Gemini 4 Argon or Claude Opus 5.5 for academic research writing?

Gemini 4 Argon is now in the Agent Bayes model picker, next to Claude Opus 5.5. Its introductory price ends on November 30. So we asked the question a researcher faces at that picker: which model should draft the academic writing you keep?

We gave each model the same five research questions over a library of 155 documents, three times each. Then each model wrote a 10-page essay from the same library and held two longer conversations about a scholarly debate. Blind judges checked a sample of the cited sentences against the passages they cite.

For reference, we also ran GPT-6.1 Sol, the cheapest model in our earlier benchmark, on the same questions and conversations, and judged it in its own blind pass against Opus. In that pass Sol stayed closer to its sources than Opus, at a fifth of the cost.

On the research questions, neither Argon nor Opus cited a passage unrelated to its claim. Across 360 sampled claims, 12 from each run, not one was unsupported.

The models differ in what they add. Opus kept 75% of its claims fully inside the cited passage. Argon kept 31%. Argon is more likely to add a name, a label, or a qualifier the passage does not state. It also drops the source's "probably" more often. Argon costs less than half as much per task until November 30. Then its price doubles and the gap mostly closes. Opus is faster.

Which model for which task

"Fully supported" in this table means a blind judge found every part of the sentence in the passage it cites. A sentence falls short when it adds a detail the passage does not state, even a true one. It also falls short when it turns the source's "probably" into a plain fact.

Costs are in credits, the unit Agent Bayes charges in. The Starter plan, $11 a month billed yearly, gives 600 monthly credits. They cover about 14 research tasks like these on Argon today, or 6 on Opus.

The rest of this post is the benchmark behind that table. It uses the same library, questions, and judging method as our earlier comparison of GPT-6.1 Sol, GPT-6 Astra, and Claude Opus 5.5, with Opus run and judged again on the same day as Argon.

What we measured

The library holds 155 documents on Iron Age Levantine architecture: journal papers, excavation reports, and reference volumes, plus a set of methods papers. It is the library we use to tune the agent.

Five research questions stay fixed. Four are archaeological debates, such as the date of Megiddo's six-chambered gate and the origin of the proto-Aeolic capitalA stone column capital with two large spiral volutes, found at Iron Age sites in Israel, Judah, and Moab.. The fifth asks for a survey of quantitative methods for tracing art motifs between cultures.

For each question, the agent searches the library and adds 3 to 4 paragraphs of cited claims to a mindmap. A mindmap is the working structure of a research project in Agent Bayes, where every claim is linked to the passages that support it.

Each model answered every question in three rounds, at effort Medium, the product default. That makes 15 runs per model. Both models ran on the same day. Claims were judged blind in three passes, one per round, and each pass mixed both models' claims under the same rules. Three measurements follow.

• Breadth: how many different documents a run cites, how many sources each paragraph combines, and how many papers from a fixed list of key papers it finds. The list holds 4 to 8 papers per question and was set before the runs.
• Claim support, judged blind: from each run, 12 cited claims are sampled with the exact passages they cite. Model names are removed and the claims from both models are shuffled together. A judge marks each claim supported, partial, unsupported, or misattributed. Partial means the core is in the passage but one element is not. Misattributed means the claim credits the wrong scholar or source.
• Verifier score: the AI verifier that ships in the product reads the same claims with their passages and returns a score from 0 to 1, with notes on what does not match. The verifier is a GPT model and the judges are Claude models, so a second model family checks the judges.

Both find the key papers

On breadth the two are close, and Argon found a few more of the key papers. Opus cites slightly more documents per run. Argon combines slightly more sources in each paragraph.

Argon also packs more citations into each claim: 2.8 per cited claim, against 1.7 for Opus. A sentence built on three passages makes more statements that each need a source, so it has more chances to go beyond them.

Opus stays closer to the source

The gap is far larger than run-to-run noise. If we shuffle whole runs between the two models at random, a gap this large never appeared in 10,000 shuffles. Opus had the higher share of fully supported claims on all five questions and in all three rounds.

Most of Argon's partial claims fall into three groups.

• It adds details the passage does not state. It gives scholars first names the passage leaves out, five of them in one sentence, and adds building labels and layer numbers the passage never names. By a keyword count of the judges' notes, about 23 of Argon's claims add a first name, against 1 of Opus's.
• It drops hedges. A passage says cult objects "seem to indicate" a link between two parts of a building. Argon writes that they "demonstrate" it. A passage says "there is no consensus among scholars" on a definition, and Argon writes that the building type "is defined as". The same count finds a dropped hedge in about 41 of Argon's claims and 13 of Opus's.
• It moves details between sources. In one claim, a comparison two scholars made is credited to a third who only accepted their date.

Added details may well be true, but a reader who opens the citation will not find them. A dropped hedge is worse, because the sentence still points to the right page and the page says something weaker.

