• Artificial Analysis lists Gemini 4 Argon (High) at 53 on its Intelligence Index, ranked #8 of 223 models. The Decoder reports GPT-6 Astra at the same 53 and Claude Opus 5.5 ahead at 58.
• The Decoder reports Argon averages 62,000 output tokens per task against 27,000 for GPT-6 Astra, more than twice as many.
• At introductory pricing, Artificial Analysis puts Argon at $1.99 per Intelligence Index task. The Decoder reports $3.26 for GPT-6 Astra, so Argon is cheaper per task for now.
• Standard pricing is double the introductory rate ($4 input and $20 output per million tokens), so the per-task advantage should be recomputed once promotional pricing ends.
• Google's headline claims (DeepSWE 77.9%, CWE-bench tie for first at 68%, Gray Swan prompt-injection leadership) are vendor-reported and were not independently checked here.

Google's Gemini 4 Argon reaches the top tier of frontier models without taking first place. Artificial Analysis scores it 53 on its Intelligence Index, and The Decoder reports that GPT-6 Astra matches that score while Claude Opus 5.5 leads at 58. It is Google's first frontier model in over seven months. The model uses more than twice as many output tokens per task as Astra, but at introductory prices it still costs less per task. Access is limited to select testers for now.

Where Argon lands on the independent index

Artificial Analysis lists Argon (High) at 53 on its Intelligence Index, ranked eighth of 223 models. The Decoder reports GPT-6 Astra at the same 53 and Claude Opus 5.5 at 58, a five-point lead. The Decoder also reports Argon at 77.5 percent on AutomationBench-AA and first place on the Arena.ai text leaderboard with 1,525 points. These figures come from the Artificial Analysis page and The Decoder's report. We did not reproduce the benchmarks.

Google's own results are narrower and vendor-chosen

Google's announcement cites 77.9 percent on DeepSWE v1.1, a tie for first at 68 percent on CWE-bench v1 for vulnerability remediation, and leadership on the Vals Index. It also claims leadership on the Gray Swan prompt-injection benchmark but gives no score. These are vendor-selected results, not independent measurements. They do not contradict the index tie, because a vendor picks the benchmarks it publishes and an aggregate weights many tasks. For security teams, the Gray Swan claim needs a published score before it can inform a decision.

Token use is high, but promotional pricing still wins per task

The Decoder reports that Argon averages 62,000 output tokens per task while GPT-6 Astra needs 27,000. Even so, Artificial Analysis lists $1.99 per Intelligence Index task for Argon at introductory rates, and The Decoder reports $3.26 for Astra. Argon is therefore cheaper per task today. That depends on the promotion: standard pricing is $4 per million input tokens and $20 per million output tokens, double the $2 and $10 introductory rates, so the comparison should be redone when pricing changes.

Access is restricted and the cyber positioning is unproven

Google says Argon is rolling out to trusted cyber defenders through its Fairwind Program, with developers, enterprises and consumers to follow, starting with paid API customers and Google AI Ultra subscribers. The post gives no general-availability date. The cyber-defense framing rests on Google's own CWE-bench and Gray Swan claims. No independent red-team evaluation of Argon appears in the origins reviewed. Buyers in security-sensitive deployments should treat the positioning as a vendor claim until reproducible results are published.

Verdict: Signal. Argon puts Google back in the frontier group, which matters for multi-model routing and vendor negotiation. It does not beat Claude Opus 5.5 on the independent index, and its per-task cost edge over Astra is tied to introductory pricing. Once access opens, run your own evaluation with token counts logged and compare cost per completed task at standard rates, not at the promotional rate.