Guides / kimi k2 pricing

Kimi K2 pricing, and the discount most teams leave unclaimed

Kimi K2 pricing on Groq is $1.00 per million input tokens and $3.00 per million output tokens, with a cached read at exactly half the input rate. On agentic traffic, where tool definitions repeat on every step, that discount is worth about a quarter of the bill. Culpa, a local-first LLM cost, margin, and forecast ledger, prices cached and uncached input separately so the unclaimed part is visible.

Why this happens

This is the agentic model on the platform, and agentic traffic carries the most repetitive input of any workload you run. Tool definitions, the system prompt and the accumulating transcript go out on every step of every run, so the repeated share of input is usually far higher than on a chat feature. It's also one of the few models here carrying a cached rate at all. Those two facts point the same way and most teams act on neither, because caching gets set up on the chat path and never on the agent path.

What this usually looks like

  • Tool definitions are resent on every agent step with no cache breakpoint in front of them.
  • Cached token share on this model reads near zero despite a highly repetitive prompt.
  • Caching was implemented for a chat feature and never extended to the agent path.
  • Input dominates the bill here and nobody has measured how much of it repeats.

Free, no card, no account

Run the free Cost Leak Scan

It shows your most expensive conversation before you install anything.

Run a free scan

Mistakes that cost the most

MistakeWhy it hurtsDo instead
Setting up caching on the chat path and stopping there.Agent traffic repeats far more input than chat does, so it's where the discount is worth most.Put the cache breakpoint after the tool definitions and the system prompt on every agent run.
Placing the accumulating transcript above the stable block.The cache breaks at the first differing token, so a growing transcript at the top voids everything.Order the prompt stable first: tool definitions, then system prompt, then the transcript.
Measuring agent cost per run without splitting cached from uncached input.The run total hides whether the discount applied, so a broken cache looks like normal spend.Report cached share per run and alert when it drops.

Run this check tonight

  1. Measure what share of a typical agent step's input is identical to the previous step's.
  2. Check where your cache breakpoint sits relative to the tool definitions.
  3. Read your cached token share on this model. Near zero on agent traffic is money left on the table.
  4. Price the repeated block at full rate and at half rate. The gap is the unclaimed discount.
  5. Confirm the transcript sits below the stable block rather than above it.

An agent workload with the cache on and off

Illustrative example

50 million input tokens a month, of which 38 million is repeated tool definitions and system prompt, plus 9 million output, on Kimi K2 Instruct at real Groq rates. The repeated share is modelled.

Uncached: (50 x $1.00) + (9 x $3.00) = $50.00 + $27.00 = $77.00
Cached on the repeated 38M at $0.50: (38 x $0.50) + (12 x $1.00) + (9 x $3.00) = $19.00 + $12.00 + $27.00 = $58.00
Saving = $19.00, which is 25% of the total
The cached rate is exactly half input, so the saving is half of whatever share repeats
At a 90% repeated share the saving would be 0.5 x 0.9 x $50.00 = $22.50

A quarter off the bill for putting the tool definitions above the cache breakpoint. The rule is simple: the saving is half your repeated input share, so measuring that share tells you the answer before you build anything.

Every number, with its confidence and source

FigureWhat it meansConfidenceSource
$1.00 / $3.00 per millionKimi K2 Instruct input and output rates on Groq, with a cached read at half inputcalculatedPrice book row for groq/moonshotai/kimi-k2-instruct-0905, $0.001 and $0.003 per 1k with cached $0.0005, effective 2026-07-02, re-verified against Groq's rendered pricing page 2026-08-01.
25%modelled saving from caching a repeated agent prefix making up 76% of inputcalculated$19.00 saved against a $77.00 uncached total, from the teardown arithmetic at real Groq rates. The repeated share is modelled.

What a generic answer can’t know

A provider can price a cached read. Whether your agent's tool definitions sit above or below the breakpoint, and what share of each step repeats the last one, are facts about your own orchestration. Culpa records them on your infrastructure, keeps your prompts and responses there, and counts the calls to run your plan.

Questions founders ask next

How much does Kimi K2 cost on Groq?

$1.00 per million input tokens and $3.00 per million output, with a cached read at $0.50 per million, exactly half the input rate. Re-verified against Groq's pricing page on 2026-08-01.

How much is the cached rate worth on agent traffic?

Half of whatever share of your input repeats. On an agent resending tool definitions and a system prompt every step, that share is often three quarters or more, which puts the saving around a quarter of the whole bill.

Why do agents benefit from caching more than chat?

Because an agent run repeats the same tool definitions and system prompt on every step, and there can be dozens of steps in one run. A chat turn repeats far less, so the same cache breakpoint returns much less on that path.

On your infrastructure

Culpa runs on your infrastructure. Your prompts and responses never leave it. Culpa counts calls to run your plan, and it fails open, so if it ever breaks your app keeps running.


How Culpa works

Find the culprit. Not just the total.

Your dashboard shows what you spent. It stops short of who spent it. Culpa shows the conversation, the user and the feature behind it.

Your prompts stay local.

Culpa runs on your own infrastructure. What you send to a model reaches us at no point.

Every dollar has a name.

Follow any charge to the conversation, the user, the feature and the customer behind it.

See the bill before it lands.

Cost your next feature before you ship it. You get the likely bill and the worst case, at best, median, p90 and p99.

Three steps to your first answer.

1

Change one base URL.

Or drop in the Python or TypeScript library.

2

Find your most expensive conversation.

In the first session, not the first week.

3

Cost your next feature before you ship it.

Base URLhttp://localhost:4545/v1Your traffic keeps flowing if Culpa ever stops.

Why the bill went up

Example dashboard

Calls traced

418,209

across 3 projects

Spend this week

$378.41

+ $182 vs last week

Failed calls

312

74% retried, and you paid for all of them

+ $182 this week traced to one culprit

Spend over 14 days

$0$20$40$60$8024262830020406
user_384report_generatorconv_91fprompt_v1894,220 tokens3 retries$6.81

Most expensive users

user_384$38.42
user_119$21.07
user_562$14.90
user_204$8.30
user_871$5.10

Next week forecast

Best$180
Median$240
p90$310
p99$395

Graded against reality. Accuracy shown as results land.

Free, no card, no account

Run the free Cost Leak Scan

It shows your most expensive conversation before you install anything.

Run a free scan

Keep reading


Sources: Groq pricing. Last reviewed 2026-08-01, rates effective 2026-07-02. Plain text version.