Gemini (Google)

From Systems analysis Wiki
Jump to navigation Jump to search

Google Gemini is a family of multimodal large language models (LLMs) developed by Google DeepMind. The Gemini models, first introduced in December 2023, are built on a neural network Transformer architecture with native support for processing and generating data across multiple modalities, including text, images, audio, video, and code.

As of February 2026, the current generation is the Gemini 3.x lineup. Architectural development is focused on integrating scalable inference-time reasoning mechanisms (inference-time scaling) and optimizing models for use within autonomous agentic systems (Agentic AI). The Gemini app has over 750 million monthly active users.

Naming and philosophy

The name "Gemini" (Latin for Twins) symbolizes the merger of two leading Google research groups — Google Brain and DeepMind — to create this project. Jeff Dean, co-technical lead of Google DeepMind, confirmed this in an official blog post (May 2024): "The twins here are the folks in the legacy Brain team and the legacy DeepMind team." The project was originally codenamed "Titan"; Dean proposed the name "Gemini" in April 2023 — the same month as the formal merger of Google Brain and DeepMind. The name also references NASA's Gemini program (1965–1968), whose role in preparing for the Apollo program resonated with the development team.

A key feature and philosophical foundation of Gemini is native multimodality. Unlike many previous models, where multimodal capabilities were layered on top of an existing text-based foundation, Gemini was designed from the ground up for simultaneous understanding, manipulation, and combination of different types of information. The Gemini 1.0 technical report (arXiv:2312.11805) confirms that the model was "trained jointly across image, audio, video, and text data." This enables the model not merely to translate data between modalities, but to form a deeper, holistic understanding of them.

Architecture and key technologies

The capabilities of Gemini models are defined by a number of fundamental architectural decisions. Google does not publish the complete low-level design of all internal Gemini components; however, public sources establish the architecture class: all models from the 1.5 family onward are sparse mixture-of-experts transformer-based models with native multimodal support (confirmed by the Gemini 2.5 Flash model card).

Native multimodal architecture

Gemini's architecture is based on the concept of early fusion. Image pixel patches, video temporal frames, audiograms, and text tokens are projected into a unified latent space. The Gemini 2.5 technical report describes this approach as "Unified Multimodal Token Interleaving." Since all tokens from different modalities are processed within a shared sequence, standard self-attention mechanisms naturally provide cross-modal integration of data from different modalities at every layer. Audio signals are processed by specialized encoders directly from the raw waveform, preserving acoustic characteristics (intonation, timbre, background noise) that are lost when using intermediate Speech-to-Text transcription systems.

For the transformer class, the fundamental operation is the attention mechanism:

Attention(Q,K,V)=softmax(QK⊤dk)V

where Q is the query matrix, K is the key matrix, V is the value matrix, and dk is the key dimensionality.

Sparse Mixture-of-Experts (MoE)

Starting with version 1.5, Gemini models use a Sparse Mixture-of-Experts (MoE) architecture. Gemini 1.0 used a dense transformer; the transition to MoE is explicitly described in the 1.5 technical report: "This is our first release from Gemini 1.5, a new family… which incorporates a novel mixture-of-experts architecture."

In the MoE architecture, standard feed-forward network (FFN) layers are replaced by a set of specialized sub-networks — "experts." For an input token x∈ℝd, the output y is computed as the weighted sum of outputs from k active experts (k≪E, where E is the total number of experts):

y=∑i∈𝒯k(x)gi(x)Ei(x)

where Ei(x) is the nonlinear function of the i-th expert, 𝒯k(x) is the index set of k selected sub-networks, and the routing weight gi(x) is computed by a learned routing function applying a Softmax function over the top-k values.

This approach allows the total parametric capacity of the model to be significantly increased while keeping computational costs (FLOPs) low, since only a subset of parameters is activated for each token. Google has not disclosed the actual parameter count of Gemini models.

Long context and in-context learning

Gemini 1.5 achieved a breakthrough by expanding the context window to 1 million tokens in production mode (with experimental testing up to 10 million tokens). This is an order of magnitude larger than previous models (e.g., GPT-4 Turbo with 128K tokens). Google reported a 99% score on the Needle In A Haystack test at a context length of 1 million tokens. For subsequent generations, long context became one of the key features of the lineup. This large-scale context enables the model to:

  • Analyze entire books, hours-long videos (up to 3 hours), or large codebases within a single query.
  • Perform in-context learning on vast amounts of data provided in the prompt, enabling highly customized responses without the need for fine-tuning.

Thinking models and inference-time compute scaling

Starting with Gemini 2.5, Google designates thinking as a separate operating mode. Official documentation defines it as an internal computational process that improves multi-step planning and reasoning. Version 2.5 models (described as "thinking models") are capable of internally generating and evaluating intermediate reasoning steps before producing a final answer. This significantly improves accuracy on complex logical and mathematical tasks.

It is important to distinguish between two mechanisms:

  • Built-in Thinking: The base mode for 2.5 and 3-series models, generating a hidden chain-of-thought (CoT). The API can return thought summaries — brief summaries of internal reasoning rather than the full stream of raw "thoughts." Starting with the 3.1 Pro model, the thinking budget is controlled by the thinking_level parameter with values from Low to Max.
  • Deep Think: A separate experimental enhanced reasoning mode that uses parallel hypothesis generation and requires significantly greater computational resources. It was announced at Google I/O on May 20, 2025, and made available to AI Ultra subscribers on August 1, 2025. Deep Think should not be conflated with the base thinking mechanism.

Agentic capabilities

Starting with version 2.0, Gemini can interact with the external world: invoke tools, perform Google Search, execute code, and control UI elements. Google explicitly positioned Gemini 2.0 as a model for the "new agentic era" with native tool use support.

As of February 2026, the Gemini API includes a formally established agentic capabilities layer with support for tools: Google Search, Google Maps, Code Execution, URL Context, Computer Use, File Search, and Live API for bidirectional real-time interaction.

