Originally published on The Daily Flare.

Gemini 4 Argon Benchmarks: How Google's New Model Stacks Up Against GPT-6 and Claude

Google says Gemini 4 Argon beats rival flagship models on most published benchmarks. Independent testers confirm parts of the story, but not all of it.

According to results published by Google, Argon finishes first or tied for first on 14 of 19 benchmarks against OpenAI's GPT-6 Astra, Claude Fable 5.1, and Claude Opus 5.5. The comparison covers enterprise knowledge work, agentic coding, science, long-context reasoning, and cybersecurity — the areas where Google is pitching the model to businesses following its launch this week.

Where Gemini 4 Argon Comes Out on Top

The widest margins sit in professional knowledge work. On Harvey's Legal Agent Benchmark, Argon scores 19.6%, nearly four times GPT-6 Astra's 5.4%. On the Vals Index, which measures AI performance across finance, coding, legal, and tax tasks, it reaches 68.9% against 63.1% for Astra and 67.0% for Opus 5.5.

In coding, Argon leads DeepSWE v1.1, a set of real-world software engineering tasks, with 77.9% versus 74.2% for Opus 5.5 and 74.1% for Astra. It also tops Vibe Code Bench at 91.9%, LVBench for long-video understanding at 91.7%, and GraphWalks, a long-context test stretching to a million tokens, at 84.2%.

Independent testing backs some of this up. Vals AI ranked Argon first among 41 models on its Vals Index at 68.90%. Artificial Analysis scored it 53 on its Intelligence Index, tied with GPT-6 Astra, and 23 points above Google's previous model, Gemini 3.1 Pro Preview.

Where Rivals Still Lead

The table is not a clean sweep. On FrontierSWE v2, a harder software engineering evaluation, GPT-6 Astra leads decisively at 65.5% to Argon's 55.0%. Claude Opus 5.5 wins Terminal-Bench 4.0, which tests terminal-driven agent work, at 66.4% versus Argon's 57.4%, and Astra leads Terminal-Bench Science 68.1% to 57.6%.

Astra also edges computer-use tasks, scoring 72.6% on the OSWorld-2.0 offline subset against Argon's 69.2%, while Opus 5.5 leads PostTrainBench for machine-learning engineering at 49.3% to 45.3%. For teams building shell-based automation pipelines, those are the exact workloads that matter most.

The Fine Print Behind the Numbers

Every one of the 19 results was run and reported by Google itself. No third party had independently reproduced the full table at launch. That is standard practice for a flagship release, but worth remembering when a vendor grades its own homework. As our earlier explainer on Gemini 4 Argon noted, the model was designed from the ground up for long-horizon "deep work," and its 1-million-token output ceiling (up from 64,000 in Gemini 3) is itself a headline feature.

The honest summary: Argon leads the benchmark count, independents confirm it is genuinely top-tier, and its rivals still own several categories that matter. The large language model race, in other words, is not over. It just got a new front-runner.