Kimi has launched a newer model, Kimi K2.5, we suggest considering this model instead.

For more information, see Comparison of Kimi K2.5 to other models and API provider benchmarks for Kimi K2.5.

Kimi K2 Thinking logo
Kimi
•

Open weights model

•

Released November 2025

Kimi K2 Thinking API Provider Benchmarking & Analysis

Model Comparison

Analysis of API providers for Kimi K2 Thinking across performance metrics including latency (time to first token), output speed (output tokens per second), price and others. API providers benchmarked include Kimi Turbo, Google Vertex, Amazon Bedrock, Novita, Microsoft Azure, Kimi.

Fastest

#1
Google VertexGoogle Vertex
175.9 t/s
#2
AzureAzure
169.2 t/s
#3
AmazonAmazon
123.7 t/s
#4
Kimi TurboKimi Turbo
104.6 t/s
#5
NovitaNovita
42.8 t/s

Output speed

Total 6 providers

Lowest Latency

#1
Google VertexGoogle Vertex
12.18 s
#2
AzureAzure
13.12 s
#3
AmazonAmazon
17.60 s
#4
Kimi TurboKimi Turbo
21.93 s
#5
NovitaNovita
49.94 s

Time to first answer token

Total 6 providers

Lowest Price

#1
Google VertexGoogle Vertex
$1.07
#2
AmazonAmazon
$1.07
#3
NovitaNovita
$1.07
#4
AzureAzure
$1.07
#5
KimiKimi
$1.07

Blended price (per 1M tokens, 3:1 Input-Output ratio)

Total 6 providers

Kimi K2 Thinking is available through 6 API providers, each offering different performance characteristics and pricing. Below is a comparison of the key metrics across providers.

  • For output speed, the top providers are Google Vertex (175.9 t/s), Azure (169.2 t/s), Amazon (123.7 t/s). Speed varies significantly across providers, with a 311% difference between the fastest and slowest.
  • For latency, Google Vertex (12.18s), Azure (13.12s), Amazon (17.60s) offer the lowest time to first token.
  • For pricing, Google Vertex (1.07), Amazon (1.07), Novita (1.07) offer the lowest blended prices per 1M tokens.
  • Google Vertex stands out as the overall leader, ranking first across all three categories: speed, latency, and pricing.

Highlights

Speed

Output tokens per second · Higher is better

End-to-End Response Time

Seconds · Lower is better

Price

USD per 1M tokens (3:1 input-output ratio) · Lower is better

Update: Default performance benchmarking workload has updated to 10k input tokens to better reflect production use cases. You can still select different workloads above.

Pricing

Pricing: Cache Hit, Input, and Output

USD per 1M tokens · Lower is better · 10,000 Input Tokens

Price per token for cached prompts (previously processed), typically offering a significant discount compared to regular input price, represented as USD per million tokens. The values shown here are the cache hit price; cache write and cache storage are billed separately and vary by provider — see "Cache pricing by provider" for detail.

Price per token included in the request/message sent to the API, represented as USD per million Tokens.

The blended bar shown here uses cache hit price only. Other caching costs differ by provider:

  • Anthropic: charges a separate cache write fee, with different rates for 5-minute and 1-hour TTLs (1-hour TTL is more expensive). Blended price charts use Anthropic cache write price for the input leg.
  • Google (Vertex/Gemini): charges a per-hour cache storage fee in addition to cache hit pricing. Some providers also use tiered pricing for prompts above 200K tokens.
  • OpenAI, DeepSeek, others: typically charge only cache hit pricing with no write or storage fee.

See Prompt Caching for the full breakdown.

Price per token generated by the model (received from the API), represented as USD per million Tokens.

Pricing: Blended Price

Blended at 7:2:1 (Cache Hit : Input : Output) · USD per 1M tokens · Lower is better

Price per token, represented as USD per million Tokens. Price is a blend of cache hit, input, and output token prices using the selected ratio.