Evolution of Gemini models

The Gemini family evolves at an exceptionally rapid pace: four major model generations were released between December 2023 and February 2026.

Gemini 1.0 (December 2023)

The first generation, establishing the foundation of native multimodality. Publicly introduced on December 6, 2023.

  • Versions: Ultra (flagship for the most complex tasks), Pro (general-purpose model), and Nano (compact for mobile devices; subdivided into Nano-1 with 1.8B parameters and Nano-2 with 3.25B).
  • Context window: 32,768 tokens for all versions.
  • Achievements: Gemini 1.0 Ultra became the first model to reach and surpass human-expert performance on the MMLU benchmark with a score of 90.04% (using the CoT@32 technique — chain-of-thought with 32 sampled chains and majority voting; under standard 5-shot prompting, the score was approximately 83.7%). It achieved SOTA results on 30 out of 32 academic benchmarks.
  • Deprecation: Gemini 1.0 Pro was deprecated on February 18, 2025.

Gemini 1.5 (February–May 2024)

A breakthrough in context length and efficiency.

  • Architecture: Transition from dense transformer to Mixture-of-Experts (MoE).
  • Context window: Up to 1 million tokens in production (2 million via waitlist for 1.5 Pro, announced at Google I/O in May 2024).
  • Versions: 1.5 Pro (announced February 2024; quality on par with 1.0 Ultra at significantly lower cost) and 1.5 Flash (lightweight and fast version, added May 2024).
  • Deprecation: All Gemini 1.5 models (Pro, Flash, Flash-8B) were shut down on September 29, 2025.

Gemini 2.0 (December 2024–February 2025)

The transition to the "agentic era."

  • Timeline: December 11, 2024 — announcement of 2.0 Flash Experimental (multimodal input, text output); February 5, 2025 — broad availability (GA) of 2.0 Flash, release of 2.0 Pro Experimental and 2.0 Flash-Lite.
  • Key innovations: Built-in agentic capabilities (tool use), native image and audio generation (initially in limited mode for early-access partners), focus on agentic scenarios.
  • Context window: Up to 2M tokens (2.0 Pro); up to 1M tokens (2.0 Flash-Lite).
  • Deprecation: 2.0 Flash and Flash-Lite models are scheduled for shutdown on June 1, 2026.

Gemini 2.5 (March–June 2025)

The first "thinking model" with configurable reasoning budgets.

  • Timeline: March 25, 2025 — announcement of 2.5 Pro Experimental; April 17 — 2.5 Flash (the first fully hybrid reasoning model with toggleable thinking); May 20 (Google I/O) — updates to 2.5 Pro and Flash, Deep Think announcement; June 17, 2025 — simultaneous GA for 2.5 Pro and 2.5 Flash; same day — 2.5 Flash-Lite preview (GA July 22). August 1 — Deep Think made available to AI Ultra subscribers.
  • Key innovations: Built-in "thinking" mechanism with configurable budgets; Deep Think as a separate enhanced mode. SOTA results on complex mathematical, logical, and software benchmarks (AIME 2025 — 86.7%, GPQA Diamond — 84.0%, Humanity's Last Exam — 18.8% without tools).
  • Context window: 1 million tokens input, up to 64,000 tokens output. The promised expansion to 2M tokens for 2.5 Pro was never confirmed as delivered during the model's lifecycle.
  • Specialized variants: Gemini 2.5 Flash Image (codename "Nano Banana," anonymously appeared on the Arena August 12, officially released August 26, 2025 — went viral for photorealistic "3D figurine" images, attracting 10 million new users); Computer Use Preview (October 7, 2025, based on 2.5 Pro); Text-to-Speech models (2.5 Flash TTS, 2.5 Pro TTS).
  • Technical report: The combined Gemini 2.X report was published on arXiv on July 7, 2025 (arXiv:2507.06261), listing over 3,300 authors and covering models 2.5 Pro, 2.5 Flash, 2.0 Flash, and 2.0 Flash-Lite.

Gemini 3.x (November 2025–February 2026)

The third generation marked the transition from basic generation to long-running agentic workflows and interdisciplinary scientific problem-solving.

  • Gemini 3 Pro (November 18, 2025): Announced by Alphabet CEO Sundar Pichai and DeepMind CEO Demis Hassabis as "Google's most intelligent model." The first Gemini model deployed to Google Search on launch day. Became the first model to break the 1,500 Elo barrier on LMArena (1,501 at launch). Results: GPQA Diamond — 91.9%; SWE-bench Verified — 76.2%; Humanity's Last Exam — 37.5% (without tools); SimpleQA — 72.1%.
  • Gemini 3 Flash (December 17, 2025): Became the default model in the Gemini app. At a price of $0.50/1M input tokens, it outperformed 3 Pro on SWE-bench Verified (78%) while using 30% fewer tokens on reasoning tasks. GPQA Diamond — 90.4%; HLE — 33.7%.
  • Gemini 3.1 Pro (February 19, 2026): The flagship model as of the publication date. The first incremental ".1" release (previous generations used .5 intervals). Key result — ARC-AGI-2: 77.1% (more than double 3 Pro's 31.1%). AIME 2025 — 91.2%; GPQA Diamond — 94.3%; SWE-bench Verified — 80.6%. Introduced a new MEDIUM thinking level via the thinking_level parameter. Dedicated endpoint gemini-3.1-pro-preview-customtools for bash terminal and custom function use. Resolved output truncation issues on long generations. Channels: Gemini App, Vertex AI, AI Studio, Gemini API, NotebookLM.
  • Gemini 3 Deep Think (updated February 12, 2026): A major update to the specialized "thinking" mode. Expanded beyond mathematics and programming: gold-medal-level results on the 2025 International Physics Olympiad (IPhO) and International Chemistry Olympiad (IChO); ARC-AGI-2 — 84.6%; Humanity's Last Exam — 48.4%; CMT-Benchmark (condensed matter theoretical physics) — 50.5%; Codeforces Elo — 3,455. The Deep Think–based research agent Aletheia autonomously solved several open problems from the Erdős problem collection (including the Erdős-1051 conjecture).

