← Back to feed

Gemini 4 Argon leads Arena.ai’s Text Arena at 1,525 ELO as of October 9, overtaking Claude Opus 4.6 high (1,505) and Claude Opus 5.5 high (1,504). The model launched September 30 under restricted access — Fairwind program participants in cybersecurity only — and has accumulated enough human-preference votes over nine days to establish a clear lead on the open leaderboard.

Alongside the Arena result, Artificial Analysis’s AA-Omniscience factual benchmark puts Argon at a 15% hallucination rate: the lowest of any model scoring 45 or higher on the Intelligence Index. GPT-6 Astra max scores 51% on the same benchmark. Argon hallucinates about a third as often as the closest frontier competitor.

Where Argon Leads

Argon also leads Vals AI’s Vals Index v2.1, scoring 68.9% accuracy across eight benchmarks — finance, coding, legal, and tax — weighted by each sector’s share of US GDP. First of 43 evaluated models. The cost: $15.68 per test and 46.55 minutes per test at standard pricing, which at $4/$20 per million input/output tokens reflects Argon’s design for long, complex professional tasks rather than short interactive queries.

On Artificial Analysis’s Intelligence Index v4.3.2, Argon scores 53 — the same as GPT-6 Astra max — at a lower per-task cost. The broad intelligence comparison is a draw. The differentiation is elsewhere.

The Hallucination Gap

AA-Omniscience penalizes wrong answers but not refusals. The scoring rewards a model that says “I don’t know” over one that confabulates. At 15%, Argon’s refusal-and-accuracy balance is calibrated toward precision. GPT-6 Astra max at 51% is calibrated toward completeness — it answers more, but a larger fraction of those answers is wrong.

For legal, finance, and security work — exactly the verticals Argon targets through the Fairwind program — a false positive from a hallucinated fact can be more expensive than a gap in coverage. The benchmark numbers match the stated design priority.

Arena ELO: What the Gap Means

Arena Text ELO accumulates from blind head-to-head comparisons by volunteer raters. A 21-point lead over Claude Opus 5.5 high (1,525 vs 1,504) at nine days post-launch is notable, though the confidence interval on a newer entrant is wider than on models with months of votes. The direction is clear; the magnitude will stabilize over the next few weeks.

Claude Opus 5.5 high and Claude Opus 4.6 high are set to the same reasoning tier — high — as the Argon entry on the leaderboard. The ELO comparison is therefore reasonably apples-to-apples.

Access and Pricing

Gemini 4 Argon remains limited to Fairwind cybersecurity program partners. Google has not disclosed a date for broader availability. Intro pricing is $2 per million input tokens and $10 per million output tokens; standard pricing is $4/$20. The 1M token input context and 1M token output limit are unusual — most frontier models cap output at 8K to 64K tokens.

The high output limit is a prerequisite for the document-drafting tasks (legal, financial reports) that define Argon’s intended use. It also explains the per-test cost figure from Vals: 46.55 minutes and $15.68 for a full professional-document evaluation is not a number designed for chat-style queries.