Google Gemini 3.5 Complete Guide 2026: Models, API, and When to Use It

Google's AI strategy has reached a new level of maturity in 2026. The Gemini 3.x model family — headlined by the newly released Gemini 3.5 Flash — represents one of the most significant jumps in cost-to-performance ratio the industry has seen. For the first time, a Flash-tier model is outperforming a Pro-tier model from the same generation on key benchmarks, while new additions like Gemini Omni and the Deep Think reasoning mode expand what the Gemini family can do.

Update — July 13, 2026: Unverified Gemini 3.5 Pro benchmark leak claims internal wins vs Fable 5 and GPT-5.6 with a July 17 rollout target — not confirmed by Google; Flash remains GA.

This guide covers everything you need to know to use Gemini 3.x effectively: the full model lineup, how to call the API in Python, what makes each model technically distinct, an honest comparison against Claude Fable 5 and GPT-5.6, and a clear answer to the question: when should you actually choose Gemini?

What Is Gemini 3.x?

Gemini is Google DeepMind's flagship AI model family, now in its third major generation. The defining characteristic of the entire Gemini line is native multimodality — every model in the family was trained on text, images, audio, and video simultaneously during pretraining, not through separate encoders or adapters bolted on after the fact. The practical result is a model that reasons across modalities as a single unified system rather than as loosely connected components.

Gemini 3.x also sits at the center of Google's product ecosystem in a way that no external provider can match. It powers AI Overviews in Google Search, the Gemini assistant in Android, Workspace AI features in Docs and Gmail, and the deep-research capabilities in NotebookLM. When you use Google products, you are already using Gemini.

The release of Gemini 3.5 Flash on June 8, 2026 — and its immediate promotion to the default model in Gemini Enterprise — signals that Google has achieved a new efficiency frontier: a model that is both fast and capable enough to be the default across their entire enterprise product line.

The Gemini 3.x Model Lineup

As of June 2026, the Gemini 3.x family consists of five distinct models, each targeting a different cost-performance point.

Gemini 3.5 Flash

The newest and most important model in the lineup. Gemini 3.5 Flash is Google DeepMind's latest generally available model, with a headline claim that sets it apart from every previous Flash-tier release: it rivals large flagship models on multiple benchmark dimensions. More pointedly, it outperforms Gemini 3.1 Pro on several key benchmarks while maintaining Flash-class latency and pricing.

This is the model Google chose to make the default for Gemini Enterprise on June 8, 2026 — a statement about where the value-to-cost balance sits in the current generation. For most workloads that previously required a Pro-tier model, 3.5 Flash is now the right starting point.

Best for: production pipelines, complex document analysis, customer-facing chat, coding tasks, high-volume workloads where quality matters, Workspace integrations.

Gemini 3.1 Pro

The high-capability flagship. Gemini 3.1 Pro is positioned for complex reasoning tasks, nuanced content generation, and workloads where you need the maximum quality ceiling regardless of cost or latency. While 3.5 Flash has narrowed the gap significantly, Pro still holds advantages on some nuanced generation and complex multi-step reasoning tasks where the additional parameter capacity makes a measurable difference.

Best for: research synthesis requiring maximum depth, complex agentic tasks, nuanced long-form writing, code generation for very large or complex codebases.

Gemini 3.1 Flash-Lite

The ultra-fast, ultra-cheap tier. Flash-Lite optimizes aggressively for speed and cost, accepting reduced reasoning depth in exchange for the lowest per-token pricing in the Gemini lineup. It is designed for workloads where you are processing millions of requests per day and quality requirements are modest.

Best for: text classification, intent routing, simple entity extraction, high-volume inference at minimal cost.

Gemini 3 Deep Think

Deep Think is not a separate model so much as a reasoning mode — Gemini's extended chain-of-thought system optimized for math, science, and formal logic. When activated, the model allocates a reasoning "scratchpad" before producing its final answer, working through the problem step-by-step. The performance gains on hard reasoning benchmarks (GPQA-Diamond, MATH, competition-level logic) are substantial, at the cost of higher latency and token usage.

Best for: mathematical problem-solving, scientific reasoning, formal logic, any task where getting the right answer matters more than response speed.

