Analysis · 2026-08-07 · 8 min read

Kimi K3 on Fireworks: Frontier Quality at $3.00/$15.00 per Million Tokens

Moonshot's Kimi K3 is the first open-weight model to top Frontend Code Arena — and it runs at $3.00/M input ($0.30/M cached) and $15.00/M output with a flat 1M-token context. Fireworks benchmarks put it at up to 5x better cost-per-task than Opus 5. Here's the pricing math and when to route to it.

TL;DR

Kimi K3 is Moonshot AI's 2.8T-parameter open-weight frontier model — the first open model to rank #1 on Arena's Frontend Code Arena (1,679 Elo), ahead of every closed model.
Pricing is flat across the full 1M-token context: $3.00 per million input tokens (cache miss), $0.30 per million input tokens (cache hit), $15.00 per million output tokens — via Moonshot AI and aggregators like OpenRouter.
Fireworks AI now serves it Day-0 with US-only serverless endpoints, Zero Data Retention, and 50% off batch mode.
Fireworks' head-to-head against Claude Opus 5 found matching accuracy at up to 5x better cost per finished task — $0.35 vs $1.61 on Terminal, $0.52 vs $1.05 on SWE.
The practical takeaway: K3 is cheap enough to be a default, not an exception — with prompt caching doing the heavy lifting on agentic loops.

!Moonshot Kimi logo

The Pricing, Plainly

Two things matter here more than the headline number.

One: there is no long-context tier. Gemini 2.5 Pro and GPT-5.x step up their rates past a context threshold. K3 does not. Token 900,000 costs exactly what token 900 costs. For document-heavy RAG, whole-repo code review, and long agent trajectories, that flat curve is worth more than a slightly lower base rate.

Two: the cache-hit rate is $0.30. At a 10x discount on repeated prefixes, an agent loop that re-sends the same system prompt, tool schema, and file context on every turn converges toward the cached rate. A 55-turn agentic run with a stable 40K-token prefix pays full price once and $0.30/M for the other 54 turns.

You can check the live rate on our AI Model Pricing Table or pull it programmatically from the public pricing dataset.

What K3 Actually Scores

The benchmark story is what makes the price interesting. From independent evaluations:

This is the pattern we keep flagging: open-weight models are no longer trailing on quality, they're trailing on *distribution*. When a $3/M model wins Frontend Code Arena outright, the "frontier tax" on closed models becomes a lot harder to justify per line item.

Cost Per Task Beats Cost Per Token

Fireworks benchmarked K3 head-to-head against Claude Opus 5 and published the cost-per-task table. This is the metric that actually shows up on your invoice:

Read the SWE row carefully: K3 takes more turns (55.6 vs 37.9) and still costs half as much per completed task. Token efficiency lost, dollar efficiency won. On Terminal the gap is 4.6x.

The accuracy delta is real — 2.1 points on SWE, 3.6 on Terminal. Whether that matters is a routing decision, not a procurement decision. For the 80% of tasks where a 2-point accuracy difference is noise, you're paying 2-5x for it.

!Anthropic Claude logo

Why It's Cheap: The Architecture

K3 isn't cheap because Moonshot is subsidizing it. It's cheap because of how little of the model runs per token.

A 2.8T-parameter model that activates 104B parameters per token bills like a 104B model and reasons like a frontier one. Moonshot's Kimi Delta Attention plus Stable Latent MoE routing delivered roughly 2.5x the scaling efficiency of Kimi K2 — which is exactly the mechanism that lets a flat 1M-context price exist at all.

Where Fireworks Fits

Moonshot recommends 64+ accelerator supernodes to self-host K3. That's not a realistic line item for most teams, and idle GPU time destroys the economics anyway — the same math we walk through in our self-host vs API breakeven calculator.

Fireworks serves it per-token instead, with a few things worth noting for regulated teams:

US-only serverless endpoints, launching with K3 first.
Zero Data Retention by default — prompts and generations are not logged for open models without explicit opt-in.
Batch mode at 50% off for anything delay-tolerant: evals, bulk image captioning, backfills.
Serverless training in private preview — LoRA fine-tuning billed per token instead of per reserved GPU-hour.

!Fireworks AI logo

Replit's Michele Catasta said K3 on Replit Design delivered "unprecedented product design and UI capabilities, at a fraction of the cost we've come to expect from frontier models." Mercor's Brendan Foody flagged the same trade collapsing: top-tier performance without giving up scale or data-residency guarantees.

How to Route It

The pattern that works — and the one Coinbase used to cut its AI bill in half:

1. Make K3 the default for coding, front-end generation, long-document analysis, and multi-step agent loops. At $3/M in / $15/M out it's a frontier model at a mid-tier price.

2. Engineer for cache hits. Put the stable system prompt, tool definitions, and shared context at the front of every request. Getting a 60% hit rate drops effective input cost from $3.00 to about $1.38/M; at 90% it's $0.57/M.

3. Batch what can wait. Evals, labeling, and backfills at 50% off turn a $15/M output rate into $7.50/M.

4. Escalate on failure, not by default. Route to Opus 5 / GPT-5.6 only when K3's output fails a validator, test suite, or confidence check — not preemptively.

5. Measure cost per merged PR, not tokens. The Fireworks table only makes sense at the task level; token counts would have told you the opposite story on SWE.

The Bigger Signal

An open-weight model taking the #1 spot on Frontend Code Arena — while pricing input at $3.00 and cached input at $0.30 — is the clearest data point yet in the shift we covered in the end of tokenmaxxing. Buyers aren't asking "which model is best?" anymore. They're asking "which model is good enough per dollar, and what do I escalate?"

For a growing share of production work, the answer to the first half of that question is now an open model you can download, fine-tune, and host wherever your compliance team says you have to.

Compare K3 side by side against Opus 5, GPT-5.6, and GLM-5.2 on the LLM Leaderboard, or run your own numbers in the cost calculator.