Head-to-head · published results only

Gemini 4 Argon vs Claude Opus 5.5 on finance benchmarks

Same token price ($4 in, $20 out per million), two different strengths. Gemini 4 Argon leads filing-based research; Claude Opus 5.5 edges ahead on Excel modeling and tax research, by less than the published margins.

Same benchmarks, same harness, as published by Vals AI Context: FAB v2: n=450 test questions, 3 runs · Oct 12, 2026 leaderboard · Vals AI · ± SEM Context: EMB: n=51 test tasks · Oct 12, 2026 leaderboard · Vals AI · ± as published, type not stated Context: Tax Agent: n=193 test questions · Oct 12, 2026 leaderboard · Vals AI · ± as published, type not stated
Measure	Gemini 4 Argon	Claude Opus 5.5	Gap	Read
Finance Agent v2	65.40% ±0.32	58.59% ±0.17	6.81 pts	Gemini 4 Argon ahead
Excel Modeling Benchmark	75.24% ±2.46	75.94% ±2.38	0.70 pts	Within published margins
Tax Agent Bench	44.56% ±3.59	45.57% ±3.59	1.01 pts	Within published margins
FAB v2 cost / time per task	$4.38 / 16m21s	$9.22 / 33m08s
API price per 1M tokens (in / out)	$4.00 / $20.00	$4.00 / $20.00

How to read this

"Within published margins" means the gap is smaller than the two published margins added together, so the order is not settled.
Both models ran in the same Vals AI harness with the same tools; your own prompts and tools may give different results.
These are research and modeling tests graded by AI judges, not trading results.

Limits that apply to each model

Gemini 4 Argon

Slowest of the top three on Finance Agent v2: 16 minutes 21 seconds per task, against 3 minutes 21 seconds for Gemini 3.8 Flash.
Under the stricter All-Pass grading, Vals names Muse Spark 1.2 and Gemini 3.8 Flash as the leaders; an All-Pass figure for Gemini 4 Argon is not shown in the page text.
One of the three AI judges on Finance Agent v2 is Gemini 3.1 Pro, from the same model family. Vals does not report a bias test for its jury.

Claude Opus 5.5

Most expensive and slowest per Finance Agent v2 task among the models in this directory: $9.22 and 33 minutes 8 seconds.
Vals ran it with Claude Opus 5 and Opus 4.8 as server-side fallbacks for refusals; the model page reports a fallback rate of 2.15% Context: n: not disclosed · Vals Index runs, Sep 2026 · Vals AI · margin: not disclosed.
Claude Sonnet 4.6, from the same family, sits on the Finance Agent v2 jury.

Primary sources

Verified
Gemini 4 Argon model page (Vals AI)Checked October 14, 2026.
Verified
Claude Opus 5.5 model page (Vals AI)Checked October 14, 2026.
Verified
Finance Agent v2 (Vals AI)Checked October 14, 2026.
Verified
Excel Modeling Benchmark (Vals AI)Checked October 14, 2026.
Verified
Tax Agent Bench (Vals AI)Checked October 14, 2026.

Last verified: Oct 14, 2026. Prices and scores change; check the primary source before relying on them. We have no affiliate links and receive nothing from the companies listed. Every percentage carries its sample size, period and source (no naked percentages).

Not financial advice. For information and education only. Benchmark scores and vendor descriptions are not trading results and do not predict returns. Nothing here is a recommendation to buy or sell any asset or to subscribe to any product. See our disclaimer.