incentive is reasonable rather than sinister: a serving provider needs evidence that a specialised derivative earns its GPU capacity before committing to it indefinitely, and a short public trial is a sane way to get that evidence. But it moves a specific risk onto the buyer. The cost of migrating onto Ember-1 is yours; the decision about whether it continues to exist is not.

That asymmetry prescribes the usage precisely:

• Evaluate it now. The rates are identical to the base model, so there is no premium to justify and no procurement case to write. The only cost is engineering time.
• Do not put it on a critical path, write it into a customer commitment, or let it become the only model a production route can call.
• Keep the fallback wired. Because Ember-1 is a derivative of open weights, base Kimi K3 remains self-hostable if the hosted variant disappears — a materially better position than a closed model withdrawn by its only supplier. Routing through a neutral gateway makes the fallback a config change rather than a deployment.

How it sits against the frontier tier

The obvious comparison is GPT-6 Sol, cut to $2/$10 on 22 September. Sol is a flat 50% cheaper per token than Ember-1. On the benchmark they share, Fireworks puts Ember-1 at 75.2% on DeepSWE 1.1 against OpenAI’s published 68.8% for Sol at max effort and 66.6% for Luna at max.

So: half again the token price, more than six points ahead on that benchmark, with a token-volume claim against its own base model that has never been measured against Sol. Anyone deciding between them has to run the comparison themselves, and the decisive variable is once again the input-to-output ratio rather than anything either vendor published.

The structural differences outweigh the arithmetic anyway. Sol is generally available and unversioned; Ember-1 has a survival clock but rests on open weights you can take elsewhere. Neither property appears on a rate card.

The thesis, stated plainly

The open-weights price floor has spent this year moving in public — stealth launches that reset the floor, a price war that inverted direction in August, and rate cards from DeepSeek and Qwen that made per-token comparison the default way buyers shop. Ember-1 is the first serious move that is invisible to that entire apparatus. Every price-comparison table, every cost dashboard keyed to dollars-per-million-tokens, and every procurement spreadsheet that ranks models by rate will show Ember-1 and Kimi K3 as identical products. They are not, and the difference is a number none of those tools measure.

That is not a criticism of Fireworks, which has been unusually transparent — it published its regressions, disclosed the preview terms, and did not dress a token-efficiency claim up as a discount. It is a warning about the instrument. Dollars per million tokens is a unit price, and unit price has never been cost. As soon as vendors start competing on consumption rather than on rate, the buyer who ranks models by rate card is optimising the one variable the market has stopped competing on.

For developers and anyone maintaining a coding-agent stack, the operational change is small: start logging tokens per completed task, not just dollars per million tokens, as a first-class metric alongside quality. That number is the only one that moves when a model like this works — and it is the same number that would have caught a silent routing downgrade at peak hours going the other way. Measure Ember-1 in the next two weeks, on your own traffic, with that metric. Then decide — knowing the vendor is deciding too.

Questions or corrections? Email Pick Right. Want the full list? See all news.