Gemini Omni

Gemini Omni is the newest addition to the family and the most architecturally interesting. It is designed from the ground up to blend text, audio, image, and video inputs into a single unified model output, going further than previous multimodal Gemini models by making all four modalities genuinely co-equal inputs. The model can reason about a video clip while cross-referencing spoken audio and text overlays, or analyze an image while listening to an audio description and referencing a text document — all within a single prompt.

Best for: video understanding and analysis, audio transcription with context, multi-modal document processing, any task that requires connecting information across text, images, audio, and video simultaneously.

Model Comparison Table

Verify current pricing at ai.google.dev/pricing before building cost models — pricing changes frequently in this market.

Gemini 3.5 Flash: What Makes It Special

The emergence of a Flash model that outperforms the previous generation's Pro model is the most important story in the Gemini 3.x launch. This is not a small efficiency gain — it represents a genuine shift in what "Flash" means as a capability tier.

Benchmark performance: Gemini 3.5 Flash outperforms Gemini 3.1 Pro on multiple key benchmarks, including coding evaluation, instruction-following, and several reasoning tasks. This is unusual. Historically, Flash models existed at a quality ceiling well below their Pro counterparts. The 3.5 generation has collapsed that gap while preserving Flash's latency and cost advantages.

Enterprise default: Google's decision to make 3.5 Flash the default model for Gemini Enterprise is the clearest possible signal of confidence. Enterprise products are where Google cannot afford quality regressions — they would hear about it immediately. Switching enterprise defaults is a statement that 3.5 Flash meets the bar for real production workloads at scale.

Speed at complexity: Flash has always been fast on simple tasks. What makes 3.5 Flash different is that it maintains that speed advantage even when handling complex documents, multi-turn conversations, and code generation tasks that would traditionally favor Pro. The latency advantage is preserved across the capability range.

Cost efficiency: For teams running high-volume production workloads, 3.5 Flash achieving Pro-level results at Flash pricing represents a meaningful reduction in AI infrastructure costs without any trade-off in output quality.

Gemini Omni: The New Multimodal Frontier

Gemini Omni represents the current frontier of native multimodal AI. While earlier multimodal models could process images alongside text, Omni treats audio, video, image, and text as genuinely co-equal modalities that the model reasons across in a unified way.

What "native multimodality" means in practice: Most early multimodal models appended visual information by encoding images and projecting them into the language model's token space. The underlying language model was still fundamentally text-centric, with vision as an add-on. Gemini Omni is trained differently — all modalities are present during pretraining, so the model builds internal representations that fuse information across modality boundaries. The result is stronger cross-modal reasoning.

Video understanding at scale: Omni can process video as a first-class input, analyzing not just frames but temporal sequences of events. This means it can answer questions like "what changed between minute 2 and minute 8 of this recording?" or "at what point in this presentation does the speaker introduce the financial projections?" These are tasks that require genuine video understanding, not just image captioning applied to keyframes.

Audio integration: Omni can process raw audio alongside other inputs, enabling transcription, speaker identification, and reasoning about tone and context in spoken content. Combined with video input, this enables analysis of recorded meetings, lectures, or interviews that previously required multiple specialized systems.

Practical use cases: customer support call analysis (audio + screen recording), educational content analysis (video + slides + transcript), product demo review (video + documentation), accessibility tools that describe visual content in audio.

Deep Think: Extended Reasoning for Hard Problems

Gemini's Deep Think mode addresses one of the fundamental challenges with large language models: they often produce confident-sounding wrong answers on problems that require careful multi-step reasoning. Extended reasoning modes — now available across the major AI providers — partially solve this by giving the model space to think before answering.

How Deep Think works: When the thinking budget is set in the API, the model allocates a scratchpad for chain-of-thought reasoning before producing its final response. The user sees only the final response, but the model's path to that answer involves working through intermediate steps, checking consistency, and revising approaches. This is analogous to showing your work on a math exam — the process of writing it out catches errors that would slip past a fast mental calculation.

Where Deep Think excels: Mathematical problem-solving (olympiad-level problems, financial modeling, statistical analysis), scientific reasoning (chemistry, physics, biology), formal logic (argument analysis, proof verification), and complex multi-step planning tasks. On benchmarks like GPQA-Diamond and MATH, the performance gap between thinking and non-thinking modes is large — often 10 to 20 percentage points.

