litellm_config.yaml

# LiteLLM Proxy Server config — routes through the blockrun-litellm sidecar.
#
# 1) Start the sidecar with ONE credential:
#       blockrun-litellm-proxy --port 4001 --api-key brk_live_...   # account credit
#    or, to pay per call from an x402 wallet (Solana is the default chain):
#       export SOLANA_WALLET_KEY=YOUR_SOLANA_PRIVATE_KEY
#       blockrun-litellm-proxy --port 4001
#
# 2) Start LiteLLM Proxy with this file:
#       litellm --config litellm_config.yaml --port 4000
#
# 3) Test:
#       curl http://localhost:4000/v1/chat/completions \
#         -H "Content-Type: application/json" \
#         -d '{"model": "gpt-5.6-terra", "messages": [{"role":"user","content":"Hi"}]}'
#
# Note: the model string takes the form `openai/<blockrun-model-id>` so LiteLLM
# treats our proxy as a generic OpenAI-compatible endpoint and forwards the
# inner provider prefix (e.g. "openai/", "anthropic/", "google/") verbatim.
#
# ──────────────────────────────────────────────────────────────────────────
# ⚠️  COST / SPEND TRACKING — REQUIRED for budgets to mean anything
# ──────────────────────────────────────────────────────────────────────────
# LiteLLM computes per-call spend (and enforces team/key budgets like
# `max_budget`) from `input_cost_per_token` × prompt_tokens +
# `output_cost_per_token` × completion_tokens. The blockrun model ids
# (`anthropic/claude-opus-4.8`, `openai/gpt-5.5`, …) are NOT in LiteLLM's
# built-in `model_cost` map, so WITHOUT the explicit `*_cost_per_token` values
# below LiteLLM prices these calls at $0 — a key/team `max_budget` (e.g. $200)
# then NEVER trips and the dashboard under-reports spend.
#
# The values below are each provider's official LIST price (what BlockRun's
# gateway also bills, 1:1) converted to $/token ($X per 1M ÷ 1e6). The sidecar
# forwards the gateway's REAL upstream token counts, so cost = list price ×
# actual tokens. Keep these in sync with the gateway's price table; update when
# a model's list price changes.
#
# ACCURACY CAVEATS — this view is for BUDGET ENFORCEMENT. The gateway-side
# reconciliation report (blockrun-sol /api/internal/reconciliation) is the
# EXACT cost source of truth, including prompt-cache discounts:
#  • Prompt-cache pricing: `cache_read_input_token_cost` /
#    `cache_creation_input_token_cost` in litellm_params are honored only on
#    newer LiteLLM versions. On older proxies (e.g. 1.70.0) these keys are
#    SILENTLY DROPPED and cache tokens are priced at the full
#    `input_cost_per_token` — a CONSERVATIVE over-count of cache reads (the
#    budget trips earlier, never later). Verify your proxy version, or rely on
#    the reconciliation report for exact cost.
#  • Streaming spend is exact only if the gateway emits a final usage frame in
#    the SSE stream; otherwise LiteLLM falls back to tiktoken ESTIMATES.
#    Non-streaming always prices off the gateway's real usage block.
# ──────────────────────────────────────────────────────────────────────────

model_list:
  - model_name: gpt-5.6-terra
    litellm_params:
      model: openai/openai/gpt-5.6-terra
      api_base: http://localhost:4001/v1
      api_key: "dummy"   # unused unless BLOCKRUN_PROXY_TOKEN is set on the sidecar
      input_cost_per_token: 0.0000025   # $2.50 / 1M
      output_cost_per_token: 0.000015   # $15 / 1M

  - model_name: claude-opus-5
    litellm_params:
      model: openai/anthropic/claude-opus-5
      api_base: http://localhost:4001/v1
      api_key: "dummy"
      input_cost_per_token: 0.000005    # $5 / 1M
      output_cost_per_token: 0.000025   # $25 / 1M
      cache_read_input_token_cost: 0.0000005      # $0.50 / 1M
      cache_creation_input_token_cost: 0.00000625 # $6.25 / 1M

  - model_name: gpt-5-mini
    litellm_params:
      model: openai/openai/gpt-5-mini
      api_base: http://localhost:4001/v1
      api_key: "dummy"
      input_cost_per_token: 0.00000025  # $0.25 / 1M
      output_cost_per_token: 0.000002   # $2 / 1M

