Gemini 4 Argon vs Claude Fable 5.1: A Full Comparison Across 17 Benchmarks

Google DeepMind's Gemini 4 Argon and Anthropic's Claude Fable 5.1 went head-to-head across 17 shared benchmarks. The result was not close—Argon took 15 of them, Fable held onto only two narrow coding wins, and an entire two categories saw Fable never show up at all.

This is the most lopsided matchup in the comparison series so far. For reference, the earlier Argon vs GPT-6 showdown was a genuine tug-of-war, whereas this one reads more like a one-sided declaration.

Knowledge Work: Argon's Overwhelming Dominance

Knowledge work is where enterprises are actually willing to pay for large language models, and it's also where Argon opened up its biggest lead. Four knowledge work tests, and Argon took all of them.

The distribution of the gaps is interesting. In general reasoning, Argon scored 68.9% versus Fable's 65.8%—a moderate gap. Financial work came in at 65.4% versus 58.9%, also within an acceptable range. But then came automation tasks, where the gap suddenly widened: Argon at 51.3%, Fable at just 31.4%—nearly a 20-point chasm. Legal work was even more dramatic—Argon at 19.6%, Fable at 6.7%, close to a threefold difference.

The legal test result deserves to be called out on its own. Agentic benchmarks of this kind involve tasks that are extremely difficult by nature, so even the winner's absolute score isn't pretty. But a threefold advantage is enough to influence procurement decisions. Any organization evaluating AI for contract review or legal retrieval would find that number hard to ignore.

Science and math tell the same story. Across the four combined machine learning engineering and science tests, Argon won every one. Lab science was the standout: Argon 88.8% versus Fable 68.6%, opening up a 20-point lead on scientific reasoning tasks. Math reasoning came in at 76.0% versus 65.6%, terminal science at 57.6% versus 52.6%, and ML engineering at 45.3% versus 40.2%. None were close enough to dispute.

What is an agentic benchmark?

The "agentic benchmarks" mentioned here are fundamentally different from traditional single-turn question-answering evaluations. Conventional benchmarks typically give a model one question and one reference answer, whereas agentic benchmarks simulate a complete workflow: the model must autonomously decompose tasks, invoke tools (such as a browser, code interpreter, or file system), maintain intermediate state across multiple steps, and ultimately deliver a verifiable result. This is why absolute scores on such tests tend to be low across the board—even leading models score under 20% on the legal work test. It's not that the models are incompetent; the tasks themselves simulate work that human experts need hours or even days to complete. It also explains why the "threefold gap" between Argon's 19.6% and Fable's 6.7% on legal work carries more signal at the procurement level than those few-point differences on general reasoning.

Long Context: Argon's Signature Strength

Long context is arguably Argon's signature advantage—Google built this model with that as a core direction. And Fable struggled most in this category.

At a 128K token context window, Argon scored 99.7% versus Fable's 91.4%. When the context was stretched to the full 1 million tokens, the gap widened sharply: Argon at 84.2%, Fable at just 65.0%. When input scale truly becomes massive, that's close to a 20-point separation. For users who need to digest enormous documents, codebases, or datasets in a single pass, this is a decisive category.

These gaps matter not just for their size, but for their consistency. Argon didn't win the knowledge work category on the luck of one test—it swept reasoning, automation, finance, and legal tasks, showing that its professional work capabilities rest on a more solid foundation rather than narrow, specialized optimization. Enterprises evaluating model deployments care precisely about this kind of breadth, because real workflows rarely stay within the boundaries of a single benchmark.

Context windows and the "needle in a haystack" evaluation logic

A context window refers to the maximum number of tokens a model can "see" at once in a single inference pass. 128K is roughly the length of a medium-sized book, while 1 million tokens is equivalent to thousands of pages of documents or a small-to-medium codebase. The key difficulty in long-context capability isn't "fitting it in" but "remembering it"—as input grows longer, the model's attention to early information gets diluted, producing the "lost-in-the-middle" phenomenon. These benchmarks therefore typically bury critical information in the middle or at the end of long texts, testing whether the model can accurately retrieve and use it. Argon's 99.7% at 128K means nearly all retrieval tasks succeeded, while the drop to 84.2% at 1 million tokens shows that even the strongest model still suffers decay from scale effects—it's just that Argon's decay curve is far gentler than Fable's.

Fable's Turf: Narrow Wins in Coding Tests

To be fair, Claude Fable 5.1 wasn't routed across every domain. Its two victories both came in agentic coding tests, and they were genuine wins, not the product of rounding.

Hard engineering went to Fable, 56.3% versus Argon's 55.0%. Terminal agent also went to Fable, 57.9% versus 57.4%. The gaps are around half a percentage point, but a win is a win. This shows that on certain demanding software engineering tasks, Fable's training gives it a real advantage—or at least a narrow lead, essentially a tie.

Argon took the other two coding tests. Real-world coding was a clear Argon victory, 77.9% versus 67.4%, a gap of more than 10 points. Vibe coding was closer, 91.9% versus 90.3%, but it was still Argon's. So the coding category ended up tied 2-2—the only category in the entire comparison that Fable didn't lose.