The cost of thinking: Deep Think is significantly slower than standard generation and uses more tokens (the thinking scratchpad counts toward usage even when not shown to the user). For tasks where a fast approximate answer is acceptable, standard mode is preferable. For tasks where accuracy is critical and latency is secondary, Deep Think is the right choice.

Versus other reasoning modes: Claude Fable 5's adaptive thinking adjusts reasoning depth dynamically. GPT-5.6 Sol has its own extended reasoning capability. Deep Think is Gemini's answer — and on math and science benchmarks specifically, it is competitive with both.

Native Multimodality: How Gemini Is Built Differently

Every model in the Gemini 3.x family shares an architectural property that distinguishes it from many competitors: all modalities are trained together from the start.

This matters because of how AI models learn. A language model trained purely on text develops rich internal representations of language — the meanings of words, the structure of arguments, the patterns of code. When you later bolt on vision by training a separate image encoder and projecting image features into the language model's token space, the vision representation has to be learned in a way that aligns with the already-fixed language representations. This works reasonably well for describing images, but it creates seams when you ask the model to reason deeply across modalities.

Gemini's native multimodal training means the model's internal representations reflect patterns learned from all four modalities simultaneously. Text, images, audio, and video each influence how the model represents the others. The practical effect is smoother cross-modal reasoning — when you ask Gemini to "explain how the diagram on slide 12 relates to the equations discussed in the audio at timestamp 34:20," it handles that as a unified request rather than two separate queries stitched together with glue code.

For developers building applications that work with mixed-media content, this matters more than any individual benchmark. The model that handles your actual use case — which may involve PDFs with embedded charts, recorded interviews with slides, or video tutorials with code overlays — is more valuable than the model with the highest number on a text-only benchmark.

The 1 Million Token Context Window

One million tokens is roughly 750,000 words — about 1,500 dense pages of text, a complete large codebase, or an entire year's worth of project documentation. All Gemini 3.x models support this context window, and it remains one of the most practically significant technical differentiators in the current AI landscape.

What a 1M context window enables:

Entire codebase analysis: You can load a 100,000-line codebase and ask "find all the places where we're not handling database connection timeouts properly." No chunking, no RAG pipeline complexity, no context loss from splitting documents — just load everything and ask.

Full book or document analysis: A 400-page technical manual, a year of financial reports, a complete legal contract — these fit comfortably within 1M tokens. You can ask questions that require synthesizing information from different parts of a long document without worrying about what got cut off.

Long conversation history: Maintain months of conversation context without lossy summarization. For customer support, research assistants, or long-running project assistants, preserving full history changes what the model can do.

Multi-document synthesis: Load 50 research papers and ask for a literature review. Compare 20 vendor proposals. Cross-reference three years of board meeting minutes. Tasks that previously required sophisticated orchestration become single API calls.

The practical limits: Long context is not free. Processing 1M tokens is significantly more expensive than processing 10K tokens, and latency increases with context length. The right approach is to use long context when you need it — for tasks where the information genuinely requires it — and use shorter context when you do not. Gemini's architecture handles long context more gracefully than most competitors, but you still pay per token.

Google Search Grounding

Gemini models accessed through the API support grounding with Google Search — the model can retrieve live web information before generating a response. This is functionally different from retrieval-augmented generation (RAG) that you build yourself: it uses Google's own search infrastructure and is handled automatically by the model.

Search grounding is useful for any task where the answer might have changed since the model's training cutoff: current pricing, recent news, live product information, regulatory changes, or anything else that evolves over time. When grounding is enabled, the model queries Search, retrieves relevant results, and synthesizes them into its response — with citations.

The integration is particularly powerful in combination with Gemini's long context: you can provide a large document corpus from your own data and let the model augment it with live web information, giving you both deep contextual understanding and current information in a single response.

