I almost made a mistake in a paragraph of my previous post
When writing about "subagent vs. teammate" in the previous post, I asked both Codex and Gemini to independently check the official documentation. Gemini made a specific claim: "Agent teams must use the Opus 4.6+ model."Neither Claude nor Codex mentioned this. After comparing the three, I ruled it out with a 2-1 vote and didn’t include that section in the article. Upon reviewing the official documentation, I found that Anthropic doesn’t have this strict requirement—it was a hallucination generated by Gemini.
This time, comparing multiple models saved me. But consider this from another angle—what if I didn’t have Codex or Gemini at my disposal and instead ran three queries using Claude’s subagent? Since all three come from the same base model and share the same training data biases, the probability that they would all say “Opus is required” is quite high. Three “consistent answers” do not equate to “the correct answer”; they only mean “Claude is consistent with itself.”
This is why multi-LLM systems cannot be replaced by subagents from the same model.
Why Codex and Gemini (A Cost-Driven Opportunity)
Before discussing the pattern, let me explain why I happen to have these two external LLMs on hand. Simply saying “use Codex and Gemini to save tokens” is a red herring—if they’re more expensive than Claude, outsourcing to them would actually be a waste of money. Actual setup:
Codex (OpenAI): New subscriptions come with a 1-month free trial. Conducting research and code reviews during the trial period incurs virtually zero marginal cost
Gemini (Google): I signed up during last year’s Google AI Pro annual subscription promotion. The price was significantly lower than Claude’s subscription plans, and it came with an extra 5TB of Google One storage (something I was planning to buy anyway). When you break it down, Gemini is essentially a “free bonus with Google One.”
The combined monthly cost of these two is still lower than adding another Claude Code subscription (by the way: according to the latest news, the $20 tier is reportedly being phased out soon), and since they are the flagship products of OpenAI and Google, respectively, their quality is on par with the top tier.
Premise of this article: Claude Code remains the primary tool. Codex and Gemini are not replacements, but rather supplements for research and cross-validation. All patterns discussed in this article are based on this premise.
Self-consistency is not the same as cross-validation
Running the same model N times is academically referred to as self-consistency(Wang et al. 2022, arxiv 2203.11171). It works well for math and logic—topics with “sharp answer distributions”—where you can simply resample and take the majority. But for fact-checking and hallucination detection, running the same model N times often leads to consistent errors because they share the same set of biases.
Switching to a three-party system—Claude + Codex + Gemini—is what yields true diversity: different training data, different reasoning styles, and different error patterns. Agreement among the three parties indicates high confidence; disagreement provides a clue worth investigating.
Pattern 1: Delegate — Outsource the entire task
The most basic use case. Hand the task off to an external LLM; how it thinks, how many rounds it runs, or how many documents it reads doesn’t consume your Claude’s context—only the final report comes back.
Actual numbers: In my previous post, I sent a claude-code-guide subagent to look up the official documentation for agent teams. That subagent consumed 63K tokens (reasoning + 5 tool calls + crawling the official docs).If I switched to running it directly on Codex: my prompt was ~500 tokens, the returned report was ~1.5K tokens, totaling ~2K tokens added to my Claude context. How many tokens Codex consumes internally is Codex’s problem, not mine (and since Codex is still in the trial period, this cost me nothing).
Claude context savings: 30x.
Suitable for: Long reasoning, research-heavy tasks, and tasks where you don’t need to see the intermediate steps—just the conclusion.
Pattern 2: Verify — Independent Cross-Validation
Instead of outsourcing the task, have multiple LLMs each perform the same task once, then compare their results. Key rules:
Independent prompts: Each LLM doesn’t know what the others are doing or what you’ve already found
No priming: Do not imply in the prompt that “I think the answer is X”
Handling Disagreements: When there is a discrepancy among the three, refer to the source of truth (official documents, source code) for a ruling
This isn’t a waste of tokens—it’s to preserve diversity. If you feed Claude’s preliminary conclusion to Codex as context, Codex will tend to confirm the conclusion you already have (an observation from SelfCheckGPT, arxiv 2303.08896).
Suitable for: Fact-sensitive claims (version numbers, feature support, API signatures), premises for technical decisions, and content intended for public release.
Pattern 3: Vote — Majority Decision
Verify is about “gathering opinions,” while Vote is about “making a decision.” Upgrade to this when you need a definitive answer rather than a reference report.
Practical rules for M=3:
2-of-3 majority decision: The majority wins
1-1-1 (three different answers): Return to SoT or add a fourth party
High-Risk Scenarios: Switch to "unanimity-or-escalate"—if the three parties disagree, escalate immediately; majority decisions are not accepted
Example: The "Must use Opus 4.6+" requirement for Gemini in the previous post was rejected 2-1 at this stage.
Suitable for: architectural decisions, version selection, security/compliance-related judgments.
Biggest pitfall: False Consensus
All three models were trained on a large amount of overlapping network data. The degree to which they “differ” is lower than expected. Research shows that when two models are both wrong, the error rate can reach 60%+—not only do they get it wrong together, but they often arrive at the exact same answer.
Three approaches to reducing false consensus:
Provide different sources: Have Claude read documentation, Codex read research papers, and Gemini perform web searches. These three data sources inherently offer diversity.
Prevent them from seeing each other’s drafts: Consensus should emerge as a post-processing result, not be induced during the prompt phase
Preserve dissent: When 2 out of 3 agree but one disagrees, spend 30 seconds reviewing the dissenting opinion. That dissenting view might actually be correct
Token Billing Breakdown
The billing boundaries for Claude Code and MCP are actually quite clear:
Billed to Claude: Your prompt to Claude + input for MCP tool calls + returned output
Not included in Claude’s billing: internal reasoning of external LLMs, tool calls, and temperature search
MCP Limits: A warning is issued if the response exceeds 10K tokens; the default max is 25K. If it’s too large, it will be truncated or written to a file
Hidden Costs: External LLMs have their own API billing. However, as mentioned earlier—Codex is free during the trial period, and Gemini is included with Google One—so this “other wallet” is practically empty. If you’re a token-based API user, you’ll need to account for this separately.
When Not to Use It
Simple tasks with fixed answers: “Which decorator should I use for this function?”—Claude can handle this on its own
Tasks requiring project context: External LLMs cannot access your CLAUDE.md, MCP skills, or file contents. Pushing the entire repo over is too costly
Low-latency requirements: Third-party serialization plus round-trip time will result in a painful user experience
Significant model strength disparity: The Self-MoA paper (arXiv 2502.00674) shows that adding a significantly weaker model actually lowers accuracy. It’s not a case of “more models is better”
Practical Recipe
Delegate-first: Have Codex write the draft, and Claude finalize it. Suitable for: technical articles, code reviews, and long-form summaries
Verify-on-publish: Have a third party verify key facts before content is published. Suitable for: Technical claims intended for public release
Vote-for-decision: Use "unanimity-or-escalate" for architectural decisions. Suitable for: One-time, high-stakes decisions
Conclusion
This article—and the previous one—were both written using this set of patterns. Codex handles technical verification, Gemini handles web research, and Claude handles integration and drafting. Only after cross-checking the opinions of all three parties did I decide which claims to include and which to exclude.
The tools are all right there, and I’ve already paid for them (or am still in the trial period). The cost of not using them isn’t immediately apparent—it only becomes evident when you publish an incorrect claim and readers point out the error. Cross-checking across multiple models is worthwhile because it catches errors while there’s still time to fix them.