Price per token for cached prompts (previously processed), typically offering a significant discount compared to regular input price, represented as USD per million tokens. The values shown here are the cache hit price; cache write and cache storage are billed separately and vary by provider — see "Cache pricing by provider" for detail.

The blended bar shown here uses cache hit price only. Other caching costs differ by provider:

  • Anthropic: charges a separate cache write fee, with different rates for 5-minute and 1-hour TTLs (1-hour TTL is more expensive). Blended price charts use Anthropic cache write price for the input leg.
  • Google (Vertex/Gemini): charges a per-hour cache storage fee in addition to cache hit pricing. Some providers also use tiered pricing for prompts above 200K tokens.
  • OpenAI, DeepSeek, others: typically charge only cache hit pricing with no write or storage fee.

See Prompt Caching for the full breakdown.

Figures represent median (P50) measurement over the past 72 hours to reflect sustained changes in performance.

Speed vs. Price

Blended at 7:2:1 (Cache Hit : Input : Output) · Output speed: output tokens per second · Price: USD per 1M tokens
Most attractive quadrant
Amazon
Azure
Google Vertex
Kimi
Kimi Turbo
Novita

Tokens per second received while the model is generating tokens (ie. after first chunk has been received from the API for models which support streaming).

Price per token, represented as USD per million Tokens. Price is a blend of cache hit, input, and output token prices using the selected ratio.

Price per token for cached prompts (previously processed), typically offering a significant discount compared to regular input price, represented as USD per million tokens. The values shown here are the cache hit price; cache write and cache storage are billed separately and vary by provider — see "Cache pricing by provider" for detail.

The blended bar shown here uses cache hit price only. Other caching costs differ by provider:

  • Anthropic: charges a separate cache write fee, with different rates for 5-minute and 1-hour TTLs (1-hour TTL is more expensive). Blended price charts use Anthropic cache write price for the input leg.
  • Google (Vertex/Gemini): charges a per-hour cache storage fee in addition to cache hit pricing. Some providers also use tiered pricing for prompts above 200K tokens.
  • OpenAI, DeepSeek, others: typically charge only cache hit pricing with no write or storage fee.

See Prompt Caching for the full breakdown.

Figures represent median (P50) measurement over the past 72 hours to reflect sustained changes in performance.

Speed

Measured by Output Speed (tokens per second)

Output Speed

Output tokens per second · Higher is better · 10,000 Input Tokens

Tokens per second received while the model is generating tokens (ie. after first chunk has been received from the API for models which support streaming).

Figures represent median (P50) measurement over the past 72 hours to reflect sustained changes in performance.

Latency vs. Output Speed

Latency: seconds to first token received · Output speed: output tokens per second · 10,000 Input Tokens
Most attractive quadrant
Size represents Price (USD per M Tokens)
Amazon
Azure
Google Vertex
Kimi
Kimi Turbo
Novita

Tokens per second received while the model is generating tokens (ie. after first chunk has been received from the API for models which support streaming).

Time to first token received, in seconds, after API request sent. For reasoning models which share reasoning tokens, this will be the first reasoning token. For models which do not support streaming, this represents time to receive the completion.

Price per token, represented as USD per million Tokens. Price is a blend of cache hit, input, and output token prices using the selected ratio.

Figures represent median (P50) measurement over the past 72 hours to reflect sustained changes in performance.

Latency

Measured by Time (seconds) to First Token

Time to First Answer Token

Seconds to first token received · Lower is better · 10,000 Input Tokens

Time to first answer token received, in seconds, after API request sent. For reasoning models, this includes the 'thinking' time of the model before providing an answer. For models which do not support streaming, this represents time to receive the completion.

Figures represent median (P50) measurement over the past 72 hours to reflect sustained changes in performance.

