Ask GPT-6.1 Sol, Claude Opus 5.5 or Gemini 4 Argon "Who Is This Researcher?" — and Watch Argon Reach for the Wrong Lab

Three frontier models were asked one question each, 448 times, with nothing but a name to work from. GPT-6.1 Sol and Claude Opus 5.5 identified almost every AI researcher they were shown. Gemini 4 Argon identified nearly as many people but described fewer of them correctly, and the reported reason is specific and useful: it repeatedly named a researcher's former employer as their current one. The evaluation was posted on 14 October 2026 by the account @scaling01; it is one posted test, not an independent benchmark, and nobody has reproduced it. The three models have very different ages — Gemini 4 Argon went generally available on the paid tier on 13 October 2026 after being announced on 30 September, Claude Opus 5.5 shipped 22 September, GPT-6.1 Sol 29 September — so the only thing genuinely new here is the finding itself, which lands two days after Argon's paid tier opened.

This is a stale-attribute problem, and it is worth understanding properly, because it is not the same class of failure as a hallucinated citation or a confidently invented paper. A model that names a researcher's former employer is not making something up. It is returning a fact that was true, from a snapshot that has moved on. The cited internal reasoning for the gap is that the "better" a model is at citing outdated information as current, the more useful it is for some tasks and the more dangerous for this one. That is a clean formulation of the trade, and it is the reason this particular evaluation is more interesting than its sample size suggests.

What the post actually measured, in its own words: 448 real researchers were shown to each model as a bare name, once, with no web access and no tools. Answers were checked against public sources — personal pages, lab bios, papers — on 14 October 2026. The chart separates two things that are usually collapsed into one score. The full bar is "recognized the right person." The solid part of the bar is "the whole answer was also accurate." The gap between the two is the failure mode: right person, wrong detail. The cohort split was 131 O​penAI researchers, 166 Anthropic, 151 G​oogle. Lab affiliation was established from 298 current profiles or lab pages, 148 lab bylines since October 2025, and 2 people who moved between labs. Excluded: 2 who left these labs, 13 who could not be confirmed, and 22 prominent researchers who work outside the three labs. All three models rejected all 50 invented names they were given — none of the three invented a person, which is the harder control and the one they passed.

What a "former employer" error actually costs you

If you are using a frontier model to prepare a briefing, draft an outreach list, or summarize who works where, this is the precise shape of the error you will get. The name will be right. The role will be plausible. The organization will be the one the person left. Nothing about the output will look wrong, which is why it survives review, and why a wrong-employer answer is worse than an obviously wrong one.

It also explains where the differences between models come from. Attribution errors of this kind track how a model's knowledge was snapshotted and how it phrases an answer when its training data contains both an old and a new affiliation. A model that has seen both facts and picks the more recent one scores well here. A model that has seen both and picks the more common one in its corpus — which for a long-tenured researcher is usually the older one — scores badly without being wrong about anything it was actually told. The evaluation's own framing matches: it is measuring whether a model treats "who is this person" as a question about the present.

What three different models means for this result, and what it does not

This is one posted evaluation by one account, and the honest reading has to start there. The sample is 448 names, the grading is the poster's own against public sources, the per-model bar figures are not fully reproducible from the chart's images, and no other lab or evaluator has run the same protocol. Treat it as a signal that a specific failure mode exists at a measurable rate in one of the three, not as a ranking to shop from.

What makes it more than noise is that the control is unusually good for a social-media evaluation. Giving each model only a name, once, with no web access, removes the retrieval path that would normally paper over a stale affiliation — a model with browsing would simply look the person up. Excluding researchers who left the labs, and separately excluding prominent people outside the labs, removes the cases where a "wrong current employer" answer is arguably correct. And the 50 invented names are the test that separates "stale knowledge" from "willing to make things up," which the three models passed together.

