Models › Text › Gemini 4 Argon
Google · Model details, evidence, and API facts
Gemini 4 Argon is Google’s flagship reasoning model for engineering
Starting provider API rate; context tiers apply
See what Gemini 4 Argon does best, then review its limits, benchmarks, and pricing below
Gemini 4 Argon handles complex software engineering, research, and cybersecurity
Every Gemini 4 Argon result below comes from a named source, labeled clearly by protocol, so each score reflects how it was measured
| Benchmark | Score | Evaluation details | Source |
|---|---|---|---|
|
AutomationBench-AA
Tests Gemini 4 Argon on 657 multi-step tasks across six simulated SaaS business domains.
|
51.3%
Guardrail-aware score
|
Version: AutomationBench (Zapier) • Split: 657 tasks across six business domains • Google DeepMind model card, October 2026 • Top score of the four models compared (GPT-6 Astra 41.4%, Claude Opus 5.5 42.5%, Claude Fable 5.1 31.4%) • Artificial Analysis's guardrail-aware AutomationBench-AA variant reads 78%. |
Methodology
Google
Provider-reported |
|
DeepSWE 1.1
Measures Gemini 4 Argon on 113 long-horizon coding tasks graded by tests.
|
77.9%
Pass@1
|
Version: DeepSWE 1.1 • Split: 113 tasks across 91 repositories • Google self-computed, highest thinking setting • Leads GPT-6 Astra 74.1% and Claude Opus 5.5 74.2%. |
Methodology
Google
Provider-reported |
|
GPQA Diamond
Tests Gemini 4 Argon on 198 Google-proof graduate-level questions in biology, chemistry, and physics to probe scientific reasoning.
|
96.4%
Accuracy (high effort)
|
Version: GPQA Diamond 198 • Split: 198 Diamond questions • Evaluator: Epoch AI Benchmarking Hub, high reasoning effort • Ranked #1 of 295 configurations • Google has not published its own GPQA Diamond figure. |
Methodology
Epoch AI
Independently run |
|
Humanity's Last Exam
Tests Gemini 4 Argon on 2,500 expert-level questions across many academic fields, with tool use enabled.
|
57.1%
Accuracy (high effort)
|
Version: Humanity's Last Exam • Evaluator: Artificial Analysis, high reasoning effort • Ranked #5 of configurations measured • Google has not published its own HLE figure; HLE-Diamond (AA) reads 56.7%. |
Methodology
Artificial Analysis
Independently run |
|
Terminal-Bench 4.0
Tests Gemini 4 Argon on terminal-based tasks spanning software engineering, system configuration, and data analysis, rewarding long-horizon tool use and error recovery.
|
57.4%
Independent task success rate
|
Version: Terminal-Bench 4.0 • Split: 66 tasks • Google self-computed, highest thinking setting • Trails Claude Opus 5.5's 66.4% • Artificial Analysis's independent run reads 57.1%, Vals AI 57.6%. |
Methodology
Google
Provider-reported |
Arena rankings capture how people vote between anonymous responses. Gemini 4 Argon’s preference standing is reported per leaderboard, since results can vary widely by task
Gemini 4 Argon ranked 11th on Arena’s WebDev leaderboard for apps in November, scoring 1683 from 4,677 votes
Gemini 4 Argon offers a 1,048,576-token context window, 65,536 output tokens, image input, and adaptive thinking, with full API pricing and tiers shown below
| Model | |
| Provider | |
| Release date | Oct 13, 2026 (generally available; announced Sep 30, 2026) |
| Catalog added | [Pyxa catalog date] |
| Context window | 1,048,576 tokens |
| Maximum output | 65,536 tokens (up to 1,048,576 with long decode) |
| Knowledge cutoff | March 2026 |
| License | Proprietary |
| Capabilities and API | |
| Input types | Text, Image, Video, Audio, PDF |
| Output types | Text |
| Reasoning | Supported (thinking levels low, medium, high; default medium) |
| Provider endpoint tool parameters | Accepts tool parameters (function calling) |
| Base provider API input | $2.00 / 1M tokens (introductory, through Nov 30, 2026) |
| Base provider cached input (read) | $0.10 / 1M tokens (implicit cached input, through Nov 30, 2026) |
| Base provider API output | $10.00 / 1M tokens (including thinking, through Nov 30, 2026) |
| Standard rate from Dec 1, 2026 | $4.00 input, $20.00 output, $0.20 cached input per 1M tokens — introductory rates double on Dec 1, 2026 |
Provider token rates are context-dependent. The base rates apply below the listed prompt thresholds; the matching tier applies at or above each threshold. These are reference provider API rates, not Pyxa AI checkout charges — access in Pyxa AI follows the allowances of your Pyxa AI plan.
Each Gemini 4 Argon benchmark and specification links to its source, so you can check where every figure comes from
Discover where Gemini 4 Argon performs best, from long-horizon coding and agents to deep daily scientific problem solving
Put Gemini 4 Argon to work on coding, research, and cybersecurity tasks