Gemini 4 Argon vs GPT-6 Astra: Same Score, but One Hallucinates 3x More

Gemini 4 Argon and GPT-6 Astra both score 53 on the Artificial Analysis Intelligence Index v4.3.2, as of 12 October 2026. The Batch reports AA-Omniscience hallucination rates of 15% for Argon (high) and 51% for Astra (max), a ratio of about 3.4x.

That gap is real only if you read it correctly. The two models run at different effort settings, the hallucination figures come from one secondary source, and the metric does not penalise refusals.

Do Argon and Astra really score the same?

On the headline index, yes: 53 each, as of 12 October 2026. Artificial Analysis lists Argon (high) at #8 of 218 models and Astra (max) at #7 of 218, so the tie is a score tie, not a rank tie. The Batch describes Argon as tying for third, which does not match those pages.

The index is a weighted average of 10 evaluations, with four categories (agents, coding, general capability, scientific reasoning) each contributing 25%. AA-Omniscience is one of the ten. A model can therefore trade a weak Omniscience result against strong agentic or coding results and land on the same composite.

All rows captured from Artificial Analysis model pages on 12 October 2026, except Argon's introductory price, which Google states rises to $4 and $20 per million tokens after the introductory period.

What are the hallucination numbers?

The Batch says Argon (high) has the "lowest hallucination rate (15 percent) of any model scoring 45 or higher on the Intelligence Index" and gives GPT-6 Astra (max) as 51 percent. Artificial Analysis's model pages list AA-Omniscience metrics but show no values for either model, so these two figures rest on The Batch alone.

One adjacent figure is on Artificial Analysis directly. Its AA-Omniscience page gives GPT-6 Astra (High) an Omniscience Index of 44, against 46 for Claude Opus 5.5 at the top. That is the High variant, not the Max variant behind the 51% figure, so the two should not be mixed.

What does AA-Omniscience measure?

It is a 6,000-question factual-knowledge test across 6 major domains, per Artificial Analysis. The arXiv paper (Jackson, Keating, Cameron and Hill-Smith, 17 November 2025) says the questions cover 42 economically relevant topics. Questions are derived from academic and industry sources by an LLM-based generation agent.

Artificial Analysis defines the hallucination rate as incorrect answers divided by all non-correct responses: incorrect plus partial plus not attempted. Accuracy is correct answers over all questions. The Omniscience Index ranges from -100 to 100, where 0 means as many correct as incorrect answers. The page gives no explicit formula.

Does a low hallucination rate mean the model knows more?

Not necessarily. The evaluation page says the index has "no penalty for refusing to answer", which means abstention is free. A model that declines when uncertain can post a low hallucination rate at modest accuracy. The paper's abstract frames this slightly differently, saying the index rewards abstention when uncertain.

Accuracy figures for Argon and Astra are not published on their Artificial Analysis pages. Without them, the 15% versus 51% comparison cannot separate two explanations: Argon knows more, or Argon abstains more. The Batch's wording, a "lowest rate among models scoring 45 or higher", is a statement about calibration among strong models, not about raw knowledge.

Is this a fair comparison of effort settings?

Only partly. Argon was measured at high reasoning and Astra at max. Argon used 110M output tokens on the index and Astra 60M, per Artificial Analysis, so Argon spent more tokens yet cost less per task ($1.99 against $3.26) because its per-token price is lower. The Batch notes that $1.99 reflects the discounted price and that the standard price works out to $3.98 per task, which would put Argon above Astra on cost once the introductory period ends.

Hallucination behaviour can change with reasoning effort, so a max-effort Argon or a high-effort Astra may produce different numbers.

What do the labs claim themselves?

Google's announcement lists self-reported results including DeepSWE v1.1 at 77.9%, AutomationBench at 51.3% (ranked #1), LVBench at 91.7% and CWE-bench v1 at 68% (tied for first). Those are Google's figures, and the post contains no hallucination rate or Artificial Analysis index number. The Batch separately reports Argon at first on Vals Index v2.1 (68.9%) and Arena's Text Arena (1,525 Elo), and says it trails Anthropic's newest models on Artificial Analysis's broad index.

For Astra, Constellation Research reports that OpenAI claimed 72.6% on an offline OSWorld 2.0 subset, according to that outlet. This is a lab-reported figure on a different benchmark and says nothing about hallucination.

Can you use either model today?

Argon is not broadly available. Google says it is rolling out to trusted cyber defenders through its Fairwind Program, with broad release undated and planned to start with paid API customers and Google AI Ultra subscribers. Artificial Analysis lists Astra as released on 3 September 2026 and available through 11 providers as of 12 October 2026. Because Argon access is gated, independent replication of the hallucination figure is limited.

What's still unclear?

• Primary-source confirmation. The 15% and 51% figures rest on The Batch alone. Artificial Analysis's pages did not show them.
• Accuracy and abstention. No accuracy or not-attempted rate is published for either model, so the 3.4x gap cannot be attributed to knowledge versus caution.
• Effort mismatch. High versus max is not a controlled comparison.
• Rank discrepancy. The Batch says tied for third; Artificial Analysis shows #8 and #7.
• Confidence intervals. No uncertainty ranges are published for the Omniscience figures or the composite index. A one-point index tie says little without them.
• Contamination. Artificial Analysis's page describes LLM-generated questions from public academic and industry sources. Labs do not disclose their training data, so any overlap with that material is unknown (inference, not a finding).

The weakest link in the headline is the framing. "Same score" is a composite tie across ten evaluations, and "3x more" is a ratio of rates on one unpenalised-refusal metric. Both are true as stated, and neither says which model to trust on your own documents.