Summary table of Gemini generations

Evolution of key characteristics of Gemini models
Generation Release year Key versions Max context window Key architectural innovations and improvements
Gemini 1.0 2023–2024 Ultra, Pro, Pro Vision, Nano-1 (1.8B), Nano-2 (3.25B) 32,768 tokens First Gemini generation, designed as a natively multimodal model family rather than separate text and vision systems. Ultra targeted maximum capability, Pro balanced capability and serving cost, while Nano introduced distilled and quantized on-device models. Gemini 1.0 Ultra became the first model reported by Google to exceed human-expert performance on MMLU under its CoT@32 evaluation (90.04%).
Gemini 1.5 2024 1.5 Pro, 1.5 Flash, 1.5 Flash-8B 2,000,000 tokens (1.5 Pro); 1,000,000 for Flash and Flash-8B Introduced a Mixture-of-Experts (MoE) architecture and a major expansion of multimodal long-context processing. 1.5 Pro initially launched with 128K/1M context and later made a 2M-token window generally available. 1.5 Flash introduced a smaller model distilled from Pro for high-throughput workloads, followed by the still smaller Flash-8B. Strong near-perfect retrieval across very long text, audio and video contexts.
Gemini 2.0 2024–2025 2.0 Flash, 2.0 Flash Thinking Experimental, 2.0 Pro Experimental, 2.0 Flash-Lite; specialized Flash Live / image-generation variants 2,000,000 tokens (2.0 Pro Experimental); 1,000,000 for Flash family Beginning of Google's “agentic era.” Added native tool use, including Search, code execution and function calling; improved spatial and multimodal reasoning; native image and steerable multilingual audio output in experimental configurations; and the Multimodal Live API for real-time audio/video interaction. Flash Thinking Experimental introduced explicit multi-step reasoning, while 2.0 Pro Experimental targeted complex reasoning and coding.
Gemini 2.5 2025 2.5 Pro, 2.5 Flash, 2.5 Flash-Lite, 2.5 Deep Think, 2.5 Flash Image (Nano Banana), 2.5 Computer Use, Flash Live / native-audio and Pro/Flash TTS variants 1,048,576 input / 65,536 output tokens for Pro, Flash and Flash-Lite First Gemini generation designed throughout as a family of hybrid reasoning (“thinking”) models. Introduced controllable thinking budgets, allowing developers to trade reasoning depth for latency and cost. Deep Think added parallel reasoning over multiple candidate hypotheses. Gemini 2.5 Flash Image, better known as Nano Banana, brought native conversational image generation and editing, while the specialized Computer Use model enabled agents to operate graphical user interfaces.
Gemini 3.0 2025 3 Pro Preview, 3 Flash Preview, 3 Deep Think, 3 Pro Image (Nano Banana Pro) 1,048,576 input / 65,536 output tokens for Pro and Flash Major expansion of multimodal reasoning, coding and agentic execution. Introduced thinking_level as a higher-level reasoning-control mechanism, alongside thought signatures for preserving reasoning state across tool calls. Gemini 3 strengthened visual and spatial reasoning, multimodal tool use and long-horizon agentic workflows. Deep Think used advanced parallel reasoning to explore multiple hypotheses simultaneously.
Gemini 3.1 2026 3.1 Pro, 3.1 Deep Think, 3.1 Flash-Lite, 3.1 Flash Image (Nano Banana 2), 3.1 Flash-Lite Image (Nano Banana 2 Lite), 3.1 Flash Live, 3.1 Flash TTS 1,048,576 input / 65,536 output tokens for Pro and Flash-Lite Refinement of the Gemini 3 architecture with substantially stronger reasoning, token efficiency, factual consistency, software engineering and reliable multi-step tool use. 3.1 Pro replaced the original 3 Pro Preview. 3.1 Deep Think, built on 3.1 Pro, specialized in advanced mathematics, physics, chemistry, research and engineering and reached 84.6% on ARC-AGI-2 in Google's reported evaluation. The 3.1 family also expanded native visual and real-time audio branches.
Gemini 3.5 2026 3.5 Flash, 3.5 Flash-Lite; specialized 3.5 Transcribe, 3.5 Transcribe Live and 3.5 Live Translate 1,048,576 input / 65,536 output tokens for Flash and Flash-Lite Shifted much of the frontier Gemini 3 capability into the lower-latency Flash tier. Improved sustained performance on coding, agentic execution and long-horizon tasks, with automatic preservation of intermediate reasoning across multi-turn interactions. Flash-Lite targeted low-cost subagents and high-volume automation. Gemini 3.5 Flash also added integrated Computer Use support.
Gemini 3.6 2026 3.6 Flash 1,048,576 input / 65,536 output tokens Production Flash update focused on greater token efficiency, reduced verbosity, lower cost, stronger code generation, spatial reasoning and more reliable planning in fast agentic loops. Continued integrated tool use and Computer Use support.
Gemini 3.7 2026 3.7 Flash 1,048,576 input / 65,536 output tokens Further improvement in real-world software engineering, web development and autonomous agent workflows, with more reliable multi-step execution and stronger design-to-code fidelity. Added advanced agentic video understanding, allowing models to navigate video timelines dynamically and request relevant frames, transcripts or audio rather than processing all media uniformly.
Gemini 3.8 2026 3.8 Flash 1,048,576 input / 65,536 output tokens Latest stable Flash generation as of September 2026, engineered for long-horizon software engineering, autonomous agents and complex enterprise workflows. Supports configurable low / medium / high thinking levels, Computer Use, Search and Maps grounding, code execution, File Search, URL context, function calling and other built-in tools while retaining Flash-class speed and cost efficiency.

