Gemini 4 Argon's 14-4 tally counts a 0.8-point Astra win the same as a 10.5-point one

The 14-4 count adds one per win and ignores the size of the win. Astra's four margins are 10.5, 10.5, 3.4 and 0.8 points ^[17]. Two Argon margins in the writeup are larger than any of them: 12.4 points at one million tokens ^[18] and 14.2 on the legal test, 19.6 to 5.4 ^[7][19]. A 0.8-point win and a 10.5-point win each add one to the tally.

The writeup does not report run counts, score variance or the agent harness behind the table. Without a spread, the 0.8-point lead on terminal agents is a thin reason to pick a model.

The writeup says the terminal rows put the model in an actual terminal, where it executes commands, fixes what breaks and tries again, much as an engineer does ^[12]. For Astra's 10.5-point gaps to carry over, your agent would have to loop through shell commands and their output in the same way, on tasks that look like hard engineering and terminal science, with tools and timeouts close to whatever the original harness used.

Argon's lead on real-world coding is 3.8 points, 77.9 to 74.1 ^[9][20]. A team whose agents spend most of their time in a shell should test Astra first. A team whose work looks like the real-world coding row should test Argon first. In our view the useful step is to replay a sample of the team's own tasks through both models in the team's own harness and score them with the team's own tests.

The writeup comes from The Daily Flare and was reposted on dev.to. It says the figures are Google's own, taken from DeepMind's evaluation methodology page, and are not third-party tests ^[14]. It says Google chose the benchmarks, so the selection may favor Argon's strengths, and that the GPT-6 scores come from the same table, measured the same way ^[15]. It also says neither company has disputed the figures ^[22]. The dev.to copy breaks off mid-sentence at "vibe codin" in the agentic coding section ^[16].