  - model_name: claude-fable-5
    litellm_params:
      model: openai/anthropic/claude-fable-5
      api_base: http://localhost:4001/v1
      api_key: "dummy"
      # claude-fable-5 is redirected to Opus 4.8 upstream; price at Opus 4.8.
      input_cost_per_token: 0.000005    # $5 / 1M
      output_cost_per_token: 0.000025   # $25 / 1M
      cache_read_input_token_cost: 0.0000005     # $0.50 / 1M (0.1x input)
      cache_creation_input_token_cost: 0.00000625 # $6.25 / 1M (1.25x input)

  - model_name: claude-haiku-4-5
    litellm_params:
      model: openai/anthropic/claude-haiku-4-5
      api_base: http://localhost:4001/v1
      api_key: "dummy"
      input_cost_per_token: 0.000001    # $1 / 1M
      output_cost_per_token: 0.000005   # $5 / 1M
      cache_read_input_token_cost: 0.0000001     # $0.10 / 1M
      cache_creation_input_token_cost: 0.00000125 # $1.25 / 1M

  - model_name: gemini-3.1-pro
    litellm_params:
      model: openai/google/gemini-3.1-pro
      api_base: http://localhost:4001/v1
      api_key: "dummy"
      input_cost_per_token: 0.000002    # $2 / 1M
      output_cost_per_token: 0.000012   # $12 / 1M

  # ── Image / video models ────────────────────────────────────────────────
  # Media models are NOT in LiteLLM's price map either. Images bill as
  # `input_cost_per_pixel × width × height × n` (flat per-image price ÷
  # 1024×1024 = 1048576); video bills as `output_cost_per_second × seconds` —
  # always pass `seconds` on POST /v1/videos or the call is priced $0.
  # Video routes through the sidecar's OpenAI-compatible Videos API
  # (POST /videos → poll GET /videos/{id} → GET /videos/{id}/content),
  # available since blockrun-litellm 0.6.0.
  - model_name: grok-imagine-image
    litellm_params:
      model: openai/xai/grok-imagine-image
      api_base: http://localhost:4001/v1
      api_key: "dummy"
      input_cost_per_pixel: 1.9073486328125e-08   # $0.02/image ÷ 1048576 px
    model_info:
      mode: image_generation

  - model_name: grok-imagine-image-pro
    litellm_params:
      model: openai/xai/grok-imagine-image-pro
      api_base: http://localhost:4001/v1
      api_key: "dummy"
      input_cost_per_pixel: 6.67572021484375e-08  # $0.07/image ÷ 1048576 px
    model_info:
      mode: image_generation

  - model_name: grok-imagine-video
    litellm_params:
      model: openai/xai/grok-imagine-video
      api_base: http://localhost:4001/v1
      api_key: "dummy"
      output_cost_per_second: 0.05                # $0.05/second
    model_info:
      mode: video_generation

  # Wildcard fallback — any BlockRun model id passed as the model name will be
  # forwarded as-is. Useful if your app already speaks BlockRun model names.
  # ⚠️  Calls matched by this wildcard are NOT priced (no per-token cost), so
  # they contribute $0 to spend/budgets. Add an explicit model_list entry with
  # `input_cost_per_token`/`output_cost_per_token` for any model you want
  # counted against a budget.
  - model_name: blockrun/*
    litellm_params:
      model: openai/*
      api_base: http://localhost:4001/v1
      api_key: "dummy"

litellm_settings:
  # BlockRun ignores most OpenAI tuning knobs (frequency_penalty, etc.) —
  # let LiteLLM drop them silently instead of warning on every request.
  drop_params: True

  # Exact chat spend, no callback needed: since blockrun-litellm 0.6.0 the
  # sidecar returns the real settled x402 charge in the
  # `x-litellm-response-cost` response header, which LiteLLM reads off
  # openai-compatible upstreams and uses as `response_cost` directly (it
  # overrides the per-token estimate above when present). The per-model
  # `*_cost_per_token` values remain as the fallback for LiteLLM versions or
  # call paths that don't capture upstream headers.

general_settings:
  # Bind LiteLLM Proxy here; clients talk to LiteLLM, LiteLLM talks to our sidecar.
  master_key: "sk-litellm-master-key-change-me"