Key results and benchmarks

As classical benchmarks (such as MMLU) have become saturated, evaluation of Gemini model performance has shifted toward abstract reasoning, scientific modeling, and autonomous software engineering tasks. Results are reported from official Google data (self-reported); comparisons are valid only when inference mode, tool use presence, sampling method (single-attempt vs. majority voting), and specific model-id all match.

Gemini model results on key benchmarks (data as of February 2026)
Benchmark Task description Gemini 2.5 Pro (Jun 2025) Gemini 3 Pro (Nov 2025) Gemini 3.1 Pro (Feb 2026) Gemini 3 Deep Think (Feb 2026)
MMLU Multitask language understanding — — — —
GPQA Diamond PhD-level science questions 84.0% 91.9% 94.3% N/A
Humanity's Last Exam Frontier domain knowledge 18.8% 37.5% 44.4% 48.4%
ARC-AGI-2 Abstract logical puzzles 4.9% 31.1% 77.1% 84.6%
SWE-bench Verified Autonomous GitHub issue resolution 63.8%* 76.2% 80.6% N/A
AIME 2025 Olympiad-level math problems 86.7% — 91.2% —
Codeforces (Elo) Competitive programming rating — — 2,887 3,455

* The 2.5 Pro result on SWE-bench was obtained with a custom agent setup.

LMArena rankings (late February 2026 snapshot)

LMArena (formerly Chatbot Arena) is an independent platform for blind pairwise voting. Rankings are dynamically recalculated; values at a model's launch date may differ from current ones.

Overall (snapshot: February 24, 2026)
Model Rating Rank Votes Note
Gemini 3.1 Pro Preview 1,500 ± 9 #3 4,060 Preliminary
Gemini 3 Pro 1,486 ± 4 #5 37,854
Gemini 3 Flash 1,473 ± 5 #7 28,847
Gemini 2.5 Pro 1,464 ± 3 #9 97,296
Gemini 2.5 Flash 1,411 ± 3 #64 96,163

At launch on November 18, 2025, Gemini 3 Pro reached a rating of 1,501 Elo, becoming the first model to break the 1,500 barrier on LMArena.

