Gemini 4 Argon tops 14 of 19 Google-run benchmarks; GPT-6 and Claude keep key wins

Google's claim and what outside testers found

Google launched Gemini 4 Argon this week and says the model finishes first or tied for first on 14 of 19 published benchmarks measured against OpenAI's GPT-6 Astra, Claude Fable 5.1 and Claude Opus 5.5. The comparison, laid out in a report from The Daily Flare republished on dev.to, spans enterprise knowledge work, agentic coding, science, long-context reasoning and cybersecurity — the areas Google is aiming at business customers.

Outside checks support part of the story. Vals AI ranked Argon first among 41 models on its Vals Index, at 68.90%. Artificial Analysis scored the model 53 on its Intelligence Index, level with GPT-6 Astra and 23 points clear of Gemini 3.1 Pro Preview, Google's previous release.

Where Argon builds its lead

The biggest gaps show up in professional knowledge work. On Harvey's Legal Agent Benchmark, Argon records 19.6% against 5.4% for GPT-6 Astra — close to a fourfold difference. On the Vals Index, which mixes finance, coding, legal and tax tasks, Argon posts 68.9% while Astra manages 63.1% and Opus 5.5 reaches 67.0%.

Software engineering also leans Argon's way on some measures. It leads DeepSWE v1.1, a suite of real-world engineering tasks, at 77.9%, ahead of Opus 5.5 at 74.2% and Astra at 74.1%. Further wins include Vibe Code Bench at 91.9%, LVBench for long-video understanding at 91.7%, and GraphWalks — a long-context evaluation running to a million tokens — at 84.2%.

Where GPT-6 and Claude hold ground

The picture is more mixed elsewhere. GPT-6 Astra dominates FrontierSWE v2, a tougher software engineering evaluation, at 65.5% versus 55.0% for Argon. Claude Opus 5.5 takes Terminal-Bench 4.0, which measures terminal-driven agent work, 66.4% to 57.4%, and Astra carries Terminal-Bench Science at 68.1% against 57.6%.

Astra is also ahead on computer-use tasks, scoring 72.6% on the OSWorld-2.0 offline subset compared with Argon's 69.2%. Opus 5.5 wins PostTrainBench, a machine-learning engineering test, 49.3% to 45.3%. For teams that build shell-based automation pipelines, these are the scores that matter, as The Daily Flare notes.

The caveats on the numbers

Every result in the 19-benchmark table was produced and published by Google itself. At launch, no third party had independently reproduced the full set. That follows standard practice for a flagship release, but it means the headline win count comes from the vendor rather than neutral evaluators.

Two design choices frame the results. Argon was built from the ground up for long-horizon “deep work”, and its output ceiling now reaches one million tokens, up from 64,000 in Gemini 3 — a significant expansion in its own right.

Why it matters

Argon tops the benchmark count and independent raters agree it belongs in the top tier, yet rivals still own categories that decide real deployments. Legal, finance and long-context workloads favour the new Google model; harder software engineering and terminal-driven agents still favour OpenAI and Anthropic. Buyers should match the model to the workload rather than trust a single leaderboard — and with margins this narrow, the frontier race has a new leader but no settled winner.