There is a tension worth naming: the reported nonsense rate — the proportion of answers that were wrong at the fringe — is essentially nil, while the recognizable inaccuracy is concentrated almost entirely in the affiliation field. A model that scored that cleanly on invented names and still misattributed employers is not confused. It is stale in exactly one dimension.

The three models are not comparable products, so read the result accordingly

• Gemini 4 Argon — announced 30 September 2026, generally available on the paid tier from 13 October at a stated $2.00 per million input and $10.00 per million output; the same evaluation that flags its attribution errors is the one that ranked it highly on recognition.

• Claude Opus 5.5 — shipped 22 September 2026 at $4.00 per million input and $20.00 per million output, down from the $5/$25 that Claude Opus 5 carried.

• GPT-6.1 Sol — shipped 29 September 2026 at $2.00 per million input and $10.00 per million output.

The age gap matters for how you read the result. Argon is the newest of the three by a wide margin, and if the failure mode is stale snapshots, the model with the most recent knowledge is the one most likely to be carrying two conflicting affiliations for the same person without a strong signal about which is current. That is a hypothesis the evaluation does not test and could not, given its design, but it is the first thing to check if you rerun it.

If you are going to use a model for people facts, test it before you trust it

This evaluation is easy to approximate on your own data, and the approximation is more useful than the aggregate. Take twenty names from your own domain — people you know have changed employers in the last eighteen months — and ask the model for each person's current role, once, with no tools. Count how many come back with the old organization. That single number tells you more about your deployment than the whole 448-name run does, because the failure is domain-specific: a model that is reliable on AI-lab researchers may be weak on your industry, or the reverse.

Two protocol details from the original are worth copying, because they are what make the result meaningful. First, no web access. A model with a search tool will simply look the person up, which is the right behavior and means the test stops measuring anything about its stored knowledge. Second, give the name only. Adding context ("this person works on robotics") lets the model infer the answer from the context rather than retrieve it, which flatters the result without fixing the weakness.

If you are running this at volume across several models rather than one, the comparison is the expensive part — the same names have to go to each model the same way, and that is where a routing layer earns its place rather than being a convenience.

Where this kind of check belongs in a pipeline

People facts are one of the few categories where a wrong answer is cheap to detect after the fact and expensive to detect before it, so the practical move is to treat the affiliation field as unverified by default and check it against a source you control. That does not mean trusting the models less across the board. It means noticing that "recognized the right person" and "described them correctly" are two different measurements, and that the second is the one you actually wanted.

On OrcaRouter this is the sort of check that runs across several models on the same key rather than one at a time. GPT-6.1 Sol and Claude Opus 5.5 — the two models this evaluation found least prone to the stale-affiliation error — are both routable here through the same endpoint, so the same twenty names go to both with a model-string change rather than a second contract, and a provider having a bad afternoon does not take the comparison down with it. Gemini 4 Argon is not on OrcaRouter; it is reachable through the vendor's own paid tier and other third-party platforms, and nothing here should be read as an offer to route it. The models we do carry are at provider list price with no mark-up, so a vendor price change reaches your bill the same day it is published.

What to watch, and what would change this reading

The result is one account's evaluation with a good control and a small sample, posted on 14 October 2026, and the honest next step is not to act on it but to look for the second run. Two things would move it from a signal to a finding: an independent evaluator repeating the protocol with a larger name list and published grading, or a vendor acknowledging that attribution between a person's former and current employer is a distinct failure class worth targeting. The 50-invented-names control passing across all three models is the detail that makes that plausible — it means the problem is not fabrication, which is hard, but refresh, which is tractable.

Until then the useful summary is narrower than a leaderboard and more actionable than one. A model that confidently reports a stale employer is not lying to you. It is answering a question about the past as though you had asked about the present, and the only way to catch it is to ask the question you actually meant.

A result like this is also a reminder that model mistakes are already documented: the AI incidents and agent failures behind models in this class are collected in the OrcaRouter AI Incident Archive , where each entry carries its own sources, severity, confidence grade and disputed flag.