Gemini

No. Date Model Note Source Developer arXiv / Technical report Model Card
1 2023-12-06 Gemini 1.0 First Gemini generation; covered Ultra, Pro and Nano variants. Designed from the ground up as a natively multimodal model family capable of processing text, images, audio, video and code. Google — arXiv:2312.11805
Google DeepMind PDF 		HTML: —
PDF: Google DeepMind PDF
2 2024-02-15 Gemini 1.5 Long-context Gemini generation based on a Mixture-of-Experts architecture. Included Gemini 1.5 Pro and Gemini 1.5 Flash; Gemini 1.5 Pro later received a generally available 2-million-token context window. Google Vertex AI model versions arXiv:2403.05530
Google DeepMind PDF 		HTML: —
PDF: Google DeepMind PDF
3 2024-08-27 / 2024-10-03 Gemini 1.5 Flash-8B Smaller and faster Flash variant aimed at high-volume and lower-complexity workloads. An experimental version appeared in August 2024; the production-ready GA version followed on 2024-10-03. Google Developers Blog Gemini API docs arXiv:2403.05530
Google DeepMind PDF 		Gemini 1.5 technical report
4 2024-12-11 Gemini 2.0 Flash First Gemini 2.0 model and the beginning of Google's “agentic era”; introduced stronger agentic capabilities, native tool use, multimodal interaction and experimental multimodal output. GA followed on 2025-02-05. Google Gemini API docs — HTML: Google DeepMind
PDF: Model Card PDF
5 2024-12-19 / 2025-01-21 Gemini 2.0 Flash Thinking Experimental Experimental test-time-compute reasoning variant of Gemini 2.0 Flash. Google released the original public preview in December 2024 and an updated gemini-2.0-flash-thinking-exp-01-21 snapshot in January 2025. Gemini API release notes Gemini API deprecations — —
6 2025-02-05 Gemini 2.0 Pro Experimental Experimental Gemini 2.0 Pro model for complex reasoning, coding and long-context workloads. The public preview endpoint was gemini-2.0-pro-exp-02-05; unlike Flash and Flash-Lite, Gemini 2.0 Pro never received a stable GA release. Google Gemini API release notes — —
7 2025-02-05 / 2025-02-25 Gemini 2.0 Flash-Lite Cost-efficient Gemini 2.0 model optimized for speed, scale and low latency. Preview released on 2025-02-05 and GA followed on 2025-02-25. Google Developers Blog Gemini API docs — HTML: Google DeepMind
PDF: Model Card PDF
8 2025-03-12 / 2025-05-07 Gemini 2.0 Flash Preview Image Generation Experimental native image-generation and conversational image-editing variant of Gemini 2.0 Flash. First demonstrated in March 2025; the dedicated gemini-2.0-flash-preview-image-generation endpoint followed in May. Gemini API release notes Gemini API docs — Parent model: Gemini 2.0 Flash Model Card PDF
9 2025-04-09 Gemini 2.0 Flash Live Public-preview Live API model for low-latency bidirectional voice and video interaction. The endpoint gemini-2.0-flash-live-001 supported streaming audio/video input and audio/text output. Gemini API release notes Gemini API deprecations — Parent model: Gemini 2.0 Flash Model Card PDF
10 2025-03-25 / 2025-06-17 Gemini 2.5 Pro First Gemini 2.5 “thinking” Pro model. Introduced a reasoning-first architecture with thinking enabled by default; the stable GA version followed in June 2025. Google Gemini API docs arXiv:2507.06261
Google DeepMind technical report PDF 		HTML: Google DeepMind
PDF: Model Card PDF
11 2025-04-17 / 2025-06-17 Gemini 2.5 Flash Hybrid-reasoning Flash model optimized for price-performance, adaptive thinking and high-volume workloads. Preview launched in April 2025 and the stable release followed in June. Google Developers Blog Gemini API docs arXiv:2507.06261
Google DeepMind technical report PDF 		HTML: Google DeepMind
PDF: Model Card PDF
12 2025-05-20 / 2025-12-10 Gemini 2.5 Flash TTS / Gemini 2.5 Pro TTS Dedicated text-to-speech models. Flash TTS emphasized low latency and cost efficiency, while Pro TTS targeted higher-fidelity professional narration. Both support controllable speech generation and single- or multi-speaker output. Gemini API release notes Gemini 2.5 Flash TTS
Gemini 2.5 Pro TTS 		— —
13 2025-05-20 / 2025-12-12 Gemini 2.5 Flash Native Audio / Flash Live Native-audio Gemini 2.5 Flash branch for low-latency bidirectional voice and video agents. Initial native-audio dialogue models were released in May 2025; a substantially updated gemini-2.5-flash-native-audio-preview-12-2025 version followed in December. Gemini API release notes Gemini API docs arXiv:2507.06261
Google DeepMind technical report PDF 		—
14 2025-06-17 / 2025-07-22 Gemini 2.5 Flash-Lite Fastest and most budget-friendly general-purpose Gemini 2.5 model, designed for high-throughput and cost-sensitive workloads. Preview launched in June and stable GA followed in July. Google Developers Blog
Google Developers Blog 		Gemini API docs arXiv:2507.06261
Google DeepMind technical report PDF 		HTML: Google DeepMind
PDF: Model Card PDF
15 2025-08-01 Gemini 2.5 Deep Think Specialized enhanced-reasoning mode/model for especially difficult mathematics, science and coding problems; used parallel reasoning to explore multiple candidate solutions. Google — arXiv:2507.06261
Google DeepMind technical report PDF 		PDF: Model Card PDF
16 2025-08-26 / 2025-10-02 Gemini 2.5 Flash Image / Nano Banana Native conversational image-generation and editing model built on Gemini 2.5 Flash. Initially released in preview in August 2025 under the “Nano Banana” nickname; GA followed in October. Google Developers Blog Gemini API docs arXiv:2507.06261
Google DeepMind technical report PDF 		Google DeepMind
17 2025-10-07 Gemini 2.5 Computer Use Specialized Gemini 2.5 model for browser/UI-control agents using screenshots, visual reasoning and generated interface actions. Google Gemini API docs arXiv:2507.06261
Google DeepMind technical report PDF 		PDF: Model Card PDF
18 2025-11-18 Gemini 3 Pro First Gemini 3 model; major upgrade in reasoning, multimodal understanding, visual/spatial reasoning, coding and agentic workflows. Initially released as Gemini 3 Pro Preview. Google Gemini API docs — Google DeepMind
19 2025-11-20 / 2026-05-28 Gemini 3 Pro Image / Nano Banana Pro Advanced native image-generation and editing model based on Gemini 3 Pro reasoning. Preview launched in November 2025; the stable gemini-3-pro-image endpoint became GA in May 2026. Google Gemini API docs — Google DeepMind
20 2025-12-04 / 2026-02-12 Gemini 3 Deep Think Advanced parallel-reasoning mode for complex mathematics, science, logic, research and engineering. Originally launched for Gemini 3 Pro in December 2025 and substantially upgraded in February 2026. The current DeepMind implementation is presented as Gemini 3.1 Deep Think and is built on Gemini 3.1 Pro. Google
Google 		Google DeepMind — —
21 2025-12-17 Gemini 3 Flash Gemini 3 Flash model optimized for frontier-class reasoning, coding and agentic performance at lower latency and cost than Pro-class models. Google Gemini API docs — Google DeepMind
22 2026-02-19 Gemini 3.1 Pro Upgraded Gemini 3.1 Pro model for complex reasoning, coding, multimodal tasks, synthesis and agentic workflows; successor to the original Gemini 3 Pro Preview. Google Gemini API docs — HTML: Google DeepMind
PDF: Model Card PDF
23 2026-02-26 / 2026-05-28 Gemini 3.1 Flash Image / Nano Banana 2 Gemini 3.1 native image-generation and editing model. Preview launched in February 2026 and the stable gemini-3.1-flash-image endpoint became GA in May. Supports multimodal conversational editing and later added video-to-image generation. Google
Gemini API release notes 		Gemini API docs — HTML: Google DeepMind
PDF: Model Card PDF
24 2026-03-03 / 2026-05-07 Gemini 3.1 Flash-Lite Lightweight Gemini 3.1 model focused on high-frequency, cost-efficient and low-latency workloads. Preview launched in March; stable GA followed in May. Google Gemini API docs — HTML: Google DeepMind
PDF: Model Card PDF
25 2026-03-11 / 2026-03-26 Gemini 3.1 Flash Live Low-latency audio-to-audio model for real-time dialogue and voice-first multimodal agents. The API preview endpoint appeared in March 2026 and was publicly announced on March 26. Google Gemini API docs — HTML: Google DeepMind
PDF: Model Card PDF
26 2026-04-13 / 2026-04-15 Gemini 3.1 Flash TTS Low-latency expressive text-to-speech model with steerable prompting and granular audio tags controlling vocal style, pacing and delivery. Google Gemini API docs — HTML: Google DeepMind
PDF: Model Card PDF
27 2026-05-19 Gemini 3.5 Flash First Gemini 3.5 model; designed for sustained frontier performance on coding, autonomous-agent and long-horizon tasks while retaining Flash-class speed. Built on the Gemini 3 Flash reasoning foundation. Google Gemini API docs — HTML: Google DeepMind
PDF: Model Card PDF
28 2026-05-19 / 2026-06-30 / 2026-08-27 Gemini Omni Flash / Gemini Omni 1.1 Flash Separate generative Gemini branch combining Gemini multimodal reasoning with native video generation and conversational video editing. Announced at Google I/O 2026; gemini-omni-flash-preview entered API public preview on June 30. The stable Gemini Omni 1.1 Flash (gemini-omni-1.1-flash) became GA on August 27 with video extension, keyframe interpolation and output up to 4K. Google I/O 2026
Google 		Gemini API docs
Gemini Omni guide 		— HTML: Google DeepMind
29 2026-05-28 Gemini 3 Pro Image (GA) Stable production release of Nano Banana Pro, replacing the original gemini-3-pro-image-preview endpoint. Gemini API release notes Gemini API docs — Google DeepMind
30 2026-06-09 / 2026-08-26 Gemini 3.5 Audio — Live Translate / Transcribe / Transcribe Live Specialized Gemini 3.5 audio family. Live Translate provides low-latency speech-to-speech translation. Transcribe provides high-accuracy non-streaming speech-to-text, while Transcribe Live provides bidirectional streaming transcription. The Transcribe models became GA on 2026-08-26. Google
Gemini API release notes 		Gemini 3.5 Transcribe
Gemini API models 		— HTML: Google DeepMind
31 2026-06-30 Gemini 3.1 Flash-Lite Image / Nano Banana 2 Lite Ultra-low-latency, cost-efficient native image-generation and editing model designed for high-volume interactive workloads. Google Gemini API docs — HTML: Google DeepMind
PDF: Model Card PDF
32 2026-07-21 Gemini 3.5 Flash-Lite Fastest and most cost-effective 3.5-class model. Optimized for high-volume, latency-sensitive tasks, document processing, translation, classification and subagent workloads; based on Gemini 3.1 Flash-Lite. Google Gemini API docs — HTML: Google DeepMind
PDF: Model Card PDF
33 2026-07-21 Gemini 3.5 Flash Cyber Cybersecurity-specialized model built on Gemini 3.5 Flash and fine-tuned for defensive vulnerability discovery, validation and automated patching. Access is more restricted than the general Gemini API models. Google DeepMind
Google 		— — —
34 2026-07-21 Gemini 3.6 Flash Production Flash update with stronger coding, knowledge-work, multimodal and spatial-reasoning performance and improved token efficiency relative to Gemini 3.5 Flash. Supports Computer Use and the main Gemini built-in tools. Google Gemini API docs — HTML: Google DeepMind
35 2026-08-13 Gemini 3.7 Flash Successor to Gemini 3.6 Flash with algorithmic improvements to the core reasoning foundation, stronger software engineering, web development and agentic workflows, plus configurable low/medium/high thinking levels. Later received agentic video understanding. Google Gemini API docs — HTML: Google DeepMind
36 2026-08-26 Gemini 3.5 Transcribe Dedicated non-streaming speech-to-text endpoint (gemini-3.5-transcribe) with language detection across more than 85 languages, speaker diarization, word-level timestamps and custom vocabulary biasing. Gemini API release notes Gemini API docs — Gemini 3.5 Audio Model Card
37 2026-08-26 Gemini 3.5 Transcribe Live Streaming speech-to-text counterpart to Gemini 3.5 Transcribe, using the Live API and WebSockets for low-latency interim and finalized transcription events. Gemini API release notes Gemini API docs — Gemini 3.5 Audio Model Card
38 2026-08-27 Gemini Omni 1.1 Flash Stable GA release of Gemini Omni Flash. Added video extension, first/last-frame interpolation and resolution control for 360p, 720p, 1080p and upscaled 4K video output. Gemini API release notes Gemini API docs — Google DeepMind
39 2026-09-02 Gemini 3.8 Flash Latest stable numbered Gemini model as of 2026-09-06. Successor to Gemini 3.7 Flash, with further improvements in software engineering, long-horizon autonomous agents, knowledge workflows and critical multi-step reasoning. Supports a 1,048,576-token input context, 65,536-token maximum output and configurable low/medium/high thinking levels. Google Gemini API docs — HTML: Google DeepMind
40 2026-09-02 Gemini 3.8 Flash Cyber Cybersecurity-specialized Gemini 3.8 variant for defensive vulnerability discovery and automated patching. Google describes it as its most capable cybersecurity model; access is restricted to trusted defenders through the Fairwind Program rather than being a normal open Gemini API endpoint. Google Google DeepMind / Fairwind — —

