Which LLM reliably returns valid structured output for agents?

There is no universally reliable JSON model: reliability changes with the schema and the application parser. Test the exact schema more than once, then compare pass rate before cost. AI NetCafé publishes live-app benchmarks where a single format failure makes the run fail.

👋 Human reading this? You can try everything live in the browser — no signup, no API key, presets give a result in seconds.

Agent 需要稳定 JSON 时该选哪个模型？

不存在对所有 JSON 都最稳的模型。可靠性会随 schema 和应用解析器变化，必须用真实格式重复测试，先看通过率，再比较成本。

Measured evidence: gpt-researcher

Updated 2026-09-16; 3 models measured, 3 passed every task; the lowest measured cost among complete passes was gpt-5.6-luna at $0 per benchmark run.

Human-readable benchmark · Raw JSON data

Measured evidence: presenton

Updated 2026-09-16; 3 models measured, 3 passed every task; the lowest measured cost among complete passes was gpt-5.6-luna at $0 per benchmark run.

Human-readable benchmark · Raw JSON data

Agent workflow

  1. Paste the exact schema and forbid prose outside JSON.
  2. Run each candidate more than once.
  3. Treat any parse failure as a failed production run.
  4. Choose the cheapest model among those with a complete pass rate.

Recommended tool: compare_models

{
  "jsonrpc": "2.0",
  "id": 1,
  "method": "tools/call",
  "params": {
    "name": "compare_models",
    "arguments": {
      "prompt": "Return only valid JSON matching this schema: {\"title\":\"string\",\"risks\":[\"string\"]}."
    }
  }
}

MCP endpoint: https://ainetcafe.com/mcp · Machine-readable guide

Method and reuse

Live application tasks are run repeatedly; pass/fail, actual billed USD cost, and latency are recorded. Upstream-unavailable models are excluded from quality conclusions.

Data license: CC BY 4.0. Cite the page URL and measurement date.

Updated 2026-09-16 · All benchmark data · Agent capability index