AI gateway

Point any OpenAI or Anthropic client at Pikeway by changing two things: the base URL and the key. Every model in the model catalog works with all three formats.

OpenAI Chat Completions

curl https://gw.pikeway.ai/ai/v1/chat/completions \
  -H "Authorization: Bearer $PIKEWAY_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "gpt-5",
    "messages": [{ "role": "user", "content": "Summarize the WBTC market in one line." }]
  }'

# X-Pikeway-Cost: 0.000412
# X-Pikeway-Tokens: 18+52

With "stream": true, the response is passed through as server-sent events. The gateway asks the upstream to include usage at the end of the stream so the call can be billed by tokens.

Anthropic Messages

curl https://gw.pikeway.ai/ai/v1/messages \
  -H "Authorization: Bearer $PIKEWAY_KEY" \
  -H "anthropic-version: 2023-06-01" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "claude-sonnet-5",
    "max_tokens": 512,
    "messages": [{ "role": "user", "content": "Hello" }]
  }'

Use with Claude Code

Claude Code speaks the Messages format, so two environment variables route it through Pikeway:

export ANTHROPIC_BASE_URL=https://gw.pikeway.ai/ai
export ANTHROPIC_AUTH_TOKEN=pk-…          # your Pikeway key
claude

Claude Code requests models by their dated Anthropic names (e.g. claude-sonnet-4-5-20250929). The gateway resolves these to the canonical catalog name (claude-sonnet-4.5), so you don't need to change any model settings.

OpenAI Responses

curl https://gw.pikeway.ai/ai/v1/responses \
  -H "Authorization: Bearer $PIKEWAY_KEY" \
  -H "Content-Type: application/json" \
  -d '{ "model": "gpt-5", "input": "Write a haiku about toll roads." }'

Model names

Use the id from the catalog or from GET /ai/v1/models as model. Models also accept aliases (such as an OpenRouter-style full id or a short name), and Anthropic-style names are normalized: a date or -latest suffix is dropped and dashes between digits become dots. Usage is always recorded under the canonical name. An unknown model returns 404 unknown model.

curl https://gw.pikeway.ai/ai/v1/models

{
  "object": "list",
  "data": [{
    "id": "claude-sonnet-5",
    "object": "model",
    "owned_by": "anthropic",
    "display_name": "…",
    "context_length": 200000,
    "pricing": { "input_per_mtok": 3, "output_per_mtok": 15 },
    "capabilities": { "tools": true, "streaming": true, "thinking": true }
  }, …]
}

Token billing

• Each model has an input and an output price per million tokens (input_per_mtok, output_per_mtok). Cost = input tokens × input price + output tokens × output price.
• Input tokens served from the vendor's prompt cache are billed at the model's cache-read price when it has one.
• Token counts come from the upstream's usage. Non-streaming responses report them in X-Pikeway-Tokens (input+output) and the cost in X-Pikeway-Cost. For streams the cost is settled when the stream ends; see it in your call log.
• Requests the upstream rejects (4xx/5xx) are not billed.
• A request needs a positive balance to start; otherwise it gets 402 insufficient balance.