Gemini 4 Argon vs GPT 6.1 and Claude Opus 5.5: Who Wins the Benchmarks?
Gemini 4 Argon takes on GPT 6.1, Claude Opus 5.5, and Muse Wonder across Deep Sweep, Terminal Bench, and Artificial Analysis scores.
What is Gemini 4 Argon and how does it compare to GPT 6.1 and Claude Opus 5.5?
Gemini 4 Argon is Google’s latest frontier model, built for long, complex tasks rather than quick chat replies. On coding benchmarks it leads or sits near the top of the pack, topping Deep Sweep (also called Deep Sue in some early coverage) among the models tested. On Artificial Analysis’s combined intelligence score it lands around 53, putting it roughly even with Fable 5.1 but slightly behind Opus 5.5, Sonnet 5.5, and Meta’s newly released Muse Wonder. It undercuts most rivals on cost per task.
TL;DR
- Deep Sweep leadership puts Gemini 4 Argon at the top of one widely-watched agentic coding benchmark, scoring 77.9%, just behind Muse Wonder’s 79% in one comparison chart and ahead of it in another cut of the same data depending on which models are included.
- Terminal Bench results show Argon at 57%, trailing both Claude Opus 5.5 (66.4%) and Muse Wonder, meaning command-line execution tasks remain a relative weak spot compared to coding generation.
- Artificial Analysis scoring places Argon around 53, on par with Fable 5.1 but behind Opus 5.5, Sonnet 5.5, and Muse Wonder on the broader intelligence index.
- Output capacity stands out with a one million token output limit, far beyond what most frontier models offer on the output side of a context window.
- Pricing undercuts rivals at launch: $2 per million input tokens and $10 per million output tokens, about half of Opus 5.5’s rate and roughly in line with Sonnet 5.5 and GPT 6.1, though prices rise to $4/$20 per million starting in December.
- 3D generation lags behind Claude Opus 5.5 and GPT 6 Astra specifically, even though Argon still beats Meta’s Muse Wonder at 3D tasks, which Meta’s own team has acknowledged as an area they didn’t optimize for.
- Agentic and legal benchmarks also favor Argon, with state-of-the-art marks on the Vals Index (covering finance, legal, and tax tasks) and on Automation Bench.
How does Gemini 4 Argon perform on coding benchmarks?
Coding is where Argon makes its strongest case. It holds the top spot on Deep Sweep, a benchmark widely used to gauge how real engineers experience a model’s agentic coding ability in practice rather than synthetic test scores alone. Depending on which comparison set gets used, Argon scores either just ahead of or just behind Muse Wonder, with both models clustered near the top of the field above Opus 5.5, Astra, and GPT 6.1 Soul.
On Vibe Code Bench, Argon scores 91.9%, the highest mark among the models compared, suggesting it’s particularly strong at generating functional, iterative code from loose prompts.
Terminal Bench tells a different story. This benchmark measures how well a model can actually write and execute terminal commands, a more operational skill than writing code in isolation. Here Argon scores 57%, trailing Opus 5.5’s 66.4% and also falling behind Muse Wonder and Astra. It’s a real gap, though not one that knocks Argon out of frontier contention, since all of these models are separated by single-digit to double-digit percentage points rather than by a generational gulf.
How does it do on agentic and reasoning tasks?
Beyond raw coding, Argon was evaluated on tasks that require operating semi-autonomously. On OS World, which tests a model’s ability to control a computer environment, Argon scores 69% against GPT 6.1 Soul’s 71.4%, a near-even result. On Agents Last Exam, Argon scores 39 versus Opus 5.5’s 38, a narrow edge.
Argon also posts state-of-the-art results on the Vals Index, a benchmark suite covering professional knowledge work like finance, legal analysis, and tax preparation, and on Automation Bench, which evaluates how well a model can carry out multi-step automated workflows. Separately, enterprise document-processing benchmarks from Box show a 14-point jump (from 71% to 85%) over the prior Gemini 3.1 Pro model across tasks like spreadsheet evaluation, PDF extraction, and parsing presentations and images, indicating the leap from Gemini 3.1 to Argon is substantial in practical knowledge-work settings, not just synthetic coding tests.
Where does Gemini 4 Argon fall short compared to rivals?
Two areas stand out as relative weaknesses. First, Terminal Bench performance lags Opus 5.5 by roughly nine points, meaning tasks requiring precise command-line execution are a comparative soft spot.
Second, and more visibly, 3D generation and game-style outputs. When used to build 3D scenes, physics simulations, or game clones using tools like Three.js or Unreal Engine, Argon produces solid but not best-in-class results. Hands-on testing found its game clones playable and its physics simulations functional, but visually and aesthetically behind Claude Opus 5.5 and GPT 6 Astra, both of which are described as producing more polished 3D output. Argon does outperform Meta’s Muse Wonder at 3D tasks specifically, an area Meta’s own head of AI, Alexander Wang, has acknowledged the team didn’t prioritize during training.
Remy is new. The platform isn't.
Remy is the latest expression of years of platform work. Not a hastily wrapped LLM.
On the Artificial Analysis combined score, Argon’s position around 53 places it in the same tier as Fable 5.1 but behind Opus 5.5, Sonnet 5.5, and Muse Wonder, suggesting that while Argon is a clear jump for Google, it isn’t an outright leader on every general-intelligence measure.
Is Gemini 4 Argon worth using over GPT 6.1 or Claude Opus 5.5?
The honest answer depends on the task. For agentic coding workflows, especially ones measured by Deep Sweep or Vibe Code Bench, Argon is competitive with or ahead of both GPT 6.1 and Opus 5.5. For terminal-heavy automation or polished 3D/visual generation, Opus 5.5 currently holds an edge. For knowledge work like legal review, tax analysis, or document extraction at scale, Argon’s state-of-the-art Vals Index and Automation Bench results plus the Box benchmark jump make it a strong contender.
Price is a meaningful factor. At launch, Argon’s API pricing ($2 per million input tokens, $10 per million output tokens) undercuts Opus 5.5 by roughly half and lands close to Sonnet 5.5 and GPT 6.1 Soul. That’s a real cost advantage for teams running high-volume coding or automation workloads, though the introductory pricing is temporary. Starting in December, rates increase to $4 and $20 per million tokens respectively, which narrows but doesn’t erase the price gap with rivals.
The million-token output ceiling is also distinctive. Most frontier models’ context windows are dominated by input capacity, with output capped much lower. An output limit around a million tokens (roughly 750,000 words) opens up tasks like generating very long documents, extensive codebases, or large structured datasets in a single pass, something none of the directly compared models currently match.
Frequently Asked Questions
What is Gemini 4 Argon’s score on Deep Sweep?
Argon scores 77.9% on Deep Sweep in one comparison set that includes Muse Wonder and GPT 6.1 Soul, placing it just behind Muse Wonder’s 79%. In an earlier comparison chart without those two models, Argon was reported as the new state-of-the-art leader on the benchmark.
How does Gemini 4 Argon’s pricing compare to GPT 6.1 and Claude Opus 5.5?
At introductory pricing, Argon costs $2 per million input tokens and $10 per million output tokens via the API, about half the cost of Opus 5.5 and roughly equal to Sonnet 5.5 and GPT 6.1 Soul. Prices increase to $4 and $20 per million tokens starting in December.
Is Gemini 4 Argon better than Claude Opus 5.5 at coding?
It depends on the specific benchmark. Argon leads on Deep Sweep and Vibe Code Bench, but Opus 5.5 leads on Terminal Bench (66.4% versus Argon’s 57%), which measures actual command execution rather than code generation alone.
Does Gemini 4 Argon beat Muse Wonder at 3D generation?
Yes. Hands-on testing found Argon produces better 3D outputs (Three.js scenes, Unreal Engine environments) than Meta’s Muse Wonder, though both fall short of Claude Opus 5.5 and GPT 6 Astra in visual polish and realism.
Where can I access Gemini 4 Argon?
It’s available to Google AI Pro and Ultra subscribers in the Gemini app and in AI mode in Search. Access inside Google’s Antigravity coding platform requires an Ultra subscription. It’s also available through Google’s API at the pricing listed above.