Scope note: The table lists named models and major publicly released variants in the main Gemini generative family, including its image, audio, Live, Computer Use, Deep Think, Cyber and Omni branches. Dated API snapshots such as -001, -002, -exp-0827 and similar revisions are not treated as separate models unless they introduced a distinct named capability. Gemini Embedding and Gemini Robotics constitute separate specialist model lines and are not included in this table.

Specialized and agentic systems

The Gemini ecosystem has been extended with models and platforms capable of performing multi-step actions in digital and physical environments.

Autonomous agents

  • Jules — an autonomous coding agent operating asynchronously in secure cloud virtual machines. It creates branches and pull requests on GitHub. Entered public beta at Google I/O on May 20, 2025 (over 140,000 code improvements during the beta period); GA on August 6, 2025. By late 2025, it had become one of the largest contributors to Google's internal repositories.
  • Project Mariner — a research prototype of a browser-based agent for multi-step web tasks. Migrated to cloud VMs supporting up to 10 parallel tasks and a "Teach & Repeat" feature. Achieved 83.5% on the WebVoyager benchmark. Computer Use capabilities were ported to the Gemini API.
  • Google Antigravity — an integrated development environment (IDE) for managing AI agents, introduced in November 2025. Agents autonomously modify code, interact with the terminal and a built-in browser, returning verifiable artifacts (e.g., code diffs) for developer approval.
  • Aletheia agent — a specialized mathematical research agent built on Gemini 3 Deep Think. Equipped with a natural-language verification module and web-search tools for literature review. In early 2026, it autonomously solved several open mathematical problems from the Erdős collection and co-authored scientific publications.

