Llama 3.3 70B: API Provider Benchmarking & Analysis
Comparison Summary
Cerebras (2266 t/s) and 
Groq (326 t/s) are the fastest providers of Llama 3.3 70B, followed by 
SambaNova (300 t/s), 
Eigen AI (177 t/s) & 
Google Vertex (151 t/s).Latency (TTFT): Groq (0.19s) and Google Vertex (0.20s) have the lowest latency for Llama 3.3 70B, followed by 
Lightning AI (0.21s), Cerebras (0.26s) & 
CompactifAI (0.29s).Blended Price ($/M tokens):
DeepInfra (Turbo, FP8) ($0.15) and 
Nebius Base ($0.20) are the most cost-effective providers for Llama 3.3 70B, followed by 
Novita ($0.20), 
Parasail (FP8) ($0.28) & Lightning AI ($0.30).Input Token Price: DeepInfra (Turbo, FP8) ($0.10) and Nebius Base ($0.13) offer the lowest input token prices for Llama 3.3 70B, followed by Novita ($0.14), Nebius Fast ($0.25) & Parasail (FP8) ($0.28).Output Token Price: Parasail (FP8) ($0.28) and Lightning AI ($0.30) offer the lowest output token prices for Llama 3.3 70B, followed by DeepInfra (Turbo, FP8) ($0.32), Nebius Base ($0.40) & 
Hyperbolic ($0.40).Highlights
Pricing
Speed vs. Price: Llama 3.3 70B Providers
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 Input & Output token prices (3:1 ratio).
Figures represent median (P50) measurement over the past 72 hours to reflect sustained changes in performance.
Latency vs. Output Speed: Llama 3.3 70B Providers
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 Input & Output token prices (3:1 ratio).
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: Llama 3.3 70B Providers
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
Measured by Time (seconds) to First Token
Time to First Token: Llama 3.3 70B Providers
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.
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 vs. Price: Llama 3.3 70B Providers
Price per token, represented as USD per million Tokens. Price is a blend of Input & Output token prices (3:1 ratio).
Seconds to receive a 500 token response considering input processing time, 'thinking' time of reasoning models, and output speed.
Figures represent median (P50) measurement over the past 72 hours to reflect sustained changes in performance.
API Features
Function (Tool) Calling & JSON Mode: Llama 3.3 70B Providers
Amazon
Snowflake
Fireworks
FriendliAI
Cloudflare
Databricks
Together.ai Turbo
Azure
ScalewayIndicates 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: Llama 3.3 70B Providers
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.