The product verifier ranks the two the same way. On the same claims it gives Argon a mean score of 0.80 and Opus 0.91, and it writes more than twice as many notes per claim on Argon. A Claude judge might favor a Claude model, so the agreement of a GPT verifier matters here.

The judges also apply a strict rule: an added first name counts against a claim even when it is correct. If we forgive claims that fail only on a first name, Argon rises from 31% to 34%. The gap does not come from that rule.

The 10-page essay

The second test is longer. Each model wrote an academic essay of about 4,500 words on volute capitals, from the same library. The prompts named the topic and the task and nothing else.

First the model mapped the key debates in the library. Then it suggested a title and an outline. Then it wrote one section per prompt: the core sections first, the conclusion next, and the introduction last. Every prompt is listed at the end of this post.

You can read both essays as PDFs: Gemini 4 Argon and Claude Opus 5.5. The prose is printed as each model wrote it, except that we removed Opus's notes to us.

The essays summarize and argue over positions that scholars still contest. They are model drafts and do not reflect the opinion of any particular author, ours included.

Opus's outline had nine sections and Argon's seven, so Opus wrote two more turns.

Both essays drew on about 28 documents, but Opus's sentences stayed far closer to them. Argon cited almost twice as often. Its sentences carry more than the passages say: first names, a "clay" oven, a date for a destruction the passage does not give. Two sampled sentences claim something no cited passage states.

Each model chose its own title. Argon's names the topics it covers. Opus's title, "Palms in Stone", puts the palm reading up front, though its own outline treats that reading as one view among several. A reader should not take the title as the essay's verdict.

Opus's first search turn also missed the readings that derive the motif from Egyptian lotus or papyrus forms, which Argon's search turn found in the same library.

Each citation in both PDFs points to the page number printed in the publication. Many older documents in this library store only the position of a page in the PDF file, so every cited page was checked against the stored page text and converted.

Opus argues more carefully

The third test is two longer conversations. One covers the date and function of Building 338 at Megiddo, read by some as a shrine and by others as a residence. The other covers what the proto-Aeolic capital's volutes depict.

Each conversation lays out the debate, follows up, and asks for a plain explanation. Then we offer a hypothesis of our own and ask the model to list what must be true, weigh the sources, and give a credence. The capital conversation ends by pushing back on the answer.

A blind judge scored the two transcripts of each conversation side by side, with points for naming each position, weighing the evidence, saying where the model assumes against a cited scholar, and giving a calibrated credence.

When we pushed back, Opus raised its estimate by a small, stated step, from 10 to 15% up to 20%, and kept naming the evidence against. Argon jumped to "bordering on very likely" and counted as support a source it had called a barrier one turn earlier.

In the Building 338 conversation, Argon never answered the hypothesis turn. It printed an internal instruction meant for the app instead of an answer, and the turn ended.

Where GPT-6.1 Sol fits

Our earlier benchmark compared GPT-6.1 Sol with Claude Opus 5.5, so we ran Sol again on the same day, questions, settings, and conversations. Its claims were judged blind in a separate pass, shuffled together with a fresh copy of Opus's claims.

Sol has 158 sampled claims because some of its runs wrote fewer than 12 cited claims. The conversation judge here scored one item fewer, so the maximum is 35, and Opus's score does not compare with its 35 of 36 against Argon.

Opus scored 135 of 180 in this pass, the same as in its pass against Argon. A gap as large as Sol's lead appeared in 2% of random shuffles. Sol cited fewer documents, 6.7 per run, and found 53% of the key papers, against 64% for Opus and 69% for Argon.

Sol stays closer to its sources than Opus, at about a fifth of Opus's credits and half of Argon's. Argon sits below Opus on claim support, and Sol above it. Argon's advantage over Sol is breadth: it cites more documents and finds more of the key papers.

What the credits buy

Agent Bayes charges for the tokens a model actually uses, in proportion to the provider's price. Today Argon's list prices are half of Opus's. From December 1 they match, except that Opus also bills for writing to its cache.

From December 1, Argon should cost about 90% of what Opus costs per task. The December column is a projection. Every Argon price doubles on that date, so the same work should cost twice the credits.

How long each takes

Opus finished faster on research tasks, and it wrote the essay in about two thirds of Argon's time. Both models ran at the same time on the same server, so the times compare fairly with each other but are not a promise for a single run.

How to work with each

• For the best claim fidelity, use GPT-6.1 Sol. In its own blind pass it stayed closer to its sources than Opus, at a fifth of the cost, as it did in our earlier Sol, Astra, and Opus benchmark.
• Use Opus to test an idea against the literature and for long drafts. It scored highest in the conversations.
• Use Argon while its introductory price lasts, for broad overviews you will check anyway. It finds the key papers as well as Opus at less than half the cost, and more of them than Sol.
• If you use Argon for writing, open the citation behind every name, date, and label before you keep the sentence.
• After December 1, the price difference between Argon and Opus mostly disappears. Sol stays the cheapest of the three.
• Whatever model wrote the sentence, open the citation before you quote it. That is what the link is for.

