Gemini 4 Argon wins 14 of 18 benchmarks against Claude Opus 5.5 in Google DeepMind's scores

The long-context results show a pattern. At 128K tokens Argon scores 99.7 and Opus 5.5 scores 90.6, a gap of 9.1 points ^[5][10]. At a million tokens the gap is 17.4 ^[3]. Argon gives up 15.5 points between the two lengths and Opus gives up 23.8 ^[11]. If the two tests measure the same skill at different lengths, as the write-up presents them, the advantage grows with input size. A team that stays under 128K sees the smaller gap.

Legal Work is the biggest margin by ratio and the weakest to lean on. Argon's score is 5.2 times Opus's ^[12]. Both scores sit below 20 ^[8], and the absolute gap is 15.8 points, under the 17.4 on long context ^[12][4].

The write-up says several contests were decided by a single point, including general reasoning and cybersecurity ^[14]. Cybersecurity is, at 68.0 to 67.0. General Reasoning is 68.9 to 67.0, a gap of 1.9 ^[13][20]. Agent Exam (39.5 to 38.2) and Vibe Coding (91.9 to 90.3) are 1.3 and 1.6 points apart ^[15]. Four of Argon's 14 wins therefore come by less than two points ^[19].

The write-up calls the set "the fairest comparison available: same tests, same scoring, published by a single lab with nothing to hide" ^[16]. The lab is Google DeepMind, and Argon is Google's model ^[1][17]. A shared harness removes one variable. We'd treat the tally as Google's published claim until a second party reproduces the long-context gap. For the numbers to transfer, your inputs would have to resemble the benchmark inputs and your agent scaffolding would have to behave like Google's. The write-up does not describe the prompts, tool setup or sampling settings.

The text available breaks off after Terminal Agents, which Opus wins 66.4 to 57.4 ^[7]. The other three Opus wins are not itemized. Argon takes both Lab Science and Math Reasoning ^[9], so whichever science tests Opus won are not among those listed.