DeepSeek vs Kimi Cost Comparison

Snapshot used here

Rates come from data/ai-provider-pricing.json, generated at 2026-08-22T10:54:07.443Z. GPU rental rates come from data/gpu/gpu-pricing.json, generated at 2026-08-22T14:22:27.451Z. Throughput measurements come from data/gpu/gpu-throughput.json, generated at 2026-07-09T12:35:09.683Z. Confirm rates on provider pages before buying, since this page does not recheck live pages.

API rates, side by side

All rates are USD per 1M tokens. Each row names the exact provider and model record, with named versions throughout.

Model	Record	Input	Output	Cache read	Context
`deepseek-v4-flash`	`deepseek`	`0.44`	`1.32`	`0.014`	`1,000,000`
`kimi-k2.5`	`moonshotai`	`0.60`	`3.00`	`0.10`	`262,144`
`kimi-k2.6`	`moonshotai`	`0.95`	`4.00`	`0.16`	`262,144`
`kimi-k2.7-code`	`moonshotai`	`0.95`	`4.00`	`0.19`	`262,144`
`deepseek-v4-pro`	`deepseek`	`1.32`	`3.96`	`0.044`	`1,000,000`
`kimi-k3`	`moonshotai`	`3.00`	`15.00`	`0.30`	`1,048,576`

The deepseek records link an official pricing page in the snapshot, while the moonshotai records carry docs URLs without a dedicated pricing URL, so the Kimi rows deserve a direct check on the Moonshot platform. Price order never ranks quality, so keep evaluation scores in a separate column.

Equal-workload math

Compare equal volumes, 1M input tokens plus 1M output tokens, with input and output kept separate:

deepseek-v4-flash: 0.44 + 1.32 = $1.76
kimi-k2.5: 0.60 + 3.00 = $3.60
kimi-k2.6: 0.95 + 4.00 = $4.95
kimi-k2.7-code: 0.95 + 4.00 = $4.95
deepseek-v4-pro: 1.32 + 3.96 = $5.28
kimi-k3: 3.00 + 15.00 = $18.00

A monthly agent workload of 10M input and 5M output tokens gives the same ranking at scale:

deepseek-v4-flash: 10 × 0.44 + 5 × 1.32 = 4.40 + 6.60 = $11.00
kimi-k2.5: 10 × 0.60 + 5 × 3.00 = 6.00 + 15.00 = $21.00
kimi-k2.7-code: 10 × 0.95 + 5 × 4.00 = 9.50 + 20.00 = $29.50
deepseek-v4-pro: 10 × 1.32 + 5 × 3.96 = 13.20 + 19.80 = $33.00
kimi-k3: 10 × 3.00 + 5 × 15.00 = 30.00 + 75.00 = $105.00

Repeated prefixes favor DeepSeek workload math. One million repeated input tokens cost 1,000,000 ÷ 1,000,000 × 0.014 = $0.014 as cache reads on deepseek-v4-flash, against $0.044 on deepseek-v4-pro, $0.10 on kimi-k2.5, $0.19 on kimi-k2.7-code, and $0.30 on kimi-k3.

Context and output limits

Context separates the families more than input price does. Both DeepSeek rows allow 1,000,000 context tokens with up to 384,000 output tokens. The Kimi K2 rows allow 262,144 context and output tokens. Kimi K3 allows 1,048,576 context tokens with up to 131,072 output tokens, the largest context in this comparison paired with the highest output rate. Long-document work should test whether retrieval over a cheaper row replaces full-context passes over Kimi K3.

Self-host sketches, modeled

Throughput measurements and modeled cost stay in separate steps. The committed throughput file has rows for deepseek-v4-pro and the Kimi K2.5 class (kimi-k2-5), and none for deepseek-v4-flash, kimi-k3, kimi-k2.6, or kimi-k2.7-code, so only two sketches below rest on committed measurements.

DeepSeek V4 Pro records 4,074.26 decode tokens per second per GPU on b200 with vLLM, fp4, 1k input and 1k output tokens, and concurrency 4096. With the B200 median of $6.79 per GPU hour across 201 catalog records:

4,074.26 × 3,600 × 730 = 10,707,153,528 tokens per month at full use, about 10,707M.

6.79 × 730 = $4,956.70 per GPU month.

4,956.70 ÷ 10,707.15 = $0.46 per 1M output tokens at full use, against the $3.96 API output rate, with a break-even near 0.46 ÷ 3.96 = 12 percent sustained use.

Kimi K2.5 records 260.58 decode tokens per second per GPU on h200-sxm-141gb with vLLM, int4, 1k input and 1k output tokens, and concurrency 64. With the H200 median of $4.59 per GPU hour across 216 catalog records:

260.58 × 3,600 × 730 = 684,804,240 tokens per month at full use, about 684.8M.

4.59 × 730 = $3,350.70 per GPU month.

3,350.70 ÷ 684.80 = $4.89 per 1M output tokens at full use, against the $3.00 API output rate, so the modeled sketch stays above API cost even at full use.

Both sketches exclude staff time, replicas, and power, and both lab rows used short 1k contexts, so production long-context cost may be higher. Size memory with the GPU VRAM fit calculator and convert GPU spend with the open model token cost calculator.

Practical conclusion

Four findings follow directly from the numbers above. DeepSeek V4 Flash is the cheapest row in this comparison on equal volumes at $1.76 per 1M input plus 1M output tokens. Kimi K2.5 at $3.60 sits between the two DeepSeek versions and undercuts V4 Pro at $5.28. Kimi K3 at $18.00 needs a workload that truly uses its 1M context to justify the rate. Self-hosting has a plausible modeled path for V4 Pro near 12 percent sustained use on B200 hardware, while the Kimi K2.5 sketch on H200 hardware does not beat its API rate under lab conditions.

These throughput figures are normalized per GPU across a serving cluster. They do not imply the model fits on one GPU; total rental spend scales with the number of GPUs required. These output-only sketches exclude API input charges and are not complete service-cost comparisons.

Frequently asked questions

Is DeepSeek always cheaper than Kimi?

No. V4 Flash undercuts every Kimi row here, while V4 Pro costs more than Kimi K2.5, K2.6, and K2.7-Code on equal volumes. Match the version, not the family name.

Is Kimi K3 worth it for long context?

Only when the workload fills the 1,048,576-token context with work that retrieval cannot replace. At $15.00 per 1M output tokens, long generations dominate the bill, so measure context use before committing.

Should either family be self-hosted?

V4 Pro shows a modeled break-even near 12 percent sustained B200 use, which invites a careful trial with measured throughput. The Kimi K2.5 sketch stays above its API rate even at full modeled use. Either path needs production measurements at real context lengths first.

DeepSeek vs Kimi Cost: V4 Flash and Pro Against K2.5, K2.7-Code, and K3