Gemini——Google's Native Multimodal Flagship: 1M Context and Scientific Reasoning Champion

TL;DR Gemini is Google DeepMind's native multimodal LLM family, famed for a 1M-token context window and native video/speech input plus scientific reasoning. 3.1 Pro tops GPQA Diamond 94.1% and ARC-AGI-2 77.1% to claim science-reasoning dual crowns, at $2/$12—1/6 of Claude. 3.7 Flash delivers near-Pro agent capability for $0.75/$3.75.

#ai-agent #llm #gemini #google-deepmind #model-family-gemini #multimodal #reasoning #model-selection

Table of Contents

Family Evolution Timeline
Two Product Lines: Closed Gemini for Revenue, Open Gemma for Ecosystem
Architecture: Two Walls—Native Multimodality and 1M Context
Gemini 3.1 Pro and 3.7 Flash—How to Choose
1. License Trap: Strongest Models Locked to Google Infra
2. Performance Position
Sub-lines & Ecosystem: A Table of All Gemini Products
Position Against Competitors
What This Means for Agent Developers
Overall
References

Series: AI Model Families (4 / 10)

GPT——Closed API for Revenue, Open GPT-OSS for Ecosystem: the Unified Routing Platform Behind the World's Largest AI Service Llama——From Open-Source Experiment to the Most Deployed Open LLM, and Meta's Closed-Source Pivot

🌏 中文版

In December 2023, Google DeepMind released Gemini 1.0—the first "native multimodal" model that integrated text, image, video, and speech from pre-training. Two and a half years later, in February 2026, Gemini 3.1 Pro topped GPQA Diamond at 94.1% to become the strongest science-reasoning model, and hit 77.1% on ARC-AGI-2 as the highest Fluid Intelligence score. At $2/$12, roughly one-sixth of Claude Opus. This is the fourth family deep-dive in the "AI 模型家族" series, tracing Gemini's evolution from 1.0 to 3.7 Flash.

For how to read the benchmark numbers cited here, see our AI model evaluation sources guide. This is part of the AI Model Landscape Overview series.

Family Evolution Timeline

Version	Date	Context	Key Milestones
Gemini 1.0 Ultra/Pro/Nano	2023-12	32K	First native multimodal model
Gemini 1.5 Pro	2024-02	1M	Industry-first 1M context
Gemini 1.5 Flash	2024-05	1M	Lightweight, speed-first
Gemini 2.0 Flash	2024-12	1M	Agent capability, tool use
Gemini 2.5 Pro	2025-03	1M	Major reasoning boost
Gemini 2.5 Flash	2025-04	1M	Cost-effective Flash
Gemini 3.0 Pro	2025-11	1M	Gemini 3 debut
Gemini 3.1 Pro	2026-02	1M	ARC-AGI-2 77.1%, GPQA 94.1%
Gemini 3.1 Flash-Lite	2026-03	1M	$0.25/$1.50, fastest & cheapest
Gemini 3.5 Flash	2026-05	1M	Agent capability near Pro
Gemini 3.6 Flash	2026-07	1M	Continuous iteration
Gemini 3.7 Flash	2026-08	1M	Smartest Flash, $0.75/$3.75

Two and a half years, 12 milestones. Gemini's through-line: from multimodality to agents, from flagship to full-size coverage. 1M context was Gemini's earliest moat, and every subsequent generation pushed harder on multimodal understanding and agent capability.

Two Product Lines: Closed Gemini for Revenue, Open Gemma for Ecosystem

To understand Gemini's 2026 layout, split it into two lines—much like GPT's dual track:

Closed API Line (AI Studio / Vertex AI): The entire Gemini 3 family runs only on Google infrastructure, no downloadable weights. Since the 1.5 era, 1M context has been standard; pricing starts at $2/$12. This line owns revenue—Google's enterprise and consumer traffic all runs here, backed by TPU hardware and a global datacenter footprint.

Open-Weight Line (Gemma 2 / 3, 2B–27B): Apache 2.0, downloadable, fine-tunable, self-hostable. This line owns ecosystem—giving developers needing self-host, fine-tuning, and data sovereignty a path, but explicitly not the frontier model.

Google wrote the Transformer paper in 2017, acquired DeepMind in 2014, and in December 2023 merged Brain and DeepMind into Google DeepMind with Gemini as the unified brand—consolidating research to counter OpenAI and Microsoft. Gemma's open source is strategic gap-filling, not a reversal—the closed-flagship main line has never loosened.