Using the Gemini API in Python

Getting Your API Key

1. Go to aistudio.google.com
2. Click "Get API key" in the left sidebar
3. Create a new API key (associate it with a Google Cloud project)
4. Store the key as an environment variable: export GOOGLE_API_KEY=your-key

Install the SDK

pip install google-generativeai

Basic Text Generation with Gemini 3.5 Flash

import google.generativeai as genai
import os

genai.configure(api_key=os.environ["GOOGLE_API_KEY"])

# Use Gemini 3.5 Flash for most production workloads
model = genai.GenerativeModel("gemini-3.5-flash")

response = model.generate_content(
    "Explain the difference between supervised and unsupervised learning "
    "in three paragraphs, suitable for a software engineer with no ML background."
)

print(response.text)

Multimodal Call: Image Analysis

import google.generativeai as genai
import PIL.Image
import os

genai.configure(api_key=os.environ["GOOGLE_API_KEY"])

model = genai.GenerativeModel("gemini-3.5-flash")

# Load an image from disk
image = PIL.Image.open("architecture_diagram.png")

response = model.generate_content([
    image,
    "This is a system architecture diagram. Identify potential single points of failure "
    "and suggest three improvements to make it more resilient."
])

print(response.text)

Long Document Analysis with 1M Context

import google.generativeai as genai
import os

genai.configure(api_key=os.environ["GOOGLE_API_KEY"])

# Upload a large PDF — the File API handles large documents
sample_pdf = genai.upload_file("annual_report_2025.pdf")

model = genai.GenerativeModel("gemini-3.5-flash")

response = model.generate_content([
    sample_pdf,
    "Summarize the key financial metrics, highlight any year-over-year risks, "
    "and list the three strategic priorities mentioned in the CEO letter."
])

print(response.text)

# Always clean up uploaded files when done
genai.delete_file(sample_pdf.name)

Activating Deep Think Mode for Hard Reasoning

import google.generativeai as genai
import os

genai.configure(api_key=os.environ["GOOGLE_API_KEY"])

# Use Deep Think for math, science, logic problems
model = genai.GenerativeModel("gemini-3.1-pro")

response = model.generate_content(
    "A factory produces widgets at a rate that doubles every 3 hours. "
    "Starting at 10 units/hour at 8 AM, how many total widgets are produced by 5 PM? "
    "Show your work step by step.",
    generation_config=genai.types.GenerationConfig(
        # thinking_budget: -1 = dynamic, or set a specific token budget
        # Higher budget = more thorough reasoning, higher cost and latency
        thinking_config=genai.types.ThinkingConfig(thinking_budget=16384)
    )
)

print(response.text)

Gemini Omni: Multimodal Video + Text Analysis

import google.generativeai as genai
import os

genai.configure(api_key=os.environ["GOOGLE_API_KEY"])

# Upload a video file for Omni analysis
video_file = genai.upload_file("product_demo.mp4")

# Wait for processing to complete
import time
while video_file.state.name == "PROCESSING":
    time.sleep(5)
    video_file = genai.get_file(video_file.name)

model = genai.GenerativeModel("gemini-omni")

response = model.generate_content([
    video_file,
    "This is a product demo video. Summarize what features are demonstrated, "
    "identify any bugs or usability issues shown, and list the three strongest "
    "selling points based on what you see."
])

print(response.text)
genai.delete_file(video_file.name)

Google Search Grounding

import google.generativeai as genai
import os

genai.configure(api_key=os.environ["GOOGLE_API_KEY"])

model = genai.GenerativeModel("gemini-3.5-flash")

# Enable Google Search as a tool for live web retrieval
tools = genai.types.Tool(google_search=genai.types.GoogleSearch())

response = model.generate_content(
    "What are the current pricing tiers for the major AI model providers "
    "as of today? Compare input token costs across Gemini, Claude, and OpenAI.",
    tools=[tools]
)

print(response.text)

Gemini 3.5 vs Claude Fable 5 vs GPT-5.6 — Honest Comparison

This is the question everyone is actually asking. Here is an honest assessment based on benchmarks and practical use cases as of June 2026. No single model wins every category.

Long Document Analysis

This is Gemini's most decisive advantage. The 1M token context window across all Gemini 3.x models means you can load entire documents without chunking. Claude Fable 5 also supports 1M tokens of context with its adaptive thinking, making this a closer race than it was with earlier Claude generations. GPT-5.6 Sol supports large context but not at the same scale.

For tasks that genuinely need to process entire large codebases, year-long archives, or multiple long documents simultaneously, Gemini's architecture and long-standing optimization for long context gives it an edge in reliability at the extremes.

Edge: Gemini, with Claude Fable 5 as a close second.

Coding

Claude Fable 5 (released June 9, 2026) leads here. Anthropic has invested heavily in instruction-following precision for code generation — the model's reliability at following complex, multi-constraint specifications without drifting is ahead of the field. Gemini 3.5 Flash is a strong second, and its context window advantage makes it easier to work with very large codebases without chunking. GPT-5.6 Sol is competitive but not the leader.