Consumer AI agents

  • Phone Automations — integration of an autonomous agent at the Android OS level (beta for Pixel 10 and Samsung Galaxy S26). Operates within a secure sandbox, capable of navigating third-party applications based on visual GUI analysis.
  • Gemini in Chrome (Auto Browse) — a browser agent for automating multi-step web tasks, available to all Chrome users since September 2025 (updated to Gemini 3 in January 2026).

Computer Use

Gemini 2.5 Computer Use models are optimized for controlling graphical user interfaces (GUIs). The system takes screenshots and action history as input, generating (x,y) coordinates for programmatic cursor simulation and keyboard input commands.

Gemini Robotics

Vision-Language-Action (VLA) and Embodied Reasoning (ER) models introduced in March 2025. These architectures process spatiotemporal information and predict 3D trajectories of robotic manipulators as a native output modality (arXiv:2503.20020).

Specialized generative models (early 2026)

  • Nano Banana 2 (Gemini 3.1 Flash Image) — released February 26, 2026; a visual model combining Flash-architecture speed with Pro-level quality. Provides strict character consistency across different scenes, native text-in-image generation (typography), and integration of SynthID cryptographic watermarks with C2PA metadata.
  • Lyria 3 — a music model integrated into the Gemini app on February 18, 2026. Generates 30-second musical compositions (including vocals and instrumentals) from text prompts, uploaded photos, or videos.
  • Veo 3.1 — a video generation model. Supports creating clips using up to three reference images ("Ingredients to Video"), transition generation between specified first and last frames, native vertical video rendering (9:16), and upscaling to 4K resolution.
  • Med-Gemini — a domain-specific model for medical tasks (arXiv:2404.18416, arXiv:2405.03162).

Applications and ecosystem

Google deeply integrates Gemini into its consumer and developer products.

Consumer products

  • Gemini app: A chatbot (formerly Bard, renamed February 8, 2024) using Gemini family models as a universal AI assistant. As of February 2026, it has over 750 million active users. Current rollout includes the 3.1 Pro model. Subscriptions: Google AI Pro ($19.99/month, replaced Google One AI Premium) and Google AI Ultra ($249.99/month, with access to Deep Think, Veo 3, and priority features).
  • Google Workspace: Gemini integration in Gmail, Docs, Sheets, and Meet for writing assistance, data analysis, and content generation (rebranded from Duet AI).
  • Google Search: The AI Overviews feature generates summary answers to complex queries using a specialized Gemini model. AI Mode, launched at Google I/O 2025, provides deep search with agentic capabilities (booking, shopping).
  • Android and Pixel: Gemini Nano (v3 on Pixel 10 with Tensor G5 chip, August 2025) runs locally on smartphones, providing smart replies, summarization, scam call detection, and accessibility features while preserving data privacy. ML Kit GenAI APIs for developers support on-device summarization, proofreading, and speech recognition.
  • NotebookLM: Evolved from a note-taking tool into a full creative platform. Joined Google Workspace in March 2025. Supports interactive Audio Overviews, Video Overviews, Mind Maps, slides, and infographics. Updated to Gemini 3 in December 2025; full 1M-token context window for chat available from February 2026.
  • Gemini Live: Camera and screen-sharing features from Project Astra became free for all Android and iOS users.

Developer platforms

  • Google AI Studio and Gemini API: Primary interfaces for accessing Gemini models via API. As of February 2026, they support capability blocks: Thinking, Thought signatures, Long context, Tools and agents (Google Search, Maps, Code Execution, URL Context, Computer Use, File Search, Deep Research, Live API).
  • Vertex AI: Enterprise platform with enhanced security and management capabilities.
  • Google Gen AI SDK: Reached GA for Python, JavaScript/TypeScript, Go, and Java by May 2025, providing unified access to the Gemini Developer API and Vertex AI. Supports Model Context Protocol (MCP).
  • Gemini CLI: A command-line tool for AI coding in the terminal (launched June 2025).
  • Interactions API: A unified interface for models and agents (beta since December 2025).

API lifecycle and version management

Gemini models in the API are categorized as stable, preview, latest, and experimental. A specific model_id and a model family are not the same thing; for production scenarios, binding to a specific version and its support timeline is critical. The API documentation maintains a deprecation registry with shutdown dates.

To support long-running autonomous tasks, the following were introduced: Session Resumption (server-side session state storage for up to 24 hours) and Context Compression (a sliding-window mechanism for automatic context compression when limits are exceeded).

In December 2025, Google reduced free-tier API quotas by approximately 92% (without prior notice), causing a sharp reaction from the developer community. Meanwhile, Gemini serving unit costs fell by 78% over 2025 through model optimizations.

Limitations and open problems

  • Hallucinations and confabulations: Models retain a tendency to generate factually incorrect information, especially when grounding features (Search Grounding) are disabled. Gemini 3.1 Pro reduced hallucination rates on the SimpleQA benchmark compared to previous versions, but the problem remains systemic across all LLMs.
  • Subconscious plagiarism: Experiments with the Aletheia agent revealed a problem where the model reproduces non-trivial proofs from its training set, presenting them as autonomous discoveries, complicating the validation of novelty in AI research.
  • Long-context degradation: When processing contexts of 1 million tokens or more, models are subject to the "Lost in the Middle" effect — reduced accuracy in retrieving facts located in the middle of a document.
  • High computational costs: Inference with maximum Deep Think settings requires significantly more time and resources (TPUs), limiting application in synchronous real-time scenarios.
  • Over-refusals: Due to strict alignment algorithms, reasoning models tend to reject legitimate requests by falsely classifying them as potentially harmful (especially in the context of code analysis and information security). Model cards also note issues with "preachy" refusal tones.
  • Reasoning limitations: Model cards for the 2.5 and 3 series list limitations in causal understanding, complex logical deduction, and counterfactual reasoning, as well as incomplete predictability in adhering to thinking budgets.

Ethical aspects and safety

The deployment of Gemini models is accompanied by a multi-layered safety system.

