Gemini 4 Argon Benchmarks: Google Claims 12 Wins. Which Are Real

Google's launch table for Gemini 4 Argon bolds Argon as the sole leader on 12 of 18 benchmarks, ties on one and trails on five, all lab-reported as of 30 September 2026. Cross-checked against Vals and Artificial Analysis on 13 October 2026, two of the 12 wins are independently confirmed as #1, two have matching scores but not the top rank, and eight have no independent figure. This is a desk cross-check of published results only.

How many wins does Google actually claim?

Google's announcement does not state a win count. The 12 comes from tallying the benchmark table on the DeepMind product page, where the leader in each row is bolded. The table compares Argon with GPT-6 Astra, Claude Fable 5.1 and Claude Opus 5.5.

Some outlets report 13 of 19, because GraphWalks has two score rows (up to 128k, and 256k to 1M). Trending Topics uses that count, while Hora de Codar uses 12 of 18. It is the same table counted two ways.

The five losses: GPT-6 Astra leads FrontierSWE v2 (65.5% vs 55.0%), Terminal-Bench Science 0.1 (68.1% vs 57.6%) and OSWorld-2.0 (72.6% vs 69.2%, offline subset, partial score). Claude Opus 5.5 leads Terminal-bench 4.0 (66.4% vs 57.4%) and PostTrainBench (49.3% vs 45.3%).

Which of the 12 wins do independent sources confirm?

Two are confirmed at #1, two have reproduced scores but a lower rank, and eight have no independent figure. The table below separates them. Google column is lab-reported; Vals column is independent, captured 13 October 2026.

The rank gap on Vibe Code Bench and Harvey's is not a contradiction. Google compared Argon with three rivals. Vals ranks it against 79 to 112 models, so a win in a four-model table can sit well below first place.

Is the Vals Index lead large enough to matter?

It is real but narrow. Vals reports Argon at 68.90% ± 0.97 on the Vals Index, with Claude Sonnet 5.5 at 67.04%, Claude Opus 5.5 at 66.97% and Claude Fable 5.1 at 65.83%. The 1.86-point gap to the runner-up is about twice the interval Vals reports for Argon.

Two details matter. Sonnet 5.5, the actual #2 on Vals, is absent from Google's table. And Vals reports Argon's cost per Vals Index test at $15.68, against $21.34 for Sonnet 5.5, $32.14 for Opus 5.5 and $28.71 for Fable 5.1. Vals ran Argon at high reasoning effort, temperature 1, with a 262,144 max output token setting. The Vals rank pool is 47 models on the model header and 41 in the 30 September update note, so quote the rank with its date.

Why can't the DeepSWE win be checked?

Because the Argon number comes from Google's own run. According to Hora de Codar, Argon's DeepSWE v1.1 score used Google's mini-swe-agent harness, while the competitor numbers came from public leaderboards and vendor reports. That is a single secondary source, and the article also says the comparison is not confirmed on the live leaderboard.

Scale matters too. Hora de Codar puts DeepSWE at 113 tasks across 91 repositories. Argon's 3.7-point lead over Opus 5.5 (77.9% vs 74.2%) is therefore roughly four tasks if each task carries equal weight (our arithmetic, not a figure from the source). Without a confidence interval, a mixed harness and a gap of that size, treat it as a lead of unknown reliability.

Where does Argon actually lose, and does independent data agree?

The clearest loss is Terminal-Bench 4.0. Google's table lists Argon at 57.4%, Astra at 58.2%, Fable 5.1 at 57.9% and Opus 5.5 at 66.4%. Vals independently lists Argon at 57.58%, rank 6 of 47, so the Argon figure holds up.

The Opus 5.5 figure is the weaker number. Trending Topics reports that Google uses the 66.4% Anthropic published, while Artificial Analysis measured 59.6% with a different setup. On that independent number, the gap to Argon would be about two points, not nine. We covered that discrepancy in Claude Opus 5.5 on Terminal-Bench 4.0, and Terminal-Bench 4.0 Explained covers why scores across versions do not carry over.

Other independent results are mixed. Vals ranks Argon #2 on Code Migration (68.17%), #3 on CyberBench v1.1 (77.86%, 0.12 behind GPT-6 Sol) and #7 of 8 on CUA-bench (4.83%).

How does Argon rank on Artificial Analysis?

Eighth of 216. Artificial Analysis lists Argon (High) at 53 on Intelligence Index v4.3.2, against a median of 27 for the models it tracks. The index has ten components, including AutomationBench-AA and Terminal-Bench 4.0.

This is a different measurement from Google's. AutomationBench-AA is not the AutomationBench variant in Google's table, so the 51.3% figure should not be compared with it. Artificial Analysis lists $1.99 per Intelligence Index task and 110M output tokens against a median of 85M. Our cost per task analysis covers what that does to real bills, and the Argon vs GPT-6 Astra hallucination comparison covers reliability.

Can you use Argon, and what does it cost?

Not generally, as of 13 October 2026. Google says Argon is rolling out to trusted cyber defenders through its Fairwind Program, with broader release planned for paid API customers and Google AI Ultra subscribers "as soon as possible" and no date given.

Google gives no end date for the introductory period. Vals appears to list the post-introductory rate, which is our inference. Artificial Analysis shows the introductory rate, and its page also shows both "No publicly available API" and an FAQ saying one API provider offers the model. Any cost-per-task figure measured at $2 and $10 may roughly double once the higher price applies.

What's still unclear?

Four things limit how far any of this can be pushed.

• Harness parity. The sources reviewed give harness detail only for DeepSWE (Argon on Google's mini-swe-agent, rivals from public and vendor reports). For most other rows it is not stated in the pages cited here, so the table mixes configurations of unknown comparability.
• Missing cells. Fable 5.1 has no score on Agent's Last Exam or OSWorld-2.0, and Opus 5.5 has none on OSWorld-2.0. Wins in rows with fewer competitors are weaker wins.
• Real-work performance. Trending Topics reports that Bloomberg found employee skepticism: strong benchmark results, weaker real-world work, especially some coding tasks. That is one secondary report of a report, and it is not a measurement.
• CWE-bench v1. Argon and Astra tie at 68% in Google's table. According to Hora de Codar, Collinear AI's leaderboard shows Opus 5.5 at 67%. With no confidence interval given, a one-point spread does not separate the tie from the near-tie.

Should you trust the 12-win headline?

Trust the direction, not the count. Independent data supports Argon as a top-tier model: first on the Vals Index and Finance Agent v2, and 8th of 216 on the Artificial Analysis index. It does not support reading "12 wins" as 12 independently established leads. Eight of them are Google's own numbers (one of those, DeepSWE, rests on a harness mix), and the table leaves out Sonnet 5.5 and GPT-6 Sol, which Vals shows near or ahead of Argon on some tests.

For a team choosing a model, the useful rows are the ones that match your workload and have an independent number: finance agents and knowledge work look strongest, terminal-style agent tasks look weaker. Rerun both on your own tasks once Argon has general availability.