# Kimi K2 pricing, and the discount most teams leave unclaimed > Kimi K2 pricing on Groq is $1.00 per million input and $3.00 output, with a cached read at half the input rate. On agent traffic that's a quarter off. URL: https://getculpa.com/moonshotai-kimi-k2-instruct-0905-pricing Last reviewed: 2026-08-01 Rates effective: 2026-07-02 ## Answer 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. ## Common mistakes - Setting up caching on the chat path and stopping there. Why it hurts: Agent traffic repeats far more input than chat does, so it's where the discount is worth most. Do instead: Put the cache breakpoint after the tool definitions and the system prompt on every agent run. - Placing the accumulating transcript above the stable block. Why it hurts: The cache breaks at the first differing token, so a growing transcript at the top voids everything. Do instead: Order the prompt stable first: tool definitions, then system prompt, then the transcript. - Measuring agent cost per run without splitting cached from uncached input. Why it hurts: The run total hides whether the discount applied, so a broken cache looks like normal spend. Do instead: Report cached share per run and alert when it drops. ## Self-check - Measure what share of a typical agent step's input is identical to the previous step's. - Check where your cache breakpoint sits relative to the tool definitions. - Read your cached token share on this model. Near zero on agent traffic is money left on the table. - Price the repeated block at full rate and at half rate. The gap is the unclaimed discount. - Confirm the transcript sits below the stable block rather than above it. ## An agent workload with the cache on and off (illustrative) 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. ## Cost figures Every figure carries its confidence and its source. No figure on this site is provider-reported. - $1.00 / $3.00 per million — Kimi K2 Instruct input and output rates on Groq, with a cached read at half input [calculated] Source: Price 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 input [calculated] Source: $19.00 saved against a $77.00 uncached total, from the teardown arithmetic at real Groq rates. The repeated share is modelled. ## FAQ Q: How much does Kimi K2 cost on Groq? A: $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. Q: How much is the cached rate worth on agent traffic? A: 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. Q: Why do agents benefit from caching more than chat? A: 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. ## Sources - Groq pricing: https://groq.com/pricing Run the free Cost Leak Scan: https://app.getculpa.com/scan?source=pseo&slug=moonshotai-kimi-k2-instruct-0905-pricing&cluster=model_pricing Machine-readable index of every guide: https://getculpa.com/api/pages Human-readable index of every guide: https://getculpa.com/guides Site overview: https://app.getculpa.com/llms.txt Privacy: 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.