Originally published on The Daily Flare.
Key takeaways
- ✓ Gemini 4 Argon beats Claude Fable 5.1 on 15 of 17 shared benchmarks, with its biggest margins in legal work, automation tasks, and chart reading.
- ✓ Fable's only wins come in two narrow coding tests, Hard Engineering and Terminal Agents, keeping the rivalry alive in software development.
- ✓ Fable 5.1 was never tested on computer use benchmarks, so that category remains an open question between the two models.
Google's Gemini 4 Argon takes on Anthropic's Claude Fable 5.1 across 17 head-to-head benchmarks, and the result is the most one-sided matchup in our comparison series.
Where the Argon versus GPT-6 battle was a genuine fight, this one is closer to a statement. Argon wins 15 of the 17 tests the two models share, often by wide margins. Fable 5.1 manages just two victories, both in coding, both narrow. And two entire categories, computer use tests that Argon completed, were never attempted by Fable at all.
That last point matters. This comparison is built on Google DeepMind's official evaluation data, the same benchmark results we used for the GPT-6 matchup. Every number below comes from that published table. Where Fable has no score, we say so plainly rather than guessing.
Argon's Commanding Lead in Knowledge Work and Science
Knowledge work is where businesses actually spend money on large language models, and it is where Argon opens up its biggest lead. Across the four knowledge work tests, Argon wins every single one.
The margins tell the story. In General Reasoning, Argon scores 68.9% against Fable's 65.8%, a solid but respectable gap. In Finance Work, it is 65.4% to 58.9%. Then the gap explodes. Automation Tasks comes in at 51.3% for Argon versus just 31.4% for Fable, a nearly 20-point chasm. Legal Work is even more dramatic: Argon at 19.6% against Fable's 6.7%, almost triple the score.
The legal result deserves attention. These agent benchmarks are scored on genuinely hard tasks, which is why even the winning numbers look low in absolute terms. But a 3x advantage on legal work is the kind of gap that decides purchasing decisions. Any firm evaluating artificial intelligence for contract review or legal research will find this number hard to ignore.
Science and math deliver more of the same. Argon takes all four tests in the combined machine learning engineering and science category. Lab Science is the standout: 88.8% for Argon against 68.6% for Fable, a 20-point margin on scientific reasoning tasks. Math Reasoning lands at 76.0% versus 65.6%. Terminal Science goes 57.6% to 52.6%, and ML Engineering 45.3% to 40.2%. None of these are close enough to dispute.
Long context is arguably Argon's signature strength, the capability Google built the model around, and Fable struggles most here. At 128K tokens, Argon scores 99.7% to Fable's 91.4%. Stretch the context to a full million tokens and the gap widens sharply: 84.2% for Argon, just 65.0% for Fable. Nearly 20 points separate them when the input gets truly enormous. For anyone processing massive documents, codebases, or datasets in a single pass, this is the decisive category.
What makes these margins significant is not just their size but their consistency. Argon does not win knowledge work on one lucky test. It wins on reasoning, automation, finance, and legal tasks alike, which suggests a genuinely stronger foundation for professional work rather than a narrow specialization. Enterprises evaluating models for deployment tend to weight exactly this kind of breadth, since real workflows rarely stay inside a single benchmark's boundaries.
Where Fable Keeps the Fight Alive
Credit where it is due: Claude Fable 5.1 is not embarrassed everywhere. Its two wins both come in agentic coding, and they are legitimate victories, not rounding errors.
Hard Engineering goes to Fable, 56.3% to Argon's 55.0%. Terminal Agents goes to Fable as well, 57.9% to 57.4%. These are razor-thin margins, half a point each, but a win is a win. They suggest that on certain demanding software engineering tasks, Fable's training gives it a genuine edge, or at minimum parity with a small advantage.
Argon still takes the other two coding tests. Real-World Coding is a clear Argon win at 77.9% versus 67.4%, a gap of more than 10 points. Vibe Coding is closer, 91.9% to 90.3%, but still Argon's. So the coding category splits 2-2, the only category in this entire comparison that Fable does not lose outright.
The multimodal and security results, meanwhile, contain Argon's single most dominant performance. Chart Reading ends 71.6% to 46.2%, a gap of more than 25 points. That is the largest margin in the whole comparison. Video Understanding goes 91.7% to 79.7%, and the Security test lands at 68.0% versus 58.0%. Fable is competitive on none of these.
Then there is the missing data. Fable 5.1 was not tested on either computer use benchmark, the Agent Exam or Real Desktop tests. Argon scored 39.5% and 69.2% on those respectively, but with no Fable numbers to compare against, the category simply does not count. It is worth noting, though, that Anthropic has historically positioned Claude as strong at computer use. The absence of scores here leaves an open question rather than a conceded defeat.
It is also worth putting Fable's position in perspective. Losing 15-2 to the current benchmark leader is not the same as being a weak model. Fable 5.1 remains a highly capable system, and its narrow coding victories show it can still trade blows with Argon on tasks that matter to developers. The gap looks enormous partly because Argon is genuinely exceptional on this evaluation suite, not because Fable has collapsed. Against most other models in the field, Fable would likely look far more competitive.
The Complete Benchmark Breakdown
Here is every shared test, side by side. Fifteen wins for Argon, two for Fable, and two tests where Fable never entered the ring.
| 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 | Finance 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 Agents | 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 Tokens | 99.7% | 91.4% | Argon |
| Long Context | 1M Tokens | 84.2% | 65.0% | Argon |
| Computer Use | Agent Exam | 39.5% | Not tested | Not contested |
| Computer Use | Real Desktop | 69.2% | Not tested | Not contested |
| Multimodal + Cybersecurity | Chart Reading | 71.6% | 46.2% | Argon |
| Multimodal + Cybersecurity | Video Understanding | 91.7% | 79.7% | Argon |
| Multimodal + Cybersecurity | Security | 68.0% | 58.0% | Argon |
The final tally: Argon 15, Fable 5.1 2, with 2 tests uncontested. By the numbers, this is the most decisive result in the series so far.
But the honest read is more nuanced than the scoreline. Fable's two coding wins, however narrow, prove it belongs in the conversation on software engineering tasks. The untested computer use category leaves a real gap in the comparison, and Anthropic's next model will surely address it. And benchmarks, however carefully designed, measure specific tasks under specific conditions. They do not capture writing quality, conversational feel, or the ecosystem around each model, all of which matter to real users.
Still, on the published evidence, the verdict is clear. Against Claude Fable 5.1, Gemini 4 Argon does not just win. It dominates.
For readers following this series, the pattern is now established: Argon leads everywhere, but the size of the lead depends on the opponent. GPT-6 pushed back hard in coding and terminals. Fable could only manage two narrow coding wins. The question for the next matchup is whether any rival can take a full category, not just a test or two.
Top comments (0)