LLMS · 122 REVIEWED
The Best Large Language Models (2026)
The general-purpose models everything else is built on: chat assistants, coding agents, and the APIs behind most AI products. They separate on three things — how well they reason, how much context they hold, and what a million tokens costs you.
Ranked by hand · 80 with a free tier · updated 2026-08-20
CATEGORY SNAPSHOT
TOP 10 COMPARED
| Tool | Artificial Analysis Intelligence Index | Pricing | API | Open weights | Best for |
|---|---|---|---|---|---|
| Claude Fable 5 | 59.9 | Paid | No | No | Creative Writing |
| NotebookLM | — | Free | No | No | Study & Research |
| Claude Opus 5 | 60.7 | Freemium | Yes | No | Best Frontier Model Overall |
| GPT-5.6 Sol | 58.9 | Paid | No | No | Professional-Grade Reasoning |
| Kimi K3 | 57.1 | Freemium | Yes | Yes | Open-Weight Frontier Performance |
| Claude Opus 4.8 | 55.7 | Enterprise | No | No | Enterprise Reasoning |
| Grok | — | Paid | Yes | No | Real-time |
| Grok 4.5 | 53.8 | Paid | No | No | Long-Running Coding Agents |
| DeepSeek | — | Freemium | Yes | Yes | Cost-Efficiency |
| Llama | — | Free | Yes | Yes | Open Source |
Scores from Artificial Analysis Intelligence Index, as of 2026-08-04. A dash means no published score for this tool.
ALL 122 LARGE LANGUAGE MODELS
Anthropic's Mythos-class creative model
Claude Fable 5 is Anthropic's Mythos-class model released on June 9, 2026, focused on creative writing, worldbuilding, and narrative depth
Why: Claude Fable 5 is notable as Anthropic's most experimental creative model. Even with limited availability, it represents an interesting direction for AI-assisted fiction and long-form creative work.
Paid
Best for Creative Writing
Visit
Google's AI Research Assistant: The Ultimate Study Tool
NotebookLM is an AI-first research and study assistant grounded in your own documents
Why: The Audio Overview feature is a viral sensation for a reason: it transforms dry study material into an engaging podcast. It is arguably the best free AI study tool available today.
Free
Best for Study & Research
Visit
Anthropic's frontier model, currently first on the Artificial Analysis Intelligence Index
Claude Opus 5 is Anthropic's flagship model, released 24 July 2026 with a 1M-token context window and five selectable effort levels (low, medium, high, xhigh, max)
Why: It is the current number one on the independent Artificial Analysis Intelligence Index, and it got there while costing less per task than the model it displaced: $2.03 average per index task against Fable 5's $2.75. The effort dial is the reason to pick it over a fixed-tier model, because one integration covers cheap high-volume calls and expensive long-horizon agent runs.
Freemium
Best Frontier Model Overall
Visit
OpenAI's top-tier model for complex professional work
GPT-5
Why: Sol ranks third overall on the Artificial Analysis Intelligence Index at 58.9, behind Claude Opus 5 and Claude Fable 5, a genuine top-tier frontier model rather than an incremental update, and OpenAI's clear flagship pick for the hardest professional-grade tasks.
Paid
Best for Professional-Grade Reasoning
Visit
Moonshot AI's 2.8-trillion-parameter open-weight flagship
Kimi K3 is Moonshot AI's open-weight model released July 16, 2026, built at roughly 2
Why: Kimi K3 is one of the most credible open-weight challengers to closed frontier models this year, aggressive enough on pricing and scale that it moved markets, Fortune covered it as a 'DeepSeek shock' moment for AI stocks.
Freemium
Best for Open-Weight Frontier Performance
Visit
Anthropic's powerful enterprise model from May 2026
Claude Opus 4
Why: Claude Opus 4.8 continues Anthropic's reputation for reliable, steerable models. It is a top choice for enterprises that need a capable assistant with strong safety characteristics and nuanced writing.
Enterprise
Best for Enterprise Reasoning
Visit
xAI's real-time AI assistant
Grok is xAI's AI assistant integrated into X (formerly Twitter) with real-time access to platform data and a more conversational, edgy tone
Why: xAI's AI assistant with unique real-time X platform integration and distinctive conversational style for social media context.
Paid
Best for Real-time
Visit
xAI's flagship coding model, trained in partnership with Cursor
Grok 4
Why: xAI positions it as its flagship 'for code and everything else,' and real benchmark data backs that up, 53.8% on the Artificial Analysis Intelligence Index (rank 7 overall). The Cursor training partnership is a distinctive angle: it's specifically tuned for long-horizon, multi-repository coding agent work, not just general chat.