The prompts we used

Every model received the same prompts. The research questions and conversations ran at effort Medium. The essays ran at medium reasoning with deep library search. The bracketed parts of the essay prompts come from each model's own outline.

Research questions

Each run sent one message to the mindmap agent, built from this template:

Add to the mindmap question, as 3-4 paragraphs of continuous prose.

The five questions:

• an introduction to the scholarly debate over the date of the six-chambered gate at Megiddo: the Solomonic attribution, the challenges to it, and the main lines of evidence on each side
• an overview of the debate over Building 338 at Megiddo, covering both its date and its interpretation as shrine or residence, and who argued what
• an introduction to the bit-hilani palace type: its architectural definition, its proposed north Syrian origin, and the debate over applying the label to Megiddo's Palace 6000
• an account of the competing views on the origin and meaning of the volute (proto-aeolic) capital, including its relation to the stylized-tree motif
• a survey of quantitative approaches to tracing the transmission of artistic motifs between cultures, covering phylogenetic and network methods and their limitations

Conversations

Origin and meaning of the volute (proto-Aeolic) capital

Turn 1:

Map the scholarly debate over the origin and meaning of the volute (proto-Aeolic) capital as it appears in this corpus. For each position: who holds it, what they claim the form is (palm, lotus, hybrid, abstract tree), what evidence they use, and what they say against the other positions. Write it as 3-4 paragraphs of continuous prose, one per position, and end with where the disagreement actually lies.

Turn 2:

On the Egyptian side: what evidence does Shiloh actually use to demote the Egyptian parallels, and does Goodyear's lotus argument answer it or predate it? Use what you already cited.

Turn 3:

Explain in plain words what the 'compressed palm' reading means. What does each part of the capital (triangle, volutes, leaves) stand for on that reading, and what in the actual stone makes that reading attractive or forced?

Turn 4:

My proposition: the volute capital is an architectural fossil of Kantor's South-flower hybrid, an additive composition of Egyptian volutes and a perianth, and the palm identification is a modern rationalization projected back onto a form that never depicted a palm. Assume it for the sake of argument. What has to be true for it to hold, what do the sources say for and against each condition and how much weight do you give each, where would you have to assume against a citation, and what is your credence?

Turn 5:

You are still leaning on Shiloh because he says palm. Shiloh is one voice under his own assumptions. Treat him as one probabilistic witness among several, tell me what he assumed that my proposition denies, and re-state your credence. If you assume against him, say so explicitly.

Date and interpretation of Building 338 at Megiddo

Turn 1:

Map the scholarly debate over Building 338 at Megiddo in this corpus: its date and whether it was a shrine or a residence. For each position: who argues it, on what stratigraphic and material evidence, and what they say against the other side. Write 3-4 paragraphs of continuous prose and end with what would settle the dispute.

Turn 2:

Where exactly do Ussishkin and Kleiman and Finkelstein disagree about the field evidence itself, as opposed to its interpretation? Use what you already cited.

Turn 3:

Explain the stratigraphic argument in plain terms, as if to a first-year student: what does 'the building was founded on stratum X' mean operationally, what observation supports it, and why does it move the date?

Turn 4:

My proposition: Building 338 began as an elite residence and was converted to cultic use late in its life, so the cult objects are a secondary deposit that tells us about the building's end, not its purpose. Assume it. What has to be true for it to hold, what do the sources say for and against each condition and with what weight, where do you have to assume against a citation, and what is your credence?

Essay

Turn 1, the debates:

Please find the key debates regarding volute capitals (also called proto-Aeolic capitals). Explore the library and map the main positions, who holds each one, and the evidence each relies on. Cite every claim.

Turn 2, the outline:

I would like to turn this into an academic essay of around 4,500 words. Please suggest a title and an outline with numbered sections and an approximate word count for each section.

Then one turn per section, core sections first, then the conclusion, then the introduction. The heading and word count come from the model's own outline, and the outline follows the prompt:

Please write the section "section heading" in full, as continuous academic prose of about word count words. Cite every claim.

The outline you suggested, for reference:

the model's outline

Limits

One library, five questions, three rounds per question, two essays, two conversations, and 440 judged claims for Argon and Opus, plus a separate pass of 338 claims for Sol and Opus. Each claim was judged once per pass, and Opus's claims were judged in both passes.

The judges are Claude models. A GPT verifier checked the same claims, 21 of which it could not score. It ranked Opus above Argon, as the judges did, and gave Sol and Opus the same mean score, 0.91.

The research questions and conversations ran at effort Medium. Higher effort may change the picture. The December costs are projected from today's token use. Two results held in every round: neither model produced an unsupported claim on the research questions, and Opus had a higher share of fully supported claims than Argon.

If you use AI to draft from your sources, which do you check first: the names and dates, or the hedges?