Artificial Analysis 评测 Gemini 4 Argon，智能指数追平 GPT-6 Astra 且成本仅 60%

📌 One-Sentence Summary

Artificial Analysis benchmarks Google DeepMind's Gemini 4 Argon at 53 on its Intelligence Index， matching GPT-6 Astra (max.) while costing 60 percent less per task under launch discounts.

📝 Summary

Artificial Analysis evaluated Google DeepMind's new Gemini 4 Argon， the first non-Flash proprietary Gemini model in more than seven months. At high reasoning it scores 53 on the Artificial Analysis Intelligence Index， tying GPT-6 Astra (max， 53) and edging GPT-6.1 Sol (max， 52) by one point， a 23-point jump over Gemini 3.1 Pro Preview (30). With a 50 percent launch discount and a 95 percent cached input discount， each Intelligence Index task costs $1.99， about 60 percent of GPT-6 Astra (max， $3.26) but 2.7x GPT-6.1 Sol (max); standard pricing would raise this to $3.98. Gains come from lower hallucination (15 percent on AA-Omniscience， the lowest among the over 45 models) and stronger agentic performance (77.5 percent on AutomationBench-AA， 57 percent on Term Bench 4). The model offers a 1 million-token context window， text， image， video， and audio input， and a new Long Decode Continuation feature. The model is rolling out to select users and is not publicly available yet.

💡 Main Points

Gemini 4 Argon matches the top model on the Intelligence Index， restoring Google to the top tier of labs.

It scores 53 at high reasoning， tying GPT-6 Astra (max， 53) and beating GPT-6.1 Sol (max， 52) by one point， a 23-point jump over Gemini 3.1 Pro Preview (30) and 12 points over Gemini 3.8 Flash at high reason.

Launch discounts make it cost-competitive， but the advantage is temporary.

At the current 50 percent launch discount and a 95 percent cached input discount， each Intelligence Index task costs $1.99， 60 percent of GPT-6 Astra (max， $3.26) but 2.7x GPT-6.1 Sol (max). Standard pricing would raise this to $3.98， about 1.2x GPT-6 Astra. The efficiency comes from lower token prices， not fewer tokens: it outputs 62k tokens per task versus 27k for GPT-6 Astra.

Agentic performance， historically a Gemini weakness， shows clear improvement.

It ranks first on AutomationBench-AA at 77.5 percent， ahead of Claude Sonnet 5.5 (max， 71.3 percent)， and reaches 57 percent on Term Bench 4， a 53-point gain over Gemini 3.1 Pro Preview， trailing only Claude Sonnet 5.5 (64 percent)， Claude Opus 5.5 (60 percent)， and GPT-6 Astra (59 percent).

It has the lowest hallucination rate among leading models， though raw accuracy is slightly lower.

On AA-Omniscience its hallucination rate is 15 percent， versus 51 percent for GPT-6 Astra (max.) and 54 percent for GPT-6.1 Sol (max.)， meaning it more often admits uncertainty. Accuracy is 50 percent， down 5 points from Gemini 3.1 Pro Preview and 13 points below GPT-6 Astra (63 percent)， leaving its overall AA-Omniscience score of 42 in line with GPT-6 Astra (43) and GPT-6.1 Sol (42).

The model ships with a large context window and a new long-output feature.

It supports a one-million-token context window and text， image， video， and audio input with a text output. The Long Decode Continuation， a new Gemini API feature， allows long-sentence responses to be paused and resumed across calls， enabling up to 1 million output tokens without request timeouts.

💬 Key Quotes

Google's new Gemini 4 Argon scores a 53 on Artificial Analysis's Intelligence Index， matching GPT-6 Astra (max.)， while costing 60 percent less per task at launch.

With a 50 percent launch discount and a 95 percent cached input discount， each Intelligence Index task costs $1.99， about 60 percent of GPT-6 Astra (max.， $3.26)， but 2.7x GPT-6.1 Sol (max.) Standard pricing would raise this to $3.98. Gains come from lower hallucination (15 percent on AA-Omniscience， the lowest of over 45 models) and stronger agentic performance (77.5 percent on AutomationBench-AA， 57 percent on Term Bench 4).

At the current launch discounts， each skill and domain task costs $3.98， about 40 percent of GPT-6 Astra (max.， $9.79) but 5x GPT-6.1 Sol (max.). If those discounts were never applied， each task would cost $4.98， about double the cost of the top model in the Smashing Business Edition (about $2.50).

In the $0.20 Plan， service costs are flat whether you use just one skill or the full suite in a single request. For requests outside the $0.20 Plan， which covers 25 percent of queries， and in the Free Plan， costs are higher for each additional skill or domain.

📊 Article Meta

AI Screening： 85

Source： AIHOT — 精选

Author： noreply@aihot.news (X：Artificial Analysis (@ArtificialAnlys))

Category： 人工智能

Language： 英文

Read Time： 6 min

Word Count： 1362

Tags：

AI 与智能应用 , 模型评测与基准 , 模型发布 , 大语言模型 (LLM) , AI Agent