Multimodal and Safety: The Birth of the Biggest Gap

In multimodal and safety tests, Argon delivered the most dominant performance of the entire matchup. Chart reading ended at 71.6% versus 46.2%—a gap of more than 25 points, the largest single-test difference in the whole comparison. Video understanding came in at 91.7% versus 79.7%, and safety at 68.0% versus 58.0%. Fable was not competitive on any of these.

Then there's the matter of missing data. Fable 5.1 did not participate in any computer use benchmarks, whether the agentic exam or the real desktop test. Argon scored 39.5% and 69.2% on those two respectively, but without Fable's comparative data, the category can't be counted toward the win-loss tally.

One relevant piece of context: Anthropic has historically positioned Claude as a model with strong computer use capabilities. The absence of scores here leaves an open question, rather than a default failure. Anthropic's next-generation model will almost certainly fill this gap.

Why computer use matters, and why it's hard to test

"Computer use" refers to a model directly operating a graphical interface—moving the mouse, clicking buttons, filling out forms, completing workflows across applications—rather than merely generating text or calling APIs. This is a key step for agents moving from "can write code" to "can do work," because it bypasses legacy software that lacks APIs. Evaluating this kind of capability is extremely expensive: it requires building real or virtual desktop environments, defining automatically verifiable task success criteria, and different models vary greatly in how they interpret interface screenshots, making reproducibility poor. Anthropic once made computer use a headline selling point in the Claude series. Fable 5.1's absence from this category may involve evaluation framework adaptation, version stability, or simply scheduling—the article framing it as "an open question rather than a default failure" is a fairly restrained judgment.

Full Benchmark Data at a Glance

Below is the side-by-side data for all 17 shared tests. Argon wins 15, Fable 5.1 wins 2, and 2 were unchallenged.

Category	Test	Argon	Claude Fable 5.1	Winner
Knowledge Work	General Reasoning	68.9%	65.8%	Argon
Knowledge Work	Automation Tasks	51.3%	31.4%	Argon
Knowledge Work	Financial Work	65.4%	58.9%	Argon
Knowledge Work	Legal Work	19.6%	6.7%	Argon
Agentic Coding	Real-World Coding	77.9%	67.4%	Argon
Agentic Coding	Hard Engineering	55.0%	56.3%	Fable 5.1
Agentic Coding	Vibe Coding	91.9%	90.3%	Argon
Agentic Coding	Terminal Agent	57.4%	57.9%	Fable 5.1
ML Engineering + Science	ML Engineering	45.3%	40.2%	Argon
ML Engineering + Science	Terminal Science	57.6%	52.6%	Argon
ML Engineering + Science	Lab Science	88.8%	68.6%	Argon
ML Engineering + Science	Math Reasoning	76.0%	65.6%	Argon
Long Context	128K Token	99.7%	91.4%	Argon
Long Context	1M Token	84.2%	65.0%	Argon
Computer Use	Agentic Exam	39.5%	Untested	Uncontested
Computer Use	Real Desktop	69.2%	Untested	Uncontested
Multimodal + Safety	Chart Reading	71.6%	46.2%	Argon
Multimodal + Safety	Video Understanding	91.7%	79.7%	Argon
Multimodal + Safety	Safety	68.0%	58.0%	Argon

Reading Beyond the Score

Final score: Argon 15, Fable 5.1 2, with 2 unchallenged. On the numbers alone, this is the most decisive result in the comparison series.

But an honest reading is more complicated than the score. Fable's two coding wins, narrow as they were, prove it still deserves a seat at the table on software engineering tasks. The untested computer use category leaves a genuine blank in the comparison. And benchmarks, no matter how carefully designed, ultimately measure specific tasks under specific conditions—they can't capture writing quality, conversational feel, or the ecosystem around each model, all of which matter just as much to real users.

It's also worth putting Fable's position in a broader context. Losing 15-2 to the current benchmark leader doesn't mean Fable is a weak model. Fable 5.1 is still a highly capable system, and its coding wins show that on tasks developers care about, it can still trade blows with Argon. Part of why the gap looks so large is that Argon is genuinely excellent on this evaluation set, rather than Fable collapsing. Against most other models on the market, Fable would be far more competitive.

For readers following this series, the pattern is now clear: Argon leads across all domains, but by how much depends on the opponent. GPT-6 gave Argon real trouble on coding and terminal tasks. Fable could only manage two narrow coding wins. The question for the next matchup is: is there any opponent that can take an entire category, not just one or two tests?

One-line Summary

Gemini 4 Argon wins 15 of 17 shared benchmarks against Claude Fable 5.1, dominating across the board.

Paragraph Summary

Google DeepMind's Gemini 4 Argon faced off against Anthropic's Claude Fable 5.1 across 17 shared benchmarks, winning by a lopsided 15-2 margin, with 2 computer use tests in which Fable did not participate. Argon swept all four knowledge work tests, opening a nearly threefold gap in legal work (19.6% vs 6.7%) and a nearly 20-point lead in automation tasks. Long context is Argon's signature strength, leading 84.2% to 65.0% at 1 million tokens; it also dominates multimodal and safety, with chart reading at 71.6% vs 46.2% marking the largest gap of the entire comparison. Fable held only hard engineering and terminal agent, both by about half a percentage point, leaving the coding category tied 2-2. The article cautions that the story beyond the score is more complex: Fable remains a strong model, and the missing computer use category leaves a real blank.