Paid
Best for Long-Running Coding Agents
Visit
The Efficiency Revolution: Frontier Intelligence at 1/100th the Cost
DeepSeek is the architect of the 'DeepSeek movement,' a fundamental shift in AI development that prioritizes extreme efficiency over raw compute
Why: DeepSeek changed the game by proving that 'expensive' doesn't always mean 'better.' We picked it because it's the first model family to offer true frontier-level reasoning (R1), general intelligence (V3), and advanced vision/OCR (VL2) with an open-weight philosophy and an API price point that makes proprietary models look obsolete.
Freemium
Best for Cost-Efficiency
Visit
Meta's open-source large language model
Llama is Meta AI's open-source large language model family with multiple versions: Llama (February 2023), Llama 2 (July 2023), Llama 3 (April 2024), Llama 3
Why: Meta's flagship open-source LLM with strong performance, extensive model sizes, and permissive licensing for research and commercial use.
Free
Best for Open Source
Visit
Anthropic's cheaper, near-Opus everyday model
Claude Sonnet 5 is Anthropic's mid-tier model released June 30, 2026, replacing Sonnet 4
Why: Sonnet 5 is the practical default for most day-to-day work: it closes much of the gap to Opus-tier performance while staying meaningfully cheaper, and Anthropic made it the automatic replacement for Sonnet 4.6 across the free and Pro tiers.
Freemium
Best for Everyday Agentic Work
Visit
European open-source and commercial LLM
Mistral AI provides high-performance large language models with both open-source and commercial offerings
Why: European LLM provider with strong open-source offerings, multilingual capabilities, and focus on data privacy and compliance.
Freemium
Best for Europe
Visit
Enterprise-focused LLM platform
Cohere provides enterprise-grade large language models including Command, Command R, Command R+, Command R7
Why: Enterprise-focused LLM with strong RAG capabilities, multilingual support, and emphasis on accuracy and safety for business use cases.
Enterprise
Best for Enterprise
Visit
Alibaba's multilingual open-source LLM
Qwen is Alibaba Cloud's family of large language models with multiple versions: Qwen-1
Why: Alibaba's high-performance multilingual LLM with strong Chinese language support, cost-efficient pricing, and comprehensive open-source availability.
Freemium
Best for Multilingual
Visit
Microsoft's efficient small language models
Microsoft Phi is a family of small, efficient language models designed for high performance with minimal parameters
Why: Microsoft's efficient small language models with strong reasoning capabilities, MIT licensing, and optimized for resource-constrained environments.
Free
Best for Efficiency
Visit
Google's open-source lightweight LLM
Gemma is Google DeepMind's family of open-source large language models, serving as lightweight versions of Gemini
Why: Google's open-source LLM family with strong performance, permissive licensing, and specialized variants for vision and medical applications.
Free
Best for Research
Visit
The 'Next DeepSeek' Movement: o1-Level Reasoning at 1/100th the Cost
Kimi k1
Why: Kimi k1.5 is the first model to prove that o1-level reasoning is achievable through efficient, open-weight architectures. We selected it because it consistently matches or exceeds Claude 4.5 in technical benchmarks (AIME, MATH-500) while offering a 2M context window and a significantly lower API price point, making frontier intelligence accessible to everyone.
Freemium
Best for Technical Reasoning
Visit
The Open Vision-Reasoner: SOTA Multimodal Performance
Qwen 2
Why: We added Qwen 2.5-VL to the Open Frontier movement because it is currently the highest-performing open-weight vision model. It proves that open source can lead in multimodal reasoning, especially for tasks requiring high-resolution OCR and long-form video understanding.
Free
Best for Open Vision Reasoning
Visit
Databricks' high-performance open-source LLM
DBRX is a mixture-of-experts transformer model developed by Databricks and Mosaic ML
Why: Databricks' high-performance open-source LLM with strong benchmark results, efficient MoE architecture, and permissive licensing.
Enterprise
Best for Performance
Visit
Meta's Open Multimodal Standard
Llama 3
Why: We included Llama 3.2 Vision because it is the most widely supported open multimodal model in the world. Its integration into almost every AI tool and framework makes it the 'default' choice for open-weight vision reasoning.
