Every Kimi K3 API Provider Compared: Pricing, Speed, and the One 50% Discount

Kimi K3 dropped its open weights on July 27, and within 48 hours nearly every major inference platform had it live. I spent a day checking every provider's rate card and performance numbers so you don't have to.

The punchline: nine providers charge the identical price. One charges half.

The full comparison table

Provider	Input /1M	Cached /1M	Output /1M	Throughput	TTFT
TeamoRouter	$1.50	$0.15	$7.50	not disclosed	—
Moonshot (official)	$3.00	$0.30	$15.00	30 tps	6.1s
OpenRouter	$3.00	$0.30	$15.00	not disclosed	—
Fireworks	$3.00	$0.30	$15.00	69 tps	2.4s
SiliconFlow	$3.00	$0.30	$15.00	not disclosed	—
Novita	$3.00	—	$15.00	19 tps	1.4s
Baseten	$3.00	—	$15.00	52 tps	3.1s
Vercel AI Gateway	$3.00	$0.30	$15.00	routes to 4 above	—
Together AI	not disclosed	—	not disclosed	—	—
Modal	not disclosed	—	not disclosed	460 tps	—

Throughput/TTFT numbers are from Vercel AI Gateway's live provider dashboard. Modal's 460 tps is from their launch-day post (modal.com blog); they also give $30/month in free compute, which is the cheapest way to just try K3.

Not serving K3 yet: Groq, DeepInfra (K2.x only).

Why is everyone at $3/$15?

K3 is a 2.8T-parameter MoE (896 experts, 16 active per token, 1M context, native vision). Serving something this big is expensive — the uniform pricing suggests margins at $3/$15 are already thin. So providers differentiate on speed instead:

• Fireworks wins raw throughput among priced providers (69 tps)
• Novita wins time-to-first-token (1.4s) and supports the full 1M output
• Baseten sits in the middle (52 tps / 3.1s)
• Moonshot's own API is, ironically, the slowest measured (30 tps / 6.1s)

If you're latency-sensitive, pick by these numbers — the price is the same anyway.

The exception: 50% off through a routing gateway

TeamoRouter is the only platform I found charging below the official rate: $1.50 input / $0.15 cached / $7.50 output — exactly half. It's a multi-model routing gateway (same idea as OpenRouter: one key, many models, OpenAI-compatible API), and K3 is currently its loss-leader.

OpenAI SDK setup:

import openai

client = openai.OpenAI(
    api_key="YOUR_TEAMOROUTER_KEY",  # from teamorouter.com
    base_url="https://api.teamorouter.com/v1",
)

resp = client.chat.completions.create(
    model="kimi-k3",
    messages=[{"role": "user", "content": "Write a binary search in Rust"}],
)
print(resp.choices[0].message.content)

There's also an Anthropic-compatible endpoint (https://api.teamorouter.com/anthropic), which matters if your tooling speaks the Claude API — more on that in a second.

What does 50% off actually mean in dollars?

Take a code-review bot doing 10K reviews/month, 50K input + 2K output tokens each:

At $3/$15:
  input:  10,000 × 50K × $3/1M  = $1,500
  output: 10,000 × 2K  × $15/1M = $300
  total:  $1,800/month

At $1.50/$7.50:
  total:  $900/month  →  $10,800/year saved

And caching stacks on top. A 40-turn agent session with a stable system prompt hits the cached rate ($0.15 vs $0.30) on nearly every turn after the first, so the effective input cost lands at a fraction of the naive number.

Is K3 even worth routing to?

Short version — yes, for coding:

• Frontend Code Arena: Elo 1679, rank #1 (above Claude Fable 5 at 1631 and GPT-5.6 Sol at 1618)
• SWE-bench Verified: 93.40%
• Terminal-Bench 2.1: 88.3%
• Artificial Analysis Intelligence Index: 57.0 (#4 overall)

Compare pricing: Claude Fable 5 runs $15/$50, GPT-5.6 Sol $10/$40. K3 at the official rate is roughly 1/3 the cost; at the discounted rate it's ~1/6 of Fable 5. For frontend-heavy work, the benchmark leader is also the cheapest option on the board.

Using it in Claude Code

export ANTHROPIC_BASE_URL="https://api.teamorouter.com/anthropic"
export ANTHROPIC_API_KEY="YOUR_TEAMOROUTER_KEY"
export ANTHROPIC_MODEL="kimi-k3"
claude

Same agent workflow, half the token bill. Works from mainland China without a VPN too (payment via Alipay/WeChat/Stripe), which none of the US providers handle.

TL;DR

• 10 platforms serve Kimi K3; 9 charge exactly $3/$15/1M ($0.30 cached)
• Speed is the real differentiator: Fireworks 69 tps, Novita 1.4s TTFT, Modal claims 460 tps (price undisclosed)
• TeamoRouter is the lone discount at $1.50/$7.50 (50% off), OpenAI- and Anthropic-compatible
• K3 itself is the #1 frontend-coding model right now (Elo 1679) at a fraction of Claude/GPT pricing

Data checked 2026-07-28. Prices move fast — verify before you commit production traffic.