End-to-End Response Time

Seconds to output 500 tokens, calculated based on time to first token, 'thinking' time for reasoning models, and output speed

End-to-End Response Time

Seconds to output 500 tokens, including reasoning model 'thinking' time · Lower is better · 10,000 Input Tokens

Seconds to receive a 500 token response. Key components:

  • Input time: Time to receive the first response token
  • Thinking time (only for reasoning models): Time reasoning models spend outputting tokens to reason prior to providing an answer. Amount of tokens based on the average reasoning tokens across a diverse set of 60 prompts (methodology details).
  • Answer time: Time to generate 500 output tokens, based on output speed

For fair comparison, the number of reasoning tokens is standardized across all providers for each model based on the model's representative query token counts.

Figures represent median (P50) measurement over the past 72 hours to reflect sustained changes in performance.

API Features

Function (Tool) Calling & JSON Mode

ModelsFunction callingJSON Mode
Kimi Turbo logoKimi Turbo
Google Vertex logoGoogle Vertex
Amazon logoAmazon
Novita logoNovita
Azure logoAzure
Kimi logoKimi

Indicates whether the provider supports function calling in their API. Function calling is also known as 'Tool Calling'.

Indicates whether the provider supports JSON mode in their API. When JSON mode is enabled, the models will always return a valid JSON object.

Context Window

Context window: tokens limit · Higher is better

Maximum number of combined input & output tokens. Output tokens commonly have a significantly lower limit (varied by model).

While models have their own context window, in cases this is limited by providers.

Summary Table of Key Comparison Metrics

Frequently Asked Questions

Common questions about Kimi K2 Thinking providers

Kimi K2 Thinking is available through 6 API providers: Kimi Turbo, Google Vertex, Amazon, Novita, Azure, and Kimi. Each provider offers different performance characteristics and pricing.

Kimi K2 Thinking is currently available through 6 API providers that we benchmark and track.

The fastest providers for Kimi K2 Thinking by output speed are Google Vertex (175.9 t/s), Azure (169.2 t/s), and Amazon (123.7 t/s). Output speed measures how quickly tokens are generated after the model starts responding.

The providers with the lowest time to first token for Kimi K2 Thinking are Google Vertex (0.82s), Azure (1.30s), and Amazon (1.43s). Lower latency means faster initial response time.

The most affordable providers for Kimi K2 Thinking by blended price are Google Vertex ($1.07 per 1M tokens), Amazon ($1.07 per 1M tokens), and Novita ($1.07 per 1M tokens). Blended price uses a 3:1 input to output token ratio.

The providers with the lowest input token pricing for Kimi K2 Thinking are Google Vertex ($0.60 per 1M input tokens), Amazon ($0.60 per 1M input tokens), and Novita ($0.60 per 1M input tokens).

The providers with the lowest output token pricing for Kimi K2 Thinking are Google Vertex ($2.50 per 1M output tokens), Amazon ($2.50 per 1M output tokens), and Novita ($2.50 per 1M output tokens).

Prices for Kimi K2 Thinking vary up to 2.7x across providers. The most affordable is Google Vertex at $1.07 per 1M tokens, while Kimi Turbo charges $2.86 per 1M tokens.

Output speed for Kimi K2 Thinking varies significantly across providers. Google Vertex is the fastest at 175.9 t/s, which is 7.3x faster than Kimi at 24.2 t/s.

5 of 6 providers support JSON mode for Kimi K2 Thinking: Kimi Turbo, Google Vertex, Novita, Azure, and Kimi.

All 6 providers of Kimi K2 Thinking support function calling (tool use).

Google Vertex leads across all key metrics for Kimi K2 Thinking, offering the fastest speed, lowest latency, and most competitive pricing.

When choosing a provider for Kimi K2 Thinking, consider: output speed (for throughput-intensive tasks), latency (for interactive applications requiring quick first responses), pricing (for cost-sensitive workloads), and API features like JSON mode or function calling.

Yes, provider performance can vary over time due to infrastructure changes, load balancing, and updates. We continuously benchmark all providers and display historical performance trends in the "Over Time" charts.

For information about Kimi K2 Thinking's intelligence, capabilities, modalities, and how it compares to other models, see the model overview page. View model overview