Docs

task-models.com is an OpenAI- and Anthropic-compatible endpoint for long-running economic work. Point any SDK at https://task-models.com/v1, use a key from your dashboard, and pick a model. The stream field is still named mergemodels.

Models

Quickstart

curl

curl https://task-models.com/v1/chat/completions \
  -H "Authorization: Bearer $MM_API_KEY" -H "Content-Type: application/json" \
  -d '{"model": "merge", "stream": true, "reasoning_effort": "high",
       "messages": [{"role": "user", "content": "Write a 90-day operating plan for a three-person logistics desk, with weekly milestones."}]}'

Python, OpenAI SDK

import os
from openai import OpenAI

client = OpenAI(base_url="https://task-models.com/v1", api_key=os.environ["MM_API_KEY"])
stream = client.chat.completions.create(
    model="merge", stream=True,
    messages=[{"role": "user", "content": "Compare three freight lanes for a regional distributor and recommend one."}],
    extra_body={"mergemodels": {"debug": True}},
)
for chunk in stream:
    extra = chunk.model_extra or {}
    if "mergemodels" in extra and extra["mergemodels"]["event"] == "swap":
        print(f"\n[swapped to {extra['mergemodels']['to']['label']}: {extra['mergemodels']['trigger']}]")
    for choice in chunk.choices:
        print(choice.delta.content or "", end="", flush=True)

Anthropic Messages

curl https://task-models.com/v1/messages \
  -H "x-api-key: $MM_API_KEY" -H "anthropic-version: 2023-06-01" -H "Content-Type: application/json" \
  -d '{"model": "merge", "max_tokens": 4096, "messages": [{"role": "user", "content": "..."}]}'

Claude Code works too: ANTHROPIC_BASE_URL=https://task-models.com, ANTHROPIC_AUTH_TOKEN=$MM_API_KEY, ANTHROPIC_MODEL=merge.

How a continuation looks in the stream

A continuation never breaks a client: text keeps arriving in the same delta.content / delta.reasoning_content fields. For clients that want to know, the OpenAI stream carries one extra chunk at the swap point, with an empty delta and a mergemodels object:

{"id": "chatcmpl-…", "object": "chat.completion.chunk", "model": "merge",
 "choices": [{"index": 0, "delta": {}, "finish_reason": null}],
 "mergemodels": {"event": "swap", "trigger": "false_refusal", "confidence": 0.91,
                 "harm": "benign", "reason": "refusal of a benign request", "source": "classifier",
                 "phase": "content", "at_tokens": 40,
                 "from": {"model": "deepseek/deepseek-v4.1-flash", "label": "DeepSeek V4.1 Flash"},
                 "to":   {"model": "abliterated-model", "label": "Continuation"}}}

• phase: request (before any output), reasoning or content: where the official model was when the task stopped.
• harm is benign when the task continues. source is classifier, or upstream_error when the continuation failed and the official model finished the answer.
• The chunk with finish_reason carries "mergemodels": {"event": "summary", …}: segments, swaps, checks and the cost in USD. Usage includes cost.
• Non-streaming responses carry the same summary as a top-level mergemodels field, plus an x-mergemodels-swapped header.
• Anthropic streams send these as SSE comment lines (: mergemodels {…}), which SDKs ignore.

Request options

When a long task continues

Spot price and billing

The price changes with the market. GET /v1/pricing returns the live price (no key needed); GET /v1/pricing/history lists recent changes.

{"object": "spot_price", "id": 1, "currency": "usd", "unit": "1M tokens",
 "prices": {"input": 0.42, "cached_input": 0.042, "output": 1.68,
            "advanced_input": 1.47, "advanced_output": 4.41}}

• A request is billed at the spot price that was live when it started.
• On merge and merge-fast, your prompt is billed once at input (cached tokens at cached_input) and every token delivered to you is billed at output. A continuation does not add a second price.
• advanced_input and advanced_output are the merge-pro rates, for calling the controlled abliterated model from the first token.
• Failed requests are free. If you disconnect, you pay for what was delivered until then. An early stop that a continuation replaces is not delivered and is not billed.

Watching requests

GET /v1/live lists your in-flight requests; GET /v1/live/<id> is an SSE stream of one of them: reasoning, answer, every classifier verdict and the swap, as they happen. The dashboard shows the same view.

Errors