GPT-6.1 Sol function tool for Sume runs: clamp the cap in code

To let GPT-6.1 Sol start a Sume Agent Completion, define a Responses API function tool, run its function_call in your own code, and send back a function_call_output. In the handler, clamp generation_spend_cap_usd to a ceiling you set, use the call_id as the Idempotency-Key, and return the run id, not a finished video. The model chooses the task; your code owns the money and the key.

The OpenAI side

OpenAI's function calling guide describes a function tool with type, name, description, parameters and strict. The model returns function_call items with a call_id, a name and JSON arguments; you reply with a function_call_output carrying the same call_id and an output. GPT-6.1 Sol is priced at $2 per million input tokens and $10 per million output tokens, with cached input at $0.10, per the OpenAI community post.

The Sume side

Agent Completions accept instruction or messages, require generation_spend_cap_usd, and answer 202 with an agent.run receipt containing id, status and status_url. Replaying an Idempotency-Key returns the original receipt, which is exactly what you want when a model retries a tool call.

import json, os, urllib.request

MAX_CAP = 5.0

def start_run(args: dict, call_id: str) -> dict:
    cap = min(float(args["generation_spend_cap_usd"]), MAX_CAP)
    if not 0 < cap <= MAX_CAP:
        raise ValueError("cap must be positive")
    body = json.dumps({"instruction": args["instruction"],
                       "generation_spend_cap_usd": cap}).encode()
    req = urllib.request.Request(
        "https://api.sume.com/v1/agent/completions", data=body, method="POST",
        headers={"Authorization": "Bearer " + os.environ["SUME_API_KEY"],
                 "Content-Type": "application/json",
                 "Idempotency-Key": call_id})
    with urllib.request.urlopen(req, timeout=30) as resp:
        run = json.load(resp)["data"]
    out = {"run_id": run["id"], "status": run["status"],
           "status_url": run["status_url"]}
    return {"type": "function_call_output", "call_id": call_id,
            "output": json.dumps(out)}

Why the clamp lives in code

Strict mode makes the arguments well-typed, not wise. If the schema allows any number, a model can ask for a 50 dollar cap; min() in the handler makes that request harmless. Sume's docs describe the cap as the substitute for the interactive spend-approval prompt that a backend caller does not have, so treat it as policy you own.

Then poll, do not wait

The tool result should return in seconds, even though the run may take minutes. Have a second tool, or your own background worker, read status_url until next_action stops being poll_status, or use a run webhook. Keep the key out of logs and tool outputs, per Safe automation.

One caveat: the key needs the agent_completions:write scope, and a key created before Agent Completions shipped does not have it. Create a new one in the dashboard, as described on Authentication.