Compare models

Context, limits, capabilities, price and use cases side by side — up to 4 models at a time. The URL holds your selection, so you can share it.

Gemini 4 ArgonGemini 3.8 FlashGemini 3.1 Pro Preview Add a modelChat with these 3

AttributeHide identical rows	Gemini 4 Argongemini/gemini-4-argon50% off	Gemini 3.8 Flashgemini/gemini-3.8-flash50% off	Gemini 3.1 Pro Previewgemini/gemini-3.1-pro-preview
Specs
Context window	1M	1M	1M
Max output	66K	66K	66K
Input modalities	TextImageAudioVideo	TextImageAudioVideo	TextImageAudioVideo
Output modalities	Text	Text	Text
Capabilities
Streaming	Supported	Supported	Supported
Tool calling	Supported	Supported	Supported
Structured output (JSON)	Supported	Supported	Supported
Image input	Supported	Supported	Supported
Thinking mode	Supported	Supported	Supported
Price / 1M tokens
Input	$ 4.00$ 2.00	$ 1.50$ 0.75	$ 2.00>= 200K$ 4.00
Output	$ 20.00$ 10.00	$ 7.50$ 3.75	$ 12.00>= 200K$ 18.00
Cache read	$ 0.20$ 0.10	$ 0.15$ 0.075	$ 0.20>= 200K$ 0.40
Cache write	$ 4.00$ 2.00	$ 1.50$ 0.75	$ 2.00>= 200K$ 4.00
What it is good at
Best for	Agentic software engineering Mathematical and scientific reasoning Multimodal document analysis Specialist knowledge work	Long-horizon software engineering Complex enterprise workflows Long multimodal workflows	Advanced coding agents Large multimodal corpora
Published benchmarks
Terminal-Bench 2.0pass_rate	—	—	68.5%effort high · Terminus 2(setup not fully disclosed)
Terminal-Bench 2.1pass_rate	—	89.4%(setup not fully disclosed)	—
Terminal-Bench 4.0pass_ratesetups differ by: pass_k, reasoning	57.4%effort high(setup not fully disclosed)	19.1%(setup not fully disclosed)	—
BrowseComp (with tools)accuracy	—	—	85.9%effort high · tools: browse, python, search(setup not fully disclosed)
DeepSWE 1.1resolved_ratesetups differ by: harness, pass_k, reasoning	77.9%effort high · mini-swe(setup not fully disclosed)	73.7%(setup not fully disclosed)	—
SWE-Bench Verifiedresolved_rate	—	—	80.6%effort high(setup not fully disclosed)
Agents' Last Exam (ALE-Claw, binary pass)pass_rate	39.5%effort high · ALE-Claw(setup not fully disclosed)	—	—
OSWorld 2.0 (partial credit)success_rate	—	59%(setup not fully disclosed)	—
OSWorld 2.0 (v2026.08.08 offline, partial score)score	69.2%effort high · Gemini CUA · 3 runs · Compaction following Claude Opus 5.5 system card(setup not fully disclosed)	—	—
Finance Agent v2score	—	61.4%(setup not fully disclosed)	—
GDPval-AA v2elo	—	1545(setup not fully disclosed)	—
GraphWalks BFS 256K to 1Mf1	84.2%effort high(setup not fully disclosed)	—	—
GraphWalks BFS up to 128Kf1	99.7%effort high(setup not fully disclosed)	—	—
PostTrainBench 1.1score	45.3%effort high · OpenCode(setup not fully disclosed)	—	—
CharXiv Reasoning (no tools)accuracy	—	86.2%(setup not fully disclosed)	—
HLE-Verifiedaccuracy	—	54.9%(setup not fully disclosed)	—
Humanity's Last Examaccuracy	—	—	44.4%effort high(setup not fully disclosed)
GPQA Diamondaccuracy	—	—	94.3%effort high(setup not fully disclosed)
LABBench 2score	88.8%effort high(setup not fully disclosed)	—	—
Terminal-Bench-Science 0.1accuracy	57.6%effort high(setup not fully disclosed)	—	—
LVBench (agentic)accuracy	—	87.8%(setup not fully disclosed)	—
LVBench (static)accuracysetups differ by: pass_k, reasoning	91.7%effort high(setup not fully disclosed)	87.1%(setup not fully disclosed)	—