Gemini 4 Argon became generally available on October 13, 2026 Since then, scores from outside evaluators have started coming in The benchmarks Google published at the September 30 announcement were its own measurements The results from two independent evaluators diverged On the Vals AI composite index it ranked No. 1 On the Artificial Analysis Intelligence Index it scored 53, 5 points below the leader’s 58
Gemini 4 Argon’s own benchmark table published by Google on September 30 — Source: Google
Artificial Analysis Intelligence Index
Artificial Analysis Intelligence Index comparison by model — Source: self-rendered from the Artificial Analysis leaderboard
Artificial Analysis is an independent evaluator that rolls multiple benchmark results into a single index Argon scored 53 at the high reasoning setting Comparing each model’s best setting, Claude Opus 5.5 leads with 58 Claude Sonnet 5.5 and Muse Wonder Full tie for second with 56 Argon scores 53, the same as GPT-6 Astra and Claude Fable 5.1
Cost per task for the 53-point models
Cost per Intelligence Index task comparison — Source: self-rendered from the Artificial Analysis leaderboard
Argon cost about $1.99 (₩2,790) per index task GPT-6 Astra max, also at 53, cost ₩4,560 Claude Fable 5.1 max cost ₩10,680 For the same 53 points, it cost 61% of what GPT-6 Astra did However, Claude Opus 5.5 at the high setting scored 54 at ₩2,550 per task Argon isn’t the model with the best score for the cost
Speed
Output speed was 99.5 tokens per second, faster than the comparison median of 86.1 tokens The first token took 44.7 seconds to arrive That’s 11 times the comparison median of 3.94 seconds! It means it spends a long time reasoning before giving an answer
Vals AI composite index
Vals AI composite index comparison of the top four models — Source: self-rendered from Vals AI model pages
Vals AI evaluates models on work-style tasks such as finance, law, medicine, and coding Argon ranked No. 1 with a composite index of 68.90% Claude Sonnet 5.5 at 67.04% and Opus 5.5 at 66.97% are second and third Running the full evaluation once with Argon cost about ₩21,950, the lowest Claude Opus 5.5 cost about ₩45,000
Rankings by benchmark
| Benchmark | Score | Rank |
|---|---|---|
| Finance Agent v2 | 65.40% | 1st / 78 |
| IOI (Informatics Olympiad) | 100% | 1st / 44 |
| Code Migration | 68.17% | 2nd / 77 |
| Terminal-Bench 4.0 | 57.58% | 6th / 48 |
| Vibe Code Bench 1-100 | 21.20% | 8th / 25 |
| CUA-bench | 4.83% | 8th / 9 |
IOI at 100% is the same score as GPT-6 Astra On CUA-bench, which operates a computer by looking at the screen, it placed only 8th of 9 models Evaluating this benchmark once cost about ₩271,300, also the highest Sonnet 5.5 scored higher on Code Migration and Vibe Code Bench v1.1
Google’s own numbers vs outside measurements
Concept image contrasting self-reported benchmarks with independent evaluations — Source: concept image · self-generated with agy
The Terminal-bench 4.0 score Google published for Argon on September 30 was 57.4% The score Vals AI measured separately was 57.58%, nearly identical In Google’s own benchmarks, the leader on this item was Claude Opus 5.5 at 66.4%
🔗 Link - https://blog.naver.com/rottensweetpotato/224429639264
Real-world reactions
The Kyunghyang Shinmun reported much praise for writing and logical reasoning, with mixed reviews on coding One Hacker News user said they’d use Argon for every task except coding There were also complaints that usage limits run out quickly even on simple questions Google said AI-written code submissions by its internal development teams rose more than 35% It said cases of reverting code after changes fell by about a third
Usage in Korea
Monthly active users of the Gemini app and ChatGPT in Korea — Source: self-rendered from WiseApp·Retail
According to WiseApp·Retail, monthly active users of the Gemini app in Korea were 7.72 million in March By July it had grown by 1.75 million to 9.47 million Over the following two months it added only 400,000 ChatGPT users in Korea in September were 24.7 million, about 2.5 times Gemini In the Gemini app, Argon is available only to Google AI Pro and Ultra subscribers Google’s announcement didn’t list the countries where it’s available There are also no public results yet measuring Argon on Korean-language benchmarks (KMMLU and others)
What to watch next
Gemini 4 Argon independent evaluation summary card — Source: self-rendered from Artificial Analysis · Vals AI
- Cost-per-task rankings after December 1, 2026, when full API pricing applies
- Official results from Arena, a vote-based ranking
- Korean-language benchmarks and the scope of Argon availability in the Gemini app in Korea
- Scores at lower reasoning settings that cut first-token wait time
Sources
- [Artificial Analysis] Gemini 4 Argon model page (Intelligence Index, speed, cost)
- [Artificial Analysis] Model leaderboard (Intelligence Index and cost per task by model)
- [Vals AI] Gemini 4 Argon model page (composite index, rankings by benchmark, cost)
- [Google] Gemini 4 Argon announcement (own benchmark table, 2026-09-30)
- [Google] Gemini 4 Argon generally available (2026-10-13)
- [Google] Argon availability in the Gemini app (2026-10-13)
- [Kyunghyang Shinmun] Reactions to Gemini 4 Argon and Korean user numbers (2026-10-14)
- Exchange rate: converted at $1 ≈ ₩1,400