Claude Haiku 4.5

Overview

Start with Claude Haiku 4.5

The model name is already filled in. Mint a key in the console, put it in an environment variable, and the call below runs as it stands — provided the organization’s wallet holds funds: a call with nothing to pay with answers 402.

curl https://api.kumorouter.com/v1/chat/completions \
  -H "Authorization: Bearer $KUMO_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
  "model": "anthropic/claude-haiku-4-5",
  "max_tokens": 1024,
  "messages": [
    {
      "role": "user",
      "content": "Explain tokens in one line."
    }
  ]
}'

The Anthropic-shaped call

curl https://api.kumorouter.com/v1/messages \
  -H "x-api-key: $KUMO_API_KEY" \
  -H "anthropic-version: 2023-06-01" \
  -H "Content-Type: application/json" \
  -d '{
  "model": "anthropic/claude-haiku-4-5",
  "max_tokens": 1024,
  "messages": [
    {
      "role": "user",
      "content": "Explain tokens in one line."
    }
  ]
}'

OpenAI — API parameters

Authorization: Bearer $KUMO_API_KEY

One completion per request. The gateway rejects frequency_penalty, presence_penalty, seed, logprobs, top_logprobs and logit_bias: each of them would change the generation or the shape of the reply, and accepting one silently would misdescribe the answer you got. n is accepted only at 1 — the protocol default and exactly what the gateway does — and an n above one is rejected. user is accepted but is not forwarded to the model, not stored and not echoed back. The same applies to strict on a function declaration and name on a tool-role message: both are accepted and not forwarded. reasoning_effort is accepted and not forwarded either: no supplier request on this platform carries a reasoning budget, so the answer comes back at the default of the model itself whatever the field says. cache_control on a message, on a content part or on a replayed tool call is accepted and not forwarded either: the supplier wire this protocol is served by has no member for a prompt-cache anchor, so the cached-token counts come back as though the anchor had not been sent. messages[].content is accepted both as a string and as a list of parts: text parts are joined in order into one text, so a request written in parts and the same request written as a string are one request. A part of kind image_url, input_audio, file or refusal is rejected naming its own type: no multimodal capability is verified on this platform for any pair, and a replayed refusal would reach the supplier as ordinary assistant speech. The developer role is accepted and forwarded as system — it is the newer name this protocol gives the instruction role. stop is array-only; tool_choice is the string mode auto, none or required, or an object naming a declared tool. json_object is refused by name: promising valid JSON says nothing about its shape.

Anthropic — API parameters

x-api-key: $KUMO_API_KEY

Uses the native Messages format. OpenAI fields frequency_penalty, presence_penalty, n, seed and response_format do not belong to it. metadata, thinking, output_config and context_management are accepted but not forwarded to the model and are outside verified support. The same applies to tool declaration fields eager_input_streaming, strict, defer_loading, allowed_callers, input_examples and max_uses; server-side tools are unsupported. Tool input_schema is carried as the original JSON Schema object. cache_control: {type: "ephemeral"} is supported on text blocks and tool declarations; cache availability depends on the model.

Identifiers

Capabilities

A row describes one protocol-and-modality pair rather than the model as a whole: a capability is proved on a particular surface, and the answer beside it may differ.

How it compares

Prices

The whole price list →

Integrations

Every link opens its own guide with this model already chosen.

Errors Rate limits Models