Edge: Claude Fable 5 for complex agentic coding. Gemini 3.5 Flash for code tasks involving large codebases.

Video Understanding

Gemini Omni leads clearly. Neither Claude Fable 5 nor GPT-5.6 Sol match Gemini's native video understanding capability. If your use case involves processing, summarizing, or analyzing video content, Gemini Omni is the obvious choice as of mid-2026.

Edge: Gemini Omni, clearly.

Speed and Cost at Scale

Gemini 3.5 Flash and Gemini 3.1 Flash-Lite are the most cost-efficient options for high-volume production workloads. Claude Fable 5 has its own speed-optimized tiers, and OpenAI's GPT-5.6 Luna tier competes on cost — but Gemini's Flash tier with 1M context at competitive pricing remains a strong option for workloads that need both scale and long context.

Edge: Gemini 3.5 Flash for long-context high-volume workloads. Competitive with GPT-5.6 Luna for short-context volume tasks.

Reasoning (Math, Science, Logic)

Deep Think gives Gemini a strong answer for formal reasoning tasks. Claude Fable 5's adaptive thinking dynamically adjusts reasoning depth. GPT-5.6 Sol has extended reasoning available across all tiers. On math olympiad-level problems and scientific reasoning benchmarks, the three models cluster closely. Your mileage will vary by specific domain and problem type.

Edge: No clear winner. Test on your specific problem type.

Instruction-Following

Claude Fable 5 leads here by a meaningful margin. This is Anthropic's core strength — reliably following complex, multi-constraint instructions without drifting, omitting requirements, or adding unsolicited content. For applications where prompt compliance is critical (structured extraction, formatted output, constrained generation), Claude has a consistent edge.

Edge: Claude Fable 5.

Google Ecosystem Integration

No competition here. If your stack is built on Google Cloud, Google Workspace, or Android, Gemini's native integration with IAM, billing, Workspace APIs, and Google Search is a major practical advantage that no other model provider can match.

Edge: Gemini, entirely.

Summary Table

The Google AI Ecosystem

Understanding Gemini means understanding how deeply it is woven into Google's product surface.

Gemini in Google Workspace

Gemini is integrated throughout Workspace — Docs, Sheets, Gmail, Slides, Meet. You can ask Gemini to draft a proposal based on previous emails, analyze a spreadsheet and surface anomalies, summarize a long thread, or generate a first draft of a presentation. With Gemini 3.5 Flash now the default for Gemini Enterprise, these features run on a significantly more capable model than before. Enterprise users on Business Plus and above get this through their Workspace subscription.

NotebookLM

NotebookLM is one of the most interesting applications of Gemini's long-context capabilities. You upload documents — PDFs, Google Docs, URLs, YouTube links — and it creates a private knowledge base you can query with natural language. Its standout feature is Audio Overviews: AI-generated podcast-style discussions of your source material with two synthetic hosts working through the content conversationally.

Researchers, analysts, students, and business professionals use NotebookLM to make dense material more accessible. The product's ability to handle large document collections reflects Gemini's long-context architecture at work.

Gemini in Android

Android ships with Gemini as the default AI assistant. Gemini can see your screen, understand the context of what you are doing in any app, take actions, and assist with tasks across the operating system. This tight hardware-software integration — the ability to observe and act on the live screen state of a device — is a capability that pure API providers cannot replicate. For mobile AI applications and Android development, the Gemini + Android pairing is unique.

Google Search AI Overviews

The AI-generated summaries at the top of Google Search results are powered by Gemini, now updated to draw on the 3.x generation. When users see a synthesized answer rather than a list of links, that is Gemini performing grounded retrieval and synthesis at unprecedented scale. From a developer perspective, the same grounding capability is available via the API's Google Search tool.

When to Choose Gemini Over Alternatives

Choose Gemini 3.5 Flash when:

You are running a production workload at scale and want the best balance of quality, speed, and cost in a single model. For most use cases that previously required Pro-tier models, 3.5 Flash now delivers equivalent or better quality at lower cost and latency.

