Gemini 4 Argon Benchmarks: Google Claims Lead on 14 of 19 Tests

Google has introduced Gemini 4 Argon, which it describes as its most capable model so far. The company's own comparisons show it beating or matching the leading competing systems on most of the tests it chose to publish. The headline figure is a lead or tie on 14 of 19 benchmarks, with the biggest gains in finance, legal work and business automation.12 The more useful detail is in which benchmarks were picked and where Argon falls short.

What Google announced

Google compares Argon against two current flagship rivals, GPT-6 Astra and Claude Opus 5.5. It says Argon outperforms or equals them on 14 of the 19 benchmarks in its published set.2 The company groups its clearest advantages under professional knowledge work:2

• multi-step financial research
• legal research and drafting
• end-to-end business automation

All available accounts agree on this framing. Both the short summary and the longer write-up describe strength in finance, legal and automation tasks as Argon's defining trait.12 Because both appear to come from the same outlet, this is really one account repeated rather than independent confirmation. For now, the narrative rests heavily on Google's own presentation.

The numbers behind the claim

Two specific results stand out:2

• AutomationBench, which tests whether a model can carry out multi-step business processes: Argon scores 51.3 percent, reportedly several points ahead of the next-best systems.
• Vals Finance Agent v2: Argon reaches 65.4 percent.

These figures need context. A 51.3 percent score on an automation benchmark means the model still fails at roughly half of the multi-step business tasks it attempts. Leading the field is not the same as being reliable. For an enterprise thinking about handing over end-to-end workflows, the main takeaway is that the whole category is still maturing, with Argon at its front edge.

The finance-agent score is similar. Sixty-five percent on agentic financial research is a meaningful result. It also implies that roughly a third of tasks still fall short. Firms that depend on accuracy will likely keep humans closely involved.

Coding: a split decision

Argon's software-engineering results are more mixed, and that matters for how the launch should be read. Google reports that Argon sets a new high on DeepSWE v1.1, a suite that measures long-horizon software-engineering work.2 It also says the model trails competitors on some terminal-based and systems-level tests.2

That split suggests Argon is tuned for sustained, multi-step reasoning across large tasks: planning, tracking state and finishing extended jobs. It appears weaker at the lower-level precision needed for command-line or systems work. This fits the broader pattern in Google's results, which favor long, workflow-style tasks over narrow technical ones.

For developers, the practical conclusion is that no single model clearly dominates coding. Teams doing infrastructure or terminal-heavy work may still prefer a competitor, while those building agents for large codebases may find Argon attractive.

Why the framing matters

The most important point may be in Google's own wording: these are results on the benchmarks the company chose to highlight.2 Vendor-selected benchmark suites are standard in AI launches, but they are not neutral. A company has every reason to stress tests where its model does well. A record of 14 wins or ties out of 19 shows Argon is competitive. It does not establish that Argon is better overall, especially since five of the 19 comparisons, including some coding tests, favor rivals.2

The emphasis on tasks that "resemble real professional workflows rather than pure academic puzzles" is also a strategic choice.2 As models approach the ceiling on traditional academic tests, AI labs are moving toward benchmarks that look like paid work: financial analysis, legal drafting, business process automation. Those are the areas where enterprise budgets sit. Google's focus on finance and legal results reads as a pitch to corporate buyers as much as a technical claim.

The takeaway

Gemini 4 Argon looks like a credible frontier model with a clear specialty. On the evidence available, it is strongest in long, multi-step professional tasks and in long-horizon software engineering. It is less dominant in systems-level coding.12 The "14 of 19" figure is real but comes from Google's own scorecard, and the absolute scores, especially on automation, show that agentic AI still fails often.

The launch looks like a strong opening position rather than a clear win. The next test is whether independent evaluations and real enterprise deployments confirm Argon's lead in finance and legal work, or whether the gap narrows once the comparisons are run by someone other than Google.