Gemini · Model guide
Gemini 4 Argon
Gemini 4 Argon is Google’s flagship reasoning model for complex software engineering, mathematical and scientific problem solving, and multi-step agents. It accepts text, images, audio, and video and returns text within a shared 1M-token input and output window.
What this model is good at
What the vendor positions it for, and which of our models to reach for instead.
VendorGoogle positions Argon for complex engineering, multi-turn investigations, multimodal document analysis, and scientific reasoning.source
Agentic software engineering · Google identifies complex engineering and multi-turn investigations as intended uses.(vendor claim)source
Mathematical and scientific reasoning · Google identifies scientific and mathematical reasoning as intended uses.(vendor claim)source
Multimodal document analysis · Google identifies analysis and synthesis of multimodal documents as intended uses.(vendor claim)source
Specialist knowledge work · Google highlights legal and financial analysis and technical writing.(vendor claim)source
- gemini/gemini-3.8-flashcheaper —Choose Gemini 3.8 Flash for multimodal and agent workflows when lower token cost matters more than using the latest flagship.source
- gemini/gemini-3.1-pro-previewpredecessor —Keep Gemini 3.1 Pro Preview when an existing Gemini 3 integration has already been validated and you want to compare it with Argon before migrating.source
Pricing and billing
Billed per token. Cached input is charged at the cache-read rate.
| Price / 1M tokens | You pay |
|---|---|
| Input | $ 4.00$ 2.00 |
| Output | $ 20.00$ 10.00 |
| Cache read | $ 0.20$ 0.10 |
| Cache write | $ 4.00$ 2.00 |
+ 2.0K × $ 10.00
Capabilities and limits
What CrossModel guarantees across every route this model can take right now.
- Context window
- 1.0M tokens
- Max output
- 65.5K tokens
- Input / output modalities
- Text + Image + Audio + Video → Text
- Streaming
- Supported
- Tool calling
- Supported
- Structured output (JSON)
- Supported
- Image input
- Supported
- Thinking mode
- Supported· always on
- Reasoning effort
- none · minimal · low · medium · high
- Available endpoints
- /v1/chat/completions · /v1/responses · /v1/messages
Published benchmarks
Scores the vendor reported, with the evaluation setup each one came from.
- Agents' Last Exam (ALE-Claw, binary pass)pass_rate·Google DeepMind · 2026-10-13effort highALE-Clawsetup not fully disclosed
Google self-computed binary pass rate; default ALE-Claw harness, five-hour window, safety filters enabled. Flagged responses return empty strings; episodes can continue. Full reproduction settings not disclosed. Methodology: https://deepmind.google/models/evals-methodology/gemini-4-argon
39.5% - OSWorld 2.0 (v2026.08.08 offline, partial score)score·Google DeepMind · 2026-10-13effort highGemini CUA3 runsCompaction following Claude Opus 5.5 system cardsetup not fully disclosed
Google self-computed partial score on offline subset, maximum over three single-attempt runs. Official 08.08 patch/evaluator; 1080p, max 500 steps; parallel tool calls, pyautogui screenshot-only observation, safety filters enabled. Not plain OSWorld 2.0 success rate. Full reproduction settings not disclosed. Methodology: https://deepmind.google/models/evals-methodology/gemini-4-argon
69.2%
- GraphWalks BFS 256K to 1Mf1·Google DeepMind · 2026-10-13effort highsetup not fully disclosed
Google self-computed F1 on 200 problems with context lengths from 256K to 1M tokens. Full reproduction settings not disclosed. Methodology: https://deepmind.google/models/evals-methodology/gemini-4-argon
84.2% - GraphWalks BFS up to 128Kf1·Google DeepMind · 2026-10-13effort highsetup not fully disclosed
Google self-computed F1 on 650 problems with context length up to 128K tokens; not comparable with the 256K-to-1M subset. Full reproduction settings not disclosed. Methodology: https://deepmind.google/models/evals-methodology/gemini-4-argon
99.7%
- PostTrainBench 1.1score·Google DeepMind · 2026-10-13effort highOpenCodesetup not fully disclosed
Google self-computed. 10-hour budget on one NVIDIA H100; weighted aggregate across four base models and seven benchmarks. Full reproduction settings not disclosed. Methodology: https://deepmind.google/models/evals-methodology/gemini-4-argon
45.3%
- LABBench 2score·Google DeepMind · 2026-10-13effort highsetup not fully disclosed
Google self-computed with Linux terminal, preinstalled bioinformatics tools, Python, R and internet access. Full reproduction settings not disclosed. Methodology: https://deepmind.google/models/evals-methodology/gemini-4-argon
88.8% - Terminal-Bench-Science 0.1accuracy·Google DeepMind · 2026-10-13effort highsetup not fully disclosed
Google self-computed with 6x verifier timeout to address verification timeout issues; differs from standard timeout runs. Full reproduction settings not disclosed. Methodology: https://deepmind.google/models/evals-methodology/gemini-4-argon
57.6%
- LVBench (static)accuracy·Google DeepMind · 2026-10-13effort highsetup not fully disclosed
Google self-computed without tools; Gemini video sampled at 1 FPS. Other models use different frame budgets due to API limits. Full reproduction settings not disclosed. Methodology: https://deepmind.google/models/evals-methodology/gemini-4-argon
91.7%
Vendor-reported numbers, not CrossModel measurements. Scores are only comparable when the benchmark version, metric and evaluation setup match, so nothing here is averaged or ranked.
Use it in your tools
Point the base URL at CrossModel and paste this model ID — every tool below has a setup guide.
Related models
Other tiers from the same vendor, plus similarly priced alternatives.
Frequently asked questions
What is Gemini 4 Argon?
How much does Gemini 4 Argon cost?
Does Gemini 4 Argon support tool calling and structured output?
Which endpoint do I call?
Can I try it without writing code?
Chat first, integrate later
Chat first, then wire it in once you like the answers.