Free
Best for Open Ecosystem Support
Visit
The Open Vision Frontier: 124B Multimodal Power
Pixtral Large is Mistral AI's flagship 124B parameter multimodal model, designed to compete directly with GPT-4o and Claude 3
Why: We added Pixtral Large because it represents the peak of European open-weight AI. It is one of the few open models that truly matches the visual reasoning depth of the top proprietary models, making it essential for the Open Frontier movement.
Freemium
Best for Complex Visual Reasoning
Visit
The Open-Source Vision Giant: 78B Multimodal Leader
InternVL 2
Why: We included InternVL 2.5 because it is a consistent leaderboard champion. It often outperforms much larger models in visual reasoning and OCR, making it a critical tool for developers who need GPT-4 level vision without the proprietary lock-in.
Free
Best for Leaderboard-Topping Vision
Visit
OpenAI's fast default ChatGPT model from May 2026
GPT-5
Why: GPT-5.5 Instant is the model most ChatGPT users will interact with by default. Its balance of speed and capability makes it a practical baseline for writing, analysis, coding help, and general assistant tasks.
Freemium
Best for Everyday ChatGPT
Visit
Cursor's agentic coding model for multi-file software engineering
Cursor Composer 2
Why: Composer 2.5 moves Cursor further from autocomplete toward genuine pair-programming autonomy. For teams already using Cursor, it is the most integrated way to turn high-level feature requests into working code across many files.
Paid
Best for IDE Autonomy
Visit
Open-source, model-agnostic terminal coding agent
OpenCode is an open-source terminal-based coding agent that connects to multiple language models and performs autonomous software engineering tasks from the command line
Why: OpenCode is the best 'bring-your-own-model' coding agent for developers who want full control. Because it is open source and runs in the terminal, it fits naturally into existing CI/CD and shell-centric workflows without locking you into a specific vendor.
Free
Best for Terminal Coding
Visit
Google's fast, capable multimodal model from I/O 2026
Gemini 3
Why: Gemini 3.5 Flash hits a practical sweet spot for developers and creators who need more capability than entry-level models but do not require the full cost of an Ultra model. Its native multimodal design makes it especially useful for mixed-media tasks.
Freemium
Best for Fast Multimodality
Visit
xAI's agentic coding CLI for autonomous software engineering
Grok Build is an agentic command-line coding assistant from xAI that understands natural-language project descriptions, generates and edits code across files, runs commands, and iterates until tasks a...
Why: Grok Build brings xAI's frontier reasoning directly into the terminal, making it a strong alternative to other agentic coding CLIs. It is particularly useful for developers already embedded in the X and xAI ecosystem who want a fast, opinionated agent.
Freemium
Best for Agentic CLI
Visit
Google's unified multimodal generation model
Gemini Omni is a single Google model announced at I/O 2026 that can generate and reason across text, images, video, and audio from unified prompts
Why: Gemini Omni represents Google's push toward a single model for all media types. For teams building multimodal products, it simplifies architecture by replacing multiple specialized endpoints with one interface.
Freemium
Best for Unified Generation
Visit
Google's personal AI agent for proactive assistance
Gemini Spark is a personal AI agent announced at Google I/O on May 19, 2026
Why: Gemini Spark is Google's answer to the emerging personal-agent category. By integrating deeply with Gmail, Calendar, Maps, and Android, it can automate everyday tasks that previously required switching between apps.
Freemium
Best for Personal Agent
Visit
Mistral's unified work and coding agent
Mistral Vibe is Mistral's unified agent for work and coding, rebranded from Le Chat and launched on May 28, 2026
Why: Mistral Vibe brings Mistral's strong European model lineage into a competitive all-in-one agent. It is a good choice for users who want a privacy-conscious alternative to US-centric assistants with solid coding skills.
Freemium
Best for European AI Assistant
Visit
The ceiling of enterprise autonomy with 1M context
Anthropic's most powerful model, designed for autonomous software engineering and complex reasoning
Why: Claude Opus 4.6 is like a super-smart digital architect. While most AI can only write short snippets, Opus can 'see' your entire project (up to 1 million words) at once. It doesn't just help you code; it can actually build complex software systems from scratch, making it the best choice for big companies that need an AI 'teammate' rather than just a chatbot.
Enterprise
Best for Autonomy
Visit
DeepSeek's open-weight model with permanent pricing
DeepSeek V4-Pro is a high-performance language model from DeepSeek
Why: DeepSeek V4-Pro stands out for combining frontier-level performance with transparent, permanent pricing and open weights. It is a practical choice for teams that want to self-host or avoid unpredictable API costs.
