AI Second Opinion — Independent Review From Another AI Lab

Give your AI agent a second opinion from a model built by a different company. Send the task and the answer your AI produced; get back a structured verdict your agent can act on: what's wrong, why it matters, how to fix it, and how confident the reviewer is.

Why a different lab

Models from the same lab tend to share blind spots. This Actor always picks a reviewer from another lab:

Answer came from	Reviewed by (default)
Claude (Anthropic)	GPT-6.1 Sol (OpenAI)
GPT, Gemini, Grok, DeepSeek, Mistral, Llama, other, or unknown	Claude Sonnet 5.5 (Anthropic)

You can choose a specific reviewer in the Advanced input — it is rejected if it comes from the same lab as the answer.

What you get

One row per review:

{
  "verdict": "disagree",
  "summary": "Revenue is miscalculated and the margin is an unsupported assumption.",
  "critical_flaws": [{
    "quote": "120 × $4.50 = $560 per day",
    "problem": "Arithmetic error: 120 × $4.50 is $540.",
    "why_it_matters": "Every downstream figure is inflated by about 3.7%.",
    "fix": "Use $540/day → $16,200/month before costs.",
    "severity": "high"
  }],
  "missed_considerations": ["Sales vary by weekday and weather", "Permits, rent, and staff costs"],
  "claims_to_verify": [{ "claim": "A typical margin is 70%", "status": "unverifiable", "note": "Depends on the cost structure." }],
  "confidence": 0.85,
  "recommendation": "Recompute revenue, then estimate costs from real inputs instead of an assumed margin.",
  "reviewer": { "provider": "anthropic", "model": "anthropic/claude-sonnet-5.5" },
  "billable": true
}

• verdict is agree, disagree (always with at least one flaw), or insufficient_evidence (with what to check).
• billable tells you whether this review was charged.

Pricing — you pay only for valid reviews

• $0.02 per valid review. A review is valid only if it passes strict checks: a real verdict, flaws with concrete fixes, a stated reason when there are no flaws, a calibrated confidence, and a reviewer from a different lab. Malformed reviews, model errors, and rejected inputs are delivered free.
• Model usage is billed separately by Apify's OpenRouter Actor to your own Apify account. On paid Apify plans that is OpenRouter's cost — about $0.003–$0.015 per review with the default reviewers ($0.008 on average in our tests). On Apify's free plan, Apify bills OpenRouter usage at 10× that rate (roughly $0.03–$0.15 per review). Model usage for a failed attempt is still billed, but our $0.02 is not.
• The Actor checks your "Maximum cost per run" before calling the model, so you never pay for a review that couldn't be delivered.

Input

Field	Required	What it does
task	yes	What the original AI was asked to do
answer	yes	The answer, plan, code, or analysis to review
context	no	Constraints, data, or sources the answer relied on
callerModelFamily	no	Which model family wrote the answer, so the reviewer comes from a different lab (claude, gpt, gemini, …, unknown)
mode	no	critique (default), fact-check, or alternatives
reviewerModel	no	A specific OpenRouter model ID to use as reviewer

Task, answer, and context together are limited to 20,000 characters.

Use it from an AI agent (MCP)

Point any MCP client at:

https://mcp.apify.com?tools=resultsmith/ai-second-opinion

Claude Code:

$claude mcp add --transport http second-opinion "https://mcp.apify.com?tools=resultsmith/ai-second-opinion"

Then tell your agent when to use it, for example: "Before you finalize a plan or a number, get a second opinion with the ai-second-opinion tool and fix anything it flags as high severity."

What it can't do

• No web access. Facts are checked against the reviewer's own knowledge. When that isn't enough it says insufficient_evidence and lists what to check, rather than guessing.
• Independent isn't the same as correct. Another lab's model can be wrong too. Treat the review as a second opinion, not ground truth.

Your data

Your task, answer, and context are sent to the reviewer model's provider through OpenRouter (via Apify's OpenRouter Actor). Don't send secrets or personal data you can't share with a model provider. The review is stored only in your run's dataset.

Feedback and usage stats

Rate a review (free). Run the Actor with feedbackRating (useful, wrong, or missing), the feedbackRunId of the review you're rating, and an optional feedbackNote. No model call, no charge. Agents can use it to report a review that missed something or flagged a non-problem.

{ "feedbackRating": "missing", "feedbackRunId": "<earlier run ID>", "feedbackNote": "Didn't catch the unit mismatch" }

Anonymous usage stats. Each run sends us counts only — verdict, whether the review was valid, model cost, duration, and how the run was started (Console, API, or a schedule) — plus a one-way hash of the input so repeated identical requests are visible. Never the task, answer, or context text. Set shareUsageStats to false to opt out.

FAQ

How do I get a second opinion on an AI agent's answer?

Send the task and the answer to this Actor (directly or as an MCP tool). A model from a different lab reviews it and returns a verdict, flaws with fixes, and a recommendation in JSON your agent can act on.

Can Claude check ChatGPT's work, or the other way round?

Yes — set callerModelFamily to the model that wrote the answer. GPT, Gemini, and other answers are reviewed by Claude; Claude answers are reviewed by OpenAI's GPT.

What does a review cost?

$0.02 for each valid review, plus the reviewer model's usage: about $0.008 on average on paid Apify plans, about 10× that on the free plan. Invalid reviews cost nothing from us.