Choose Gemini Omni when:

Your use case involves video. There is no serious alternative here — Gemini Omni's native video understanding is ahead of the field by a meaningful margin as of mid-2026. Audio analysis, multi-modal document processing, and any task requiring cross-modal reasoning across all four modalities also belong to Omni.

Choose Gemini 3.1 Pro when:

You have measured a quality gap between 3.5 Flash and Pro on your specific task, and quality improvement justifies the cost and latency difference. For some nuanced generation tasks and the most complex reasoning scenarios, Pro still holds an edge.

Choose Gemini Deep Think when:

The task involves math, formal logic, scientific reasoning, or any problem where accuracy matters more than speed and you can measure the quality improvement. Deep Think is a specialized tool, not a general-purpose upgrade.

Choose Gemini 3.1 Flash-Lite when:

You are processing very high volumes of simple tasks — classification, routing, extraction — where per-token cost is the primary constraint and reasoning depth is secondary.

Choose Gemini over Claude or GPT when:

You are building on Google Cloud or Google Workspace, and native ecosystem integration saves you infrastructure complexity. You need to process very long documents (over 200K tokens). Your use case involves video. Cost at scale with long context is a primary concern.

Consider Claude Fable 5 when:

Instruction-following precision is critical, coding reliability is paramount, or your team is already deeply invested in Anthropic's tool use and caching patterns.

Consider GPT-5.6 Sol when:

You are already in the OpenAI ecosystem and the third-party tooling advantage outweighs the capability differences, or when OpenAI's specific Sol/Terra/Luna tier structure fits your use case and budget.

Safety and Content Policy

Gemini models apply Google DeepMind's safety filters by default. These operate across four threshold levels (Block None, Block Few, Block Some, Block Most) for categories including sexually explicit content, dangerous instructions, hate speech, and harassment. Developers can configure these thresholds via the API for appropriate use cases with the required permissions.

Google AI Studio shows safety ratings on every response — a useful debugging tool when prompts hit filters unexpectedly. For enterprise deployments via Vertex AI, additional compliance controls, audit trails, and content filtering customization are available.

Google publishes model cards and safety evaluations for each major Gemini release, covering red-teaming results, bias evaluations, and capability assessments relevant to safe deployment.

Accessing Gemini: Where to Start

Google AI Studio (Start Here)

aistudio.google.com is the fastest path to a working API key. Sign in with a Google account, click "Get API key," and you have free-tier access to all Gemini 3.x models with generous rate limits for experimentation. The interactive playground lets you test prompts, upload files, and compare model outputs before writing any code.

Vertex AI (Production and Enterprise)

For production deployments on Google Cloud, Vertex AI provides Gemini access with SLAs, VPC Service Controls, customer-managed encryption keys, audit logging, and region control. If your data governance requirements or compliance obligations require it, Vertex AI is the right infrastructure. The model IDs and Python SDK work the same way — you swap the API key for a service account and configure the Cloud project.

Gemini Advanced (Consumer)

Gemini.google.com / Gemini Advanced is the consumer product — useful for personal productivity and experimentation, but not the path for developers who need API access or enterprise controls.

Getting Started: A Practical Checklist

If you want to start using Gemini 3.x today:

1. Get an API key from aistudio.google.com — free, no credit card required
2. Install google-generativeai (pip install google-generativeai)
3. Start with Gemini 3.5 Flash — it handles most workloads well at the best cost point
4. Test the 1M context window with a large document and see what opens up for your use case
5. Try Gemini Omni if your workflow involves video or audio
6. Turn on Deep Think only when you have a reasoning task where accuracy is the primary concern
7. If you are building for production on Google Cloud, evaluate Vertex AI for the compliance and observability features

Gemini 3.5 Flash changes the calculus for anyone who has been defaulting to Pro-tier models out of habit. The 3.x generation — particularly the combination of 3.5 Flash for general tasks, Omni for multimodal tasks, and Deep Think for reasoning tasks — gives developers a genuinely differentiated toolkit that no single competitor fully matches.

Read next

• Gemini 3.5 Pro benchmark leak — July 17 target, unverified vs Fable/GPT-5.6 (Jul 2026)
• Meta Llama 4: The Open-Source AI Model Guide 2026
• AI Benchmarks in 2026: The Complete Guide to MMLU, GPQA, SWE-bench, and Beyond
• Claude API Guide: Models, Pricing, and Best Practices 2026