Architecture: Two Walls—Native Multimodality and 1M Context

Native Multimodality

Gemini has been native multimodal since v1—text, image, video, speech trained together from pre-training, not a vision adapter bolted onto a text model after the fact.

This means Gemini truly "understands" video and speech, not just transcribing them to text first. 3.1 Pro supports:

Text: standard LLM capability
Image: understanding, analysis, reasoning
Video: directly processes video clips (not screenshots)
Speech: native speech understanding and generation
PDF: direct document parsing

1M Context Window

February 2024, Gemini 1.5 Pro became the industry's first 1M-token model. Not a gimmick—1M tokens ≈ 10 full novels, 30K lines of code, or five hours of speech transcription.

By 3.1 Pro, 1M context is mature and stable. Note: beyond 200K tokens, pricing doubles ($4/$18). If your workload routinely exceeds 200K, effective cost is far above $2/$12.

Deep Think

Gemini's deep reasoning mode, analog to Claude's extended thinking or GPT's reasoning effort. 3.1 Pro's Deep Think jumped ARC-AGI-2 from 31.1% (3 Pro) to 77.1%—one of 2026's largest single capability leaps.

Gemini 3.1 Pro and 3.7 Flash—How to Choose

February's flagship vs August's workhorse:

Item	Gemini 3.1 Pro	Gemini 3.7 Flash	Gemini 3.1 Flash-Lite
Positioning	Flagship reasoning, science/research	Agent workhorse, balanced	High throughput, lowest cost
Input ($/MTok)	$2 (≤200K) / $4 (>200K)	$0.75 (promo thru 2026/12)	$0.25
Output ($/MTok)	$12 (≤200K) / $18 (>200K)	$3.75 (promo thru 2026/12)	$1.50
Context caching	$0.20 / $0.40	$0.075	$0.025
Context	1M	1M	1M
Max output	64K	64K	64K
Deep Think	✓	✓	✓
Native multimodality	✓ (text / img / video / speech / PDF)	✓	✓ (text / img / video, speech extra)

Pricing and specs from Gemini Developer API Pricing and Gemini API Models.

License Trap: Strongest Models Locked to Google Infra

Google mirrors OpenAI—closed flagship + small open model dual track:

Closed flagship (all Gemini 3.1): API only (AI Studio / Vertex AI), no weights, license = Google Terms of Service
Open weights (Gemma 2 / 3, 2B–27B): Apache 2.0, downloadable, fine-tunable, self-hostable—but only small-mid, not frontier
Enterprise: Gemini on Vertex AI governed by Google Cloud contracts; data-processing location selectable per region

The catch: Gemini's strongest models are almost entirely bound to Google infrastructure. Want 1M context or native video understanding? Must go through Google API or Vertex. Gemma is Apache 2.0 but at 27B is far behind Gemini 3.1 Pro. If you need "Apache 2.0 + frontier quality," Gemini can't deliver—look at DeepSeek (MIT) or Qwen (mostly Apache).

Another trap: Google ecosystem lock-in—Gemini's depth advantages (search, YouTube, Maps, Workspace) only shine inside Google's world. Outside, it reverts to an ordinary frontier model. And a pricing trap: the 200K threshold—once total input exceeds 200K tokens, the entire request (including output) is billed at the long-context rate. 199K input → $2/$12, 201K → $4/$18 (entire request, not just the excess). Plus Gemini 3.7 Flash's $0.75/$3.75 expires 2026/12/31, then reverts to $1.50/$7.50.

Performance Position

Metric	Gemini 3.1 Pro	Comparison
GPQA Diamond	94.1%	GPT-5.6 Sol 94.6%; Claude Fable 5 92.6%; Opus 4.8 92%
ARC-AGI-2	77.1% (Deep Think)	Claude Opus 4.8 68.8%; Sonnet 4.6 60.4%; GPT-5.2 52.9%
SWE-bench Verified	80.6%	Claude Opus 5 96%; DeepSeek V4 Pro 96.4%; Fable 5 95%—Gemini lags ~15pp
MMMU-Pro	80.5~82%	Qwen3-VL-235B 69.3%; GPT-5.4 81%
DocVQA	92%	Qwen3-VL-235B 96.5%; GPT-5.4 95%—Qwen open actually stronger on docs

