AIs Have Egos: What Happened When 6 Frontier Models Graded Each Other

They gave themselves higher marks than their peers did. They got more generous as soon as they saw a famous name. They favored whoever spoke first. And two models from the same family graded so much alike that they could pass for one judge.

These aren’t anecdotes. They come from 3,570 peer scores collected in the October 2026 PeerRank run, and the numbers are below.

The setup: six models, three jobs each

Every model played three roles. It wrote questions, it answered everyone’s questions, and it graded everyone’s answers.

1. Ask. Each model wrote 20 questions across five categories (creative, current events, factual, practical, reasoning), for 120 in total.
2. Answer. All six models answered all 120 questions. Current-events questions came with the same web search results for every model.
3. Judge. Each model scored every answer from 1 to 10, three separate times:

• Blind + shuffled: names hidden, order random. This is the fair baseline, called the peer score.
• Names shown: order random, model names visible.
• Fixed order: names hidden, answers always in the same order.

The models were GPT-6 Astra, Claude Opus 5.5, Claude Fable 5.1, Gemini 3.8 Flash, Grok 4.7 and DeepSeek V4.1 Flash. Comparing the three judging modes shows how much a model’s identity and its position changed its grade.

Ego: five of six models liked their own work best

GPT-6 Astra scored its own answers 0.76 points higher than the other five models scored them. On a 10-point scale, that moves an answer from “good” to “great” based only on who wrote it.

Only two models came close to judging themselves honestly. Claude Opus 5.5 inflated itself by under a tenth of a point. Grok 4.7 actually rated itself slightly lower than its peers did, and it was also the only model that did significantly worse on questions it had written itself (−0.64 points). It was writing questions it couldn’t answer.

Brand bias: a famous name is worth points

All six models scored higher once judges could see the model’s name. The answers stayed exactly the same. The only thing that changed was the label next to them.

The brand premium wasn’t spread evenly. The three top-ranked models got the biggest boosts, so reputation added to a lead that already existed. AI judges, like human ones, give the benefit of the doubt to names they already trust.

Position bias: it pays to go first

Only the answer shown first got a boost. With names hidden and the order fixed, the first answer gained 0.13 points over its fair score, while answers in slots 3, 4 and 5 lost between 0.16 and 0.19.

People do the same thing. Interviewers, wine tasters and talent-show judges all reward whoever shows up first. The models seem to have picked up the habit along with everything else they learned from us.

Family resemblance: the two Claudes judge like twins

Claude Opus 5.5 and Claude Fable 5.1 agreed with each other more than any other pair of judges did (r = 0.89). Every pairing across different companies landed between 0.56 and 0.80.

This isn’t loyalty, since neither Claude knew which answers were the other’s. It’s shared upbringing: models from the same lab tend to share a sense of what a good answer looks like.

The twist: the toughest critic came out on top

The model that graded others most harshly is the one its peers rated highest. Claude Opus 5.5 gave the lowest average score of the six judges (7.50), barely inflated its own score, and still finished first. DeepSeek V4.1 Flash was the most generous judge (8.27) and finished fifth.

The top three are separated by less than a quarter of a point, and the overall winner doesn’t win every category. GPT-6 Astra led in creative, practical and current-events questions, while Opus led in factual and reasoning questions. Current events were the hardest category for every model, even with web search: the average score for “who won the most recent FIFA World Cup” was just 3.9 out of 10.

What this means if you let AI judge AI

AI already grades AI everywhere: eval pipelines, agent self-review, reward models, “pick the best of five drafts.” These results show a single AI judge carries the same biases a single human judge would. Here are four practical rules:

1. Never let a model grade itself. A self-score can run three-quarters of a point high.
2. Hide the names. Brand alone was worth up to a quarter of a point.
3. Shuffle the order. Coming first is an advantage that has nothing to do with quality.
4. Use judges from different families. Two judges from the same lab act like one judge counted twice.

The bigger lesson: there’s no single best AI, only the best AI for a specific task, judged by a panel you can trust. That’s why at Caura we build to stay model-agnostic. The best model this month may not be the best next month, and your agents shouldn’t be stuck with either one.