# DeepSeek — https://deepseek.com
# Models: 4

[provider]
id = "deepseek"
display_name = "DeepSeek"
api_key_env = "DEEPSEEK_API_KEY"
base_url = "https://api.deepseek.com/v1"
key_required = true

[[models]]
id = "deepseek-v4-pro"
display_name = "DeepSeek V4 Pro"
tier = "frontier"
context_window = 1000000
max_output_tokens = 384000
input_cost_per_m = 0.435
output_cost_per_m = 0.87
supports_tools = true
supports_vision = false
supports_streaming = true
supports_thinking = true
aliases = ["deepseek-v4", "deepseek-pro"]

[[models]]
id = "deepseek-v4-flash"
display_name = "DeepSeek V4 Flash"
tier = "smart"
context_window = 1000000
max_output_tokens = 384000
input_cost_per_m = 0.14
output_cost_per_m = 0.28
supports_tools = true
supports_vision = false
supports_streaming = true
supports_thinking = true
# Thinking mode is on by default and the API rejects multi-turn requests
# where assistant turns containing tool_calls don't echo back the original
# reasoning_content. See librefang/librefang#4842.
reasoning_echo_policy = "echo"
aliases = ["deepseek-flash"]

[[models]]
id = "deepseek-chat"
display_name = "DeepSeek V3"
tier = "smart"
context_window = 64000
max_output_tokens = 8192
input_cost_per_m = 0.32
output_cost_per_m = 0.89
supports_tools = true
supports_vision = false
supports_streaming = true
aliases = ["deepseek", "deepseek-v3"]

[[models]]
id = "deepseek-reasoner"
display_name = "DeepSeek R1"
tier = "frontier"
context_window = 64000
max_output_tokens = 8192
input_cost_per_m = 0.55
output_cost_per_m = 2.19
supports_tools = false
supports_vision = false
supports_streaming = true
supports_thinking = true
# R1 returns reasoning_content in responses but the API rejects multi-turn
# requests that carry it on previous assistant messages — drivers must strip.
reasoning_echo_policy = "strip"
aliases = ["deepseek-r1"]