Gemini 4 Argon just posted 51.3% on Zapier’s AutomationBench, 8.8 points clear of Claude Opus 5.5 and 9.9 clear of GPT-6 Astra. Google announced the model on September 30, 2026. And the Gemini 4 Argon benchmark numbers since launch tell one consistent story: it leads every public long-horizon agentic scoreboard, including the Vals Index at 68.9% against Opus 5.5’s 67.0%, plus a state-of-the-art 77.9% on DeepSWE v1.1 for long-horizon software engineering. It does not lead everything.
On FrontierSWE v2 it sits at 55.0% while the rest of the field scores between 56.3% and 65.5%.
That split matters more than any single trophy if you build agents for a living.
The scoreboard, straight from the tables
The suites doing the work here are third-party benchmarks: the Vals Index, which measures model performance across finance, coding, legal. And tax work; Zapier’s AutomationBench, which measures end-to-end execution across core business functions; Harvey’s legal agent benchmark; and DeepSWE v1.1 for real-world, long-horizon software engineering. The comparison table itself lives on Google’s model card. So the framing is the vendor’s even when the tests are not. The scores are the interesting part.
| Benchmark | Gemini 4 Argon | Claude Opus 5.5 | GPT-6 Astra |
|—|—|—|—|
| Vals Index | 68.9% | 67.0% | 63.1% |
| AutomationBench (Zapier) | 51.3% | 42.5% | 41.4% |
| Vals Finance Agent v2 | 65.4% | 58.6% | 53.5% |
| Harvey’s Legal Agent | 19.6% | 3.8% | 5.4% |
| DeepSWE v1.1 | 77.9% | 74.2% | 74.1% |
| FrontierSWE v2 | 55.0% | 62.3% | 65.5% |
Start with the Vals Index, because it is the headline everyone quotes. Argon’s 68.9% beats Opus 5.5 by 1.9 points, with Claude Fable 5.1 at 65.8% and GPT-6 Astra at 63.1% trailing. That is a win, not a rout, and I would not rebuild anything around a 1.9-point margin.
The agent suites are where the gaps actually open. Argon’s 65.4% on Vals Finance Agent v2 leads Opus 5.5 by 6.8 points and GPT-6 Astra by 11.9. On Harvey’s Legal Agent Benchmark it scored 19.6% against 5.4% for GPT-6 Astra and 3.8% for Opus 5.5, more than triple GPT-6 Astra’s result. Before you screenshot that number, read it in absolute terms: the best model on the board still fails roughly four out of every five legal agent tasks. Winning a benchmark and being ready to run your contracts unattended are different universes.
Long-horizon agentic work is where the margin lives
DeepSWE v1.1 measures performance on real-world, long-horizon software-engineering tasks. And Google’s announcement calls Argon’s 77.9% a new state of the art. Opus 5.5 scored 74.2% and GPT-6 Astra 74.1% on the same suite. So the margin at the top is 3.7 points.
Why you should care about that specific suite more than the rest: agents die in the middle.
In the pipelines I run, failures are almost never “the model wasn’t smart enough.” They are “the model forgot what it was doing at step twelve of fifteen.” Long-horizon scores are the closest public proxy for that exact failure mode, the one that costs you a night of compute and a morning of cleanup. AutomationBench is the same capability wearing office clothes: execute a whole business function end to end instead of answering one question well.
One honest flag on scope.
The launch conversation also pitched Argon’s physical-world reasoning, meaning world-modeling claims about reconstructing events and inferring hidden objects and intentions. I cannot find a public, citable score for any of that in the verified material. So I am not grading it. If those claims hold up under independent tests, they matter most for agents that act in physical space rather than in text.
Until a number exists, it is a claim, not a result.
Two numbers that should slow your migration down
The row nobody puts in the launch thread is FrontierSWE v2, where Argon scored 55.0% while the other three models in the same table posted 65.5%, 62.3%, and 56.3%. Per that table’s column order, used identically across every other benchmark, the 65.5% belongs to GPT-6 Astra and the 62.3% to Claude Opus 5.5. Argon trails the entire field there. So “Argon wins coding” is not proven; “Argon wins long-horizon work” is.
The second catch is provenance.
Vals, Zapier, and Harvey own their benchmarks, but the comparison table being shared this week is Google’s.
The vendor picked which suites to headline, and that is how every launch in this industry works. It is selection, not fraud. You still grade the model on your own tasks, as a 1.9-point Vals margin is small enough that your prompt and your agent scaffolding decide the winner on any given day.
What I would do if I ship agents
I have not put Argon through anything I run yet.
And I would not migrate a working pipeline for 3.7 points on anyone’s word.
What I would do is prepare, since this leaderboard is telling you something real about task shape.
The wins cluster on multi-step, end-to-end work, and the one public loss clusters elsewhere.
– Split your recurring agent tasks into two piles: long-horizon multi-step work versus short single-pass output. That split is now a model-selection decision, not a footnote.
– Write an eval for your top three tasks before you switch anything. If you cannot tell which model won your own eval, a leaderboard will not tell you either.
– Route per task, not per vendor. The AutomationBench and DeepSWE numbers say long-horizon work goes to Argon when you can call it; the FrontierSWE row says keep your incumbent for the work it already passes.
– Do not rewrite a pipeline that ships for a margin you have not reproduced yourself.
For small operators, the AutomationBench gap is the one that maps to work you actually bill for: reconciliations, document workflows, the multi-step back-office stuff that eats your afternoons.
An 8.8-point jump in end-to-end execution is the difference between an agent you babysit and one you let run. That is worth your attention this quarter.
Start this week. List every recurring multi-step agent task you run and mark which model currently handles it.
That list is your migration plan. And the day Argon reaches your API key you will know within 24 hours whether the leaderboards hold on your work.