Freemium
Best for Predictable Pricing
Visit
Moonshot's specialized coding model
Kimi K2
Why: Kimi K2.7-Code is one of the strongest coding models from a Chinese AI lab, with particular strength in long-context understanding and bilingual code tasks. It is a good addition for teams evaluating global coding models.
Freemium
Best for Bilingual Coding
Visit
MiniMax's 1M-context agentic frontier model
MiniMax M3 is a 1-million-token-context agentic frontier model released on May 31, 2026
Why: MiniMax M3's 1M context window makes it competitive for tasks that require digesting entire codebases, books, or video transcripts in a single pass. It is a strong option for long-context agentic applications.
Freemium
Best for 1M Context
Visit
StepFun's 198B MoE vision-language model
StepFun Step 3
Why: Step 3.7 Flash offers a competitive Chinese-frontier multimodal model with an MoE architecture that balances capability and inference cost. It is a useful option for vision-language applications and for teams exploring alternatives to US models.
Freemium
Best for Efficient VLM
Visit
80B parameter open-weight coding powerhouse
Alibaba's latest open-weight model specialized for coding
Why: Qwen3-Coder-Next is the best 'Private Brain' for coders. Most AI tools send your secret code to the internet, but this one can live entirely on your own computer. It's just as smart as the big paid tools, but it keeps your work 100% private and safe.
Free
Best for Open Coding
Visit
The frontier model for complex reasoning and software architecture
OpenAI's most advanced model to date, featuring a 2M context window and specialized training for complex multi-step reasoning
Why: GPT-5.3 Codex is the 'World's Smartest Planner.' While other AI tools are good at chatting, this one is built for solving huge, difficult problems like planning how a whole software system should work. It has a massive memory (2 million words) so it never loses track of the big picture.
Paid
Best for Reasoning
Visit
The industry standard for coding and nuanced instruction following
Anthropic's flagship model, optimized for high-speed coding and perfect adherence to complex XML-based system prompts
Why: Claude 4.6 Sonnet is the 'Perfect Student' for following directions. It is famous for doing exactly what you ask without getting confused. It also has a special 'Computer Use' feature where it can actually move the mouse and type on your screen to do chores for you.
Paid
Best for Coding
Visit
Native multimodal intelligence with a 10M context window
Google's most powerful multimodal model, capable of processing hours of video, thousands of lines of code, or massive document sets in a single prompt
Why: Gemini 3 Ultra offers an unbeatable 10M token context window, allowing it to process entire project histories, hours of video, or massive codebases in a single prompt. Its native multimodal intelligence makes it the only model capable of 'seeing' and 'hearing' complex data sets with the same level of depth as it reads text, providing a unique advantage for large-scale data analysis.
Paid
Best for Context
Visit
The conversational search engine that replaced traditional search
Perplexity uses frontier LLMs to browse the web in real-time and provide cited, accurate answers to complex queries
Why: Perplexity AI is the 'Death of the Search Engine.' Instead of giving you a list of 10 links to click on, it just reads the whole internet for you and gives you a single, cited answer. It's like having a personal researcher who never sleeps.
Freemium
Best for Research
Visit
AI search engine for peer-reviewed scientific research
Consensus searches over 200 million scientific papers to provide evidence-based answers
Why: The 'Truth' layer for AI. It solves the hallucination problem in research by grounding every answer in peer-reviewed science.
Freemium
Best for Science
Visit
One multimodal model for text, vision, audio, and video reasoning
Nemotron 3 Nano Omni is NVIDIA's compact-but-capable multimodal stack for agentic workflows: one family of endpoints that accept text, images, audio, or video (depending on route) and return text answ...
Why: If your product roadmap says 'agents that see and hear the world,' Omni is built for that integration story, fewer moving parts than bolting Whisper + CLIP + LLM together by hand.
Paid
Best for Agents
Visit
Alibaba's open-source MoE flagship with thinking modes
Qwen 3 is a 2025 open-weight Mixture-of-Experts model family from Alibaba Cloud, ranging from 0
Free
Best for Open-Source Agents
Visit
Alibaba's strongest vision-language model
Qwen-VL-Max is a high-performance vision-language model from Alibaba, capable of understanding images, charts, and documents, and answering questions about them
Freemium
Best for Vision-Language
Visit
Open mathematical reasoning specialist
Qwen-Math is a family of open-weight models specialized for mathematical reasoning and problem solving, derived from Qwen 2
Free
Best for Math Reasoning
Visit
Fast, low-cost