Gemini 4 Argon leads Zapier’s workflow test, but trails on long coding tasks
The Aggregate Digest — Thursday, October 1, 2026
Gemini 4 Argon leads Zapier’s workflow test, but trails on long coding tasks
Gemini 4 Argon completes more business workflows than GPT-6 in Zapier’s testing, making it a candidate for automating sales, marketing and support work. It also performs well on scientific Python problems but trails the leaders on long engineering projects.
Third overall, with a narrow margin. Argon enters The Aggregate at 2031 Elo. Claude Opus 5.5 scores 2028. Argon’s estimated standard error is 23 points, much larger than their three-point gap.
The Aggregate, October 1, 2026. Displayed model ranks include private models and omit human baselines. Argon’s estimated standard error is 23 Elo points.
Office workflows are the strongest result. At High reasoning effort, Argon leads Zapier’s AutomationBench with a 51.3% pass rate, against GPT-6 at Max effort on 41.4%. These tasks require connected apps and tools to produce an operational result. Argon also leads Artificial Analysis’ separate AutomationBench-AA evaluation. That evaluation uses a different completion metric, so its percentages cannot be compared directly with Zapier’s.
Zapier AutomationBench, October 1, 2026. Exact reasoning settings shown; pass rate, higher is better. Artificial Analysis’ related evaluation uses a different metric.
Scientific coding is a strength. On Artificial Analysis’ SciCode, which tests Python functions for laboratory research problems, Argon High passes 61.81% of subproblems. GPT-6 Max passes 56.48% and the earlier Gemini 3.8 Flash High passes 56.60%.
Long engineering tasks expose a gap. FrontierSWE V2 gives each task a 20-hour budget and averages five runs. Argon places fifth on 54.96%, behind GPT-6 Max on 65.51%. This evaluator does not label Argon’s reasoning effort. For these extended engineering tasks, GPT-6 still has the stronger result.
FrontierSWE V2, October 1, 2026. Mean score over five runs, with a 20-hour task budget. Argon and Sonnet have no reasoning setting specified on this board.