Direct comparison:

Metric	Gemini 3.1 Pro	Claude Fable 5	GPT-5.6 Sol	DeepSeek V4 Pro
GPQA Diamond	94.1%	92.6%	94.6%	—
ARC-AGI-2	77.1%	—	—	—
SWE-bench Verified	80.6%	95%	~95%	96.4%
MMMU-Pro	82%	—	—	—
Context	1M	1M	1.05M	1M
Max output	64K	128K	128K	128K
Native video/speech	✓	✗	✗	✗
Output pricing ($/MTok)	$12	$50	$30	$0.87

Gemini's unique combo is science reasoning + native multimodality + low pricing, but 64K output cap and SWE-bench lag are real limits.

Sub-lines & Ecosystem: A Table of All Gemini Products

Sub-line	Representative	Positioning
Flagship reasoning	Gemini 3.1 Pro / Flash	Highest intelligence / high efficiency
Lightweight	Gemini 3.1 Flash-Lite	High speed, low price, large-scale classification
Open weights	Gemma 3 (2B–27B)	Apache 2.0, self-hostable
Image	Imagen 4	Text-to-image
Speech / audio	Chirp / Lyria	Transcription / music generation
Video	Veo 3	Text-to-video
Embeddings	Gecko / text-embedding	RAG backbone
Dev tools	AI Studio / Vertex AI Agent Builder	Agent infra
Consumer product	Gemini App / NotebookLM	World's second-largest AI app

Two observations:

Multimodal breadth is Gemini's most underestimated asset. Text, image, video, speech, code—Gemini's native multimodal coverage is the most complete of any competitor, fused from pre-training (not bolted adapters). For agent scenarios needing "understand video + hear speech + read charts," Gemini currently stands alone.

Infra advantage is a hidden moat. Google owns the world's largest TPU cluster, giving Gemini better inference cost and availability than most competitors. But Gemini's fate is deeply bound to Google Cloud—risk for enterprises unwilling to depend on a single cloud.

Position Against Competitors

Placing Gemini in the 2026 landscape:

vs Claude: Gemini leads on science reasoning (GPQA 94.1%) and native multimodality; Claude is steadier on coding (SWE-bench 95%) and agentic reliability
vs GPT-5.6: GPT stronger on agentic tasks (BrowseComp, Terminal-Bench); Gemini leads on science and multimodality; general capability split
vs DeepSeek V4: DeepSeek matches on SWE-bench Verified at 1/40 price, MIT self-hostable. Gemini's edge is 1M context and multimodal breadth
vs Open source (Llama 4 / Qwen / Kimi / Mistral): Gemini quality leads, but price is 10–50× open, and the strongest model isn't open

What This Means for Agent Developers

Science research & reasoning → Gemini 3.1 Pro: GPQA 94.1%, ARC-AGI-2 77.1%—strongest for deep science reasoning
Video & speech understanding → Gemini 3.1 Pro: native video/speech input, no transcription step. Unique for video analysis, voice assistants
Long-document handling → Gemini 3.1 Pro: 1M context shines on bulk documents, especially with context caching to cut costs
Google ecosystem integration → If already on Workspace / GCP, Gemini integration is natural
High value → Gemini 3.1 Pro at $2/$12 is 1/6 of Claude, 1/2.5 of GPT, yet parity on science
High-throughput classification/summary → Gemini 3.7 Flash: $0.75/$3.75 (promo), 1M context
Long-horizon coding agents → SWE-bench Verified 80.6% trails Claude ~15pp; 64K output cap constrains complex tasks. For this scenario, prefer Claude Opus 5 or DeepSeek V4 Pro
Local deployment → Flagship locked to Google infra; for self-host use Gemma (non-frontier) or Qwen / DeepSeek

Overall

Gemini's story is "trading infrastructure advantage for pricing and multimodal leadership." Google's TPU hardware and global datacenter network let Gemini deliver 1M context and native multimodality at $2/$12—something Claude and GPT cannot match.

But Gemini has hard limits: 64K output cap, SWE-bench lag, agent stability behind Claude. Its best fit is the science reasoning + multimodal understanding + high value combo—if your scenario needs deep scientific reasoning or video/speech understanding, Gemini is the most cost-effective choice.