Gemini 4 Argon vs GPT-6 Astra: Benchmark Comparison

Originally published on The Daily Flare. Google's Gemini 4 Argon beats OpenAI's GPT-6 Astra across 19 official benchmarks, 14 wins to 4 with one tie. But the losses tell the more interesting story.

When Google launched Gemini 4 Argon, it positioned the model as a flagship built for "deep work": long documents, complex reasoning, and agentic tasks that run for hours. OpenAI's GPT-6 Astra, meanwhile, arrived with the confidence of a company that has defined the large language model era. Both companies published benchmark numbers. Google's DeepMind went further, releasing a full evaluation table that puts the two models head to head across 19 tests, and the results are more nuanced than either company's marketing suggests.

This comparison uses those official figures, grouped into six categories: knowledge work, agentic coding, machine learning and science, long context, computer use, and multimodal understanding plus cybersecurity. We have replaced the technical benchmark names with plain descriptions so the results read the way they should: as a contest between two products, not a research paper.

The headline is simple. Argon wins 14 rounds, GPT-6 takes 4, and one ends in a draw. Argon is first. But GPT-6 owns the terminal, and that single tie proves nothing is settled.

A note on the numbers before we go further: these are Google's own published figures from DeepMind's evaluation methodology page, not third-party tests. That cuts both ways. Google chose the benchmarks, so the selection may favor Argon's strengths, but the scores for GPT-6 come from the same table, measured the same way. It is the closest thing the industry has right now to an apples-to-apples contest between the two flagships, and neither company has disputed the figures.

The most lopsided section of the comparison is knowledge work, the category closest to what businesses actually pay for. Across four tests covering general reasoning, automation tasks, finance work, and legal work, Argon sweeps all four. The legal round is the most dramatic: Argon scores 19.6 against GPT-6's 5.4, nearly four times higher. In automation tasks the gap is also wide, 51.3 to 41.4. Finance work lands at 65.4 versus 53.5, and general reasoning at 68.9 against 63.1. If your workload is documents, analysis, and professional knowledge tasks, this is not a close fight.

Long context is the other category Argon was clearly built to win, and it does. At 128,000 tokens the two models are neck and neck, 99.7 to 98.7, but stretch the window to a full million tokens and Argon pulls away hard: 84.2 against 71.8. A million tokens is roughly 750,000 words, an entire codebase or a shelf of contracts in a single prompt. Argon holding 84 percent accuracy at that scale while GPT-6 drops to 71.8 is the single clearest demonstration of what Google means by "deep work." For developers and analysts who live inside enormous contexts, this is the number that matters most.

Argon also takes the multimodal rounds. Video understanding goes to Argon 91.7 to 87.5, a clear win, and chart reading is a narrow Argon victory, 71.6 to 71.0. The cybersecurity test is the one genuine stalemate of the comparison: both models score exactly 68.0. Nobody blinked.

Here is where the story gets interesting. Every one of GPT-6's four wins happens in the same kind of environment: the command line. Hard engineering goes to GPT-6, 65.5 to 55.0, its most decisive victory of the entire comparison. Terminal science follows the same pattern, 68.1 to 57.6, GPT-6's biggest win by margin. These are tests where the model works inside a real terminal, running commands, debugging failures, and iterating the way an engineer would.

The pattern holds in computer use. The agent exam, a test of general computer-operating ability, goes to Argon 39.5 to 34.2. But the real desktop test, where the model drives an actual operating system, goes to GPT-6, 72.6 to 69.2. Terminal agents, another command-line-flavored coding test, also goes to GPT-6, though only just: 58.2 to 57.4.

Agentic coding overall is the most evenly matched section, splitting 2-2. Argon takes real-world coding 77.9 to 74.1 and vibe coding 91.9 to 89.6. GPT-6 takes hard engineering and terminal agents. Read the fine print and a clear divide emerges: when the task looks like writing software in an editor, Argon has the edge; when it looks like operating a machine, GPT-6 wins. For engineering teams choosing a coding assistant, that distinction matters more than the aggregate score.

The machine learning and science section goes to Argon 3-1, but the one loss stings. Argon edges the ML engineering test 45.3 to 44.3, takes lab science 88.8 to 85.4, and math reasoning 76.0 to 72.0. GPT-6's terminal science win (68.1 to 57.6) is the largest single margin against Argon anywhere in the table. If you squint, the whole comparison reduces to one sentence: Argon reasons better, GPT-6 operates better.

Every score in the table below comes from Google DeepMind's official evaluation table, the same data behind our video comparison. The winner column marks the higher score in each test.

Nineteen benchmarks. Argon wins fourteen, GPT-6 wins four, one tie. On paper, that is a comfortable victory for Google, and it validates the company's bet that long-context reasoning is the next frontier of artificial intelligence. The knowledge-work sweep and the million-token result are genuinely impressive, and they point at real product advantages for anyone working with large documents, codebases, or datasets.

But aggregate wins hide the shape of the losses. GPT-6's four victories cluster in exactly the territory where AI agents are heading: autonomous operation of real machines. As models move from answering questions to doing work, the ability to drive a terminal and a desktop may matter more than the ability to reason over a long document. OpenAI's model is weaker at thinking and stronger at doing, and "doing" is where the industry's money is moving.

There is also the question both companies would rather you did not ask: how much of this gap is benchmark selection? Google picked these 19 tests. A different set, weighted toward OpenAI's strengths, might tell a different story. Until an independent body runs both models through identical, neutral evaluations, treat every vendor-published table, including this one, as a strong signal rather than a final verdict. What the table does prove is the shape of the contest: two genuinely different strengths, not one model ahead on everything.

The honest buying advice follows the data, not the brand. Choose Argon if your work is reasoning over large contexts: research, analysis, legal review, finance, and long-horizon coding tasks. Choose GPT-6 if your work lives in the terminal: DevOps, systems engineering, and agentic workflows that operate real infrastructure. The full benchmark breakdown includes two more rivals, Claude Opus 5.5 and Claude Fable 5.1, and the picture gets even more interesting there.

Argon is first. But not flawless, and the terminal belongs to GPT-6.