# Config file for llm: https://github.com/simonw/llm 
# Prices vary for respective API providers. Default prices may not reflect current prices.

# [Structure template]
# {PROVIDER}:
#   {MODEL_NAME}:
#     model_family: {MODEL_FAMILY_NAME}
#     model_alias: {MODEL_API_NAME}
#     model_context_length: {MODEL_WINDOW}
#     model_params:
#       {custom parameter arguments}. For example:
#       max_completion_tokens: 
#       think: (OPTIONAL)
#       temperature:
#       top_p:
#       stop:

# GEMINI
google:
  gemma-4-31b-it:
    model_family: gemma4
    model_alias: gemma-4-31b-it
    model_context_length: 256000
    model_params:
      max_output_tokens: 16384 # max: 65536
      temperature: 1.0
      top_p: 1.0
      # thinking_level: medium <- temporarily unavailable
    cost_usd_mtok:
      input: 0.00
      output: 0.00
  gemma-4-26b-a4b-it:
    model_family: gemma4
    model_alias: gemma-4-26b-a4b-it
    model_context_length: 1048576
    model_params:
      max_output_tokens: 16384 # max: 65536
      temperature: 1.0
      top_p: 1.0
      # thinking_level: medium <- temporarily unavailable
    cost_usd_mtok:
      input: 0.00
      output: 0.00
  gemini-3.1-pro-preview:
    model_family: gemini-3
    model_alias: gemini-3.1-pro-preview
    model_context_length: 1048576
    model_params:
      max_output_tokens: 16384 # max: 65536
      temperature: 1.0
      top_p: 1.0
      thinking_level: medium # low/medium/high
    cost_usd_mtok:
      input: 2.00
      output: 12.00
  gemini-3.1-flash-lite-preview:
    model_family: gemini-3
    model_alias: gemini-3.1-flash-lite-preview
    model_context_length: 1048576
    model_params:
      max_output_tokens: 16384 # max: 65536
      temperature: 1.0
      top_p: 1.0
      thinking_level: medium # low/medium/high
    cost_usd_mtok:
      input: 0.25
      output: 1.50
  gemini-3-pro-preview:
    model_family: gemini-3
    model_alias: gemini-3-pro-preview
    model_context_length: 1048576
    model_params:
      max_output_tokens: 16384 # max: 65536
      temperature: 1.0
      top_p: 1.0
      thinking_level: medium # low/medium/high
    cost_usd_mtok:
      input: 2.00
      output: 12.00
  gemini-3-flash-preview:
    model_family: gemini-3
    model_alias: gemini-3-flash-preview
    model_context_length: 1048576
    model_params:
      max_output_tokens: 16384 # max: 65536
      temperature: 1.0
      top_p: 1.0
      thinking_level: medium # low/medium/high
    cost_usd_mtok:
      input: 0.50
      output: 3.00
  gemini-2.5-pro:
    model_family: gemini-2.5
    model_alias: gemini-2.5-pro
    model_context_length: 1048576
    model_params:
      max_output_tokens: 16384 # max: 65,536
      temperature: 1.0
      top_p: 1.0
      thinking_budget: 8192 # ~medium
    cost_usd_mtok:
      input: 1.25
      output: 10.00
  gemini-2.5-flash:
    model_family: gemini-2.5
    model_alias: gemini-2.5-flash
    model_context_length: 1048576
    model_params:
      max_output_tokens: 16384 # max: 65,536
      temperature: 1.0
      top_p: 1.0
      thinking_budget: 8192 # ~medium
    cost_usd_mtok:
      input: 0.30
      output: 2.50
  gemini-2.5-flash-lite:
    model_family: gemini-2.5
    model_alias: gemini-2.5-flash-lite
    model_context_length: 1048576
    model_params:
      max_output_tokens: 16384 # max: 65,536
      temperature: 1.0
      top_p: 1.0
      thinking_budget: 8192 # ~medium
    cost_usd_mtok:
      input: 0.10
      output: 0.40
  gemini-2.0-flash:
    model_family: gemini-2.0
    model_alias: gemini-2.0-flash
    model_context_length: 1048576
    model_params:
      max_output_tokens: 8192
      temperature: 0.0
      top_p: 1.0
    cost_usd_mtok:
      input: 0.10
      output: 0.40
  gemini-2.0-flash-lite:
    model_family: gemini-2.0
    model_alias: gemini-2.0-flash-lite
    model_context_length: 1048576
    model_params:
      max_output_tokens: 8192
      temperature: 0.0
      top_p: 1.0
    cost_usd_mtok:
      input: 0.075
      output: 0.30