General frameworks

Secure AI Framework (SAIF) is Google's general approach to AI system safety (announced June 2023), forming the development context but not being a Gemini-specific standard. Frontier Safety Framework v3 (September 2025) covers CBRN, cybersecurity, ML R&D, harmful manipulation, and an exploratory approach to misalignment risks.

Gemini-specific measures

  • Model cards are the primary sources of information on limitations and safety for specific models. They contain sections on Intended Usage and Limitations, Ethics and Content Safety, and Frontier Safety. The Gemini 3 Pro model card confirmed that the model did not reach any Critical Capability Levels (CCLs) across CBRN and cybersecurity domains.
  • Bias and toxicity testing: Analysis and mitigation of bias in training data and content generation.
  • Red teaming: Attack simulation to identify vulnerabilities and undesirable behavior. Independent misalignment testing found "some upticks in situational awareness" but no critical risks.

Safety probes

To prevent the generation of harmful content, hidden activation classification is used. To address signal loss in long contexts, the MultiMax architecture is employed: the probe extracts the maximum value across all layers H for each token j in sequence ni:

fMultiMax(Si)=∑h=1Hmaxj∈[ni][vh⊤yi,j]

Probes are combined with base models into cascade classifiers, improving filtering accuracy at low computational cost (arXiv:2601.11516).

Cryptographic watermarking (SynthID)

Audio data generated through the Live API and images (from Nano Banana / Flash Image models) are watermarked using the SynthID algorithm. An invisible watermark is embedded at the pixel or audio spectrum level, enabling machine detection of generated content. The Nano Banana 2 model (February 2026) integrates SynthID with C2PA metadata.

Thinking and the question of transparency

Models with thinking mode (2.5/3 series) can return thought summaries — brief summaries of internal reasoning rather than the full stream of intermediate tokens. This provides a certain level of transparency, but has been criticized for hiding actual "raw" reasoning chains behind simplified summaries.

Regulatory aspects

Under the EU AI Act, Google signed the EU AI Code of Practice (published July 10, 2025) alongside OpenAI and Anthropic. Gemini is classified as a general-purpose AI (GPAI) model with systemic risk, which entails additional safety obligations (effective August 2, 2025).

Competitive landscape

The November–December 2025 period became the most compressed competitive cycle in AI history: Gemini 3 Pro (November 18), Claude Opus 4.5 by Anthropic (November 24), and GPT-5.2 by OpenAI (December 11) were all released within 24 days. As of February 2026, no single model dominates all categories: Gemini 3 Pro leads LMArena in text, vision, search, and multilingual; GPT-5.2 leads in pure math (100% AIME 2025 without tools) and SWE-bench Pro; Claude Opus 4.5 competes on SWE-bench Verified. In terms of API pricing, Gemini is approximately 42% cheaper than GPT-5 for comparable calls.

Business metrics

According to Alphabet's Q4 2025 earnings report (published February 4, 2026): Google Cloud revenue was $17.7 billion for the quarter (+48% year-over-year); operating margin was 29.9%; Cloud backlog reached $240 billion (doubled year-over-year). Over 120,000 enterprises use Gemini. In January 2026, Apple announced plans to integrate Gemini into Siri. Google processes over 10 billion tokens per minute via the API. Google's internal AI agents generate approximately 50% of the company's own code. Capital expenditures for 2026 are planned at $175–185 billion (nearly double 2025's $91.45B).

Bibliography

Primary Gemini technical reports

  • Gemini Team, Google (2023). Gemini: A Family of Highly Capable Multimodal Models. arXiv:2312.11805.
  • Gemini Team, Google (2024). Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context. arXiv:2403.05530.
  • Comanici, G. et al. (2025). Gemini 2.5: Pushing the Frontier with Advanced Reasoning, Multimodality, Long Context, and Next Generation Agentic Capabilities. arXiv:2507.06261.

Specialized models and applications

  • Saab, K. et al. (2024). Capabilities of Gemini Models in Medicine. arXiv:2404.18416.
  • Yang, L. et al. (2024). Advancing Multimodal Medical Capabilities of Gemini. arXiv:2405.03162.
  • Gemini Robotics Team (2025). Gemini Robotics: Bringing AI into the Physical World. arXiv:2503.20020.
  • Feng, T., Trinh, T., Bingham, G. et al. (2026). Semi-Autonomous Mathematics Discovery with Gemini: A Case Study on the Erdős Problems. arXiv:2601.22401.
  • DeepMind Research Team (2026). Building Production-Ready Probes For Gemini. arXiv:2601.11516.
  • Fu, Y., Wang, X., Tian, Y., Zhao, J. (2025). Deep Think with Confidence. arXiv:2508.15260.

Background literature (surveys and methods)

  • Wei, J. et al. (2022). Chain-of-Thought Prompting Elicits Reasoning in Large Language Models. arXiv:2201.11903.
  • Zhang, Z. et al. (2022). Automatic Chain of Thought Prompting in Large Language Models. arXiv:2210.03493.
  • Zhang, Z. et al. (2023). Multimodal Chain-of-Thought Reasoning in Language Models. arXiv:2302.00923.
  • Cai, W. et al. (2024). A Survey on Mixture of Experts in Large Language Models. arXiv:2407.06204.
  • Dai, Z. et al. (2019). Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context. arXiv:1901.02860.
  • Ding, J. et al. (2023). LongNet: Scaling Transformers to 1,000,000,000 Tokens. arXiv:2307.02486.
  • Yin, S. et al. (2024). A Survey on Multimodal Large Language Models. arXiv:2306.13549.
  • Wang, X. et al. (2023). Large-scale Multi-Modal Pre-trained Models: A Comprehensive Survey. arXiv:2302.10035.
  • Chen, Q. et al. (2025). Towards Reasoning Era: A Survey of Long Chain-of-Thought for Reasoning Large Language Models. arXiv:2503.09567.

Official Google blog posts

Retrieved from "https://systems-analysis.info/eng/index.php?title=Gemini_(Google)&oldid=162"