OutYet reporting
Gemini 4 Argon’s benchmark lead depends on the task
Google’s model card shows strong results in some areas and clear losses in others. Independent Vals AI data also puts Argon behind two Claude models on its latest overall index.
Google’s Gemini 4 Argon is built around long tasks: its API documentation specifies a shared 1 million-token window for input, thinking, and output, plus adjustable low, medium, and high thinking levels. The practical appeal is control over how much reasoning a request gets and how much material can fit in one interaction. But the large window is not a promise of uniformly better answers. Google’s own model card reports a mixed benchmark picture, including wins on DeepSWE v1.1 and GraphWalks, and losses on FrontierSWE v2, Terminal-Bench 4.0, and the offline OSWorld-2.0 subset.
The API docs also make the cost and migration tradeoffs unusually concrete. Introductory prices are $2 per million input tokens and $10 per million output tokens through November 30, then rise to $4 and $20; thinking tokens count as output. Implicit caching discounts cached input by 95 percent, but explicit context caching is unsupported. Google says temperature, top_p, and top_k are ignored, and recommends testing all three thinking levels. Teams moving an existing integration therefore need to check request parameters, caching behavior, output limits, and cost assumptions instead of treating the new model as a drop-in upgrade.
Google’s model card compares Argon with GPT-6 Astra, Claude Fable 5.1, and Claude Opus 5.5 using results reported as of October 2026. Argon leads the card’s DeepSWE v1.1, Vibe Code Bench, and GraphWalks rows, but GPT-6 Astra scores higher on FrontierSWE v2, Terminal-Bench Science, and OSWorld-2.0, while Opus 5.5 leads Terminal-Bench 4.0 and PostTrainBench. Those are provider-published comparisons, so they are useful for identifying where Google claims an advantage, not a substitute for running the same workload and setup yourself.
The independent Vals AI index offers a less favorable overall comparison than Google’s release-era framing. In its October 22 update, Vals ranks Claude Sonnet 5.5 first at 70.85 percent, Claude Opus 5.5 second at 70.18 percent, and Argon third at 69.24 percent. The index combines finance, coding, legal, and tax tasks, weighted by their share of U.S. GDP, and uses six private and two public benchmarks. Vals also notes that some Sonnet and Opus component tasks used fallback models after provider refusals; counting those as failures would put Argon first. That qualification matters, as does the fact that this is one composite with proprietary components, not a universal measure of model quality.
For technical teams, the useful takeaway is to choose a workload before choosing a headline score. Argon’s extended context and controllable thinking may help with long agent runs, but large output budgets can make cost rise quickly, and Google lists possible hallucinations, slowness, and timeouts as limitations. The independent index reports a $13.37 cost per test for Argon under its own benchmark setup, while the API’s advertised token price changes in December. Those numbers measure different things. A fair pilot should pin the model ID and date, compare matched prompts and tools against the current model, and track task success, retries, latency, and total billed tokens at each thinking level.
Related models
Sources
- What’s new in Gemini 4 Argon · Google AI for Developers
- Gemini 4 Argon Model Card · Google DeepMind
- Vals Index Leaderboard and Methodology · Vals AI