Gemini 4 Argon at high scores 53 on the Artificial Analysis Intelligence Index. That ties GPT-6 Astra at max, and Fable 5.1 at max. Opus 5.5 still leads the chart at 58. GPT-6.1 Sol at max is 52.
One index task costs $1.99 at the introductory price. That is 60% of Astra's $3.26, and 2.7 times Sol's $0.72. Google's list price starts December 1, 2026. On that price, Artificial Analysis puts the same task at $3.98, about 1.2 times Astra. The tie is real. The cheaper task lasts through November 30.
The 53 is high, the top thinking level. On gemini-4-argon the default is medium. This write-up does not have an index for low or medium. The figure is also the September 30 run. Google opened the public API on October 13. Artificial Analysis has not published a second index for that build. The September chart still marks the bar as not publicly available.
Google announced Argon on September 30 for a limited set of cyber defenders, then opened it on October 13 to the paid Gemini API, to enterprises, and to Ultra and Pro subscribers in the Gemini app and AI Mode. The API page says it is off the free tier and off the Google AI Pro plan. The id is gemini-4-argon. A second id, gemini-4-argon-cyber-permissive, is the Fairwind model, and its default thinking level is high. The $2 and $10 price in Google's footnote applies to the public model.
Artificial Analysis, September run. Argon at high is 53, level with Astra and Fable. Opus is 58. The chart still marks Argon as not publicly available. That mark is from before the October 13 release.
The tie is a lower hallucination rate, plus agents
The index is 23 points above Gemini 3.1 Pro Preview, which sits at 30, and 12 points above Gemini 3.8 Flash at high.
On AA-Omniscience, Argon scores 42, next to Astra at 43 and Sol at 42. The hallucination rate is 15%, the lowest Artificial Analysis records for any model at 45 or above on the index. Astra at max is 51%. Sol at max is 54%. Accuracy is 50%, 13 points below Astra's 63%, and 5 points below Gemini 3.1 Pro Preview. The model admits gaps more often. It also gets fewer answers right than Astra.
On agents, AutomationBench-AA puts Argon first at 78%, 7 points ahead of Sonnet 5.5 at max, 71%. Terminal-Bench 4.0 is 57%, up from 4% for Gemini 3.1 Pro Preview on the same chart, and behind Sonnet at 64%, Opus at 60%, and Astra at 59%. AA-Briefcase is 1494 Elo. The rubric pass rate is 65%, the highest Artificial Analysis has recorded on that board. Analytical Quality is 1576 Elo. Presentation Quality is 1308 Elo.
Google's September post also cites AutomationBench, at 51.3%. That is Google's figure. It is not the 78%.
Artificial Analysis. Argon leads AutomationBench-AA. On Terminal-Bench 4.0 it sits behind Sonnet, Opus, and Astra, and far above Gemini 3.1 Pro Preview.
The discount is doing the work the tokens are not
Argon uses about 62,000 output tokens per index task. Astra at max uses about 27,000. The lower bill is the token price, and then the promotion. It is not a shorter answer.
The cost chart starts at Gemini 3.1 Pro Preview, $0.67 a task. At the standard Argon price the task is $3.98. The 50% promotion cuts that to $1.99. Cached input is 95% off: $0.10 per million tokens through November 30, and $0.20 from December 1. Output, including thinking, is $10 now and $20 after.
Artificial Analysis. The standard-price task is $3.98. The introductory price brings it to $1.99.
What Google measured on its own
Google ranks Argon first on the Vals Index and does not print the score in the launch posts. On DeepSWE v1.1 it reports 77.9%. On CWE-bench v1 it reports a tie for first at 68%. On LVBench, long video, it reports 91.7%. Those numbers stay in Google's column.
The October 13 post is mostly internal work: Chrome, a chemistry package, Gemma inference, fusion, and a robotics demo. Treat that as the lab's account.
The window is one budget, not two
Input, thinking, and the answer share 1,048,576 tokens. A single request stops at 65,536 output tokens, thinking included, or 32,768 on a non-streaming HTTP call. Long decode continues past that, up to the shared ceiling minus the input. The Gemini API does this with background execution. The Agent Platform does it with a continuation token, on v1beta1.
Explicit context caching is not supported. Implicit caching discounts a repeated prefix by 95%. The Gemini API docs set that prefix at 4,096 tokens. The Agent Platform docs set it at 6,144. thinking_budget and a thinking level of MINIMAL return a 400 on the Agent Platform. temperature, top_p, and top_k are ignored.
In practice
- Read the 53 as high, on the September run. Medium is what gemini-4-argon sends if you leave the control alone. Opus at 58 is still ahead of this tie.
- Use the $1.99 task through November 30, 2026. On December 1 the list price is $4 and $20, and Artificial Analysis prices the task at $3.98, about 1.2 times Astra. Cached input moves from $0.10 to $0.20. The public footnote is for the public model.
- For a callable model at this index, Astra's max is the same 53 at $3.26 a task, with about 27,000 output tokens. Argon uses about 62,000. Sol's max is 52 at $0.72.
- On agents, take Artificial Analysis's AutomationBench at 78% and Terminal-Bench at 57%. Google's AutomationBench figure is 51.3%. DeepSWE at 77.9% and CWE-bench at 68% are Google's.
- On knowledge, the gain is the 15% hallucination rate, against 51% for Astra. Accuracy is 50%, against 63% for Astra. The Omniscience score, 42, sits next to Astra's 43.
- Budget one shared window of 1,048,576 tokens. A long answer needs long decode. Set max_output_tokens high enough that thinking does not end the response. Put the static prefix first. The cache floor is 4,096 tokens in the Gemini API docs and 6,144 in the Agent Platform docs.
Sources: Gemini 4 Argon, September 30, the October 13 release, the Gemini API guide, the Agent Platform guide, and Artificial Analysis. The index of 53, the $1.99 and $3.98 tasks, the 62,000 output tokens, Omniscience, AutomationBench-AA, Terminal-Bench, and Briefcase are Artificial Analysis, at high, on the September run. The $2 / $10 price through November 30, the $4 / $20 price from December 1, the default medium, and the shared 1,048,576-token window are Google's. DeepSWE, CWE-bench, LVBench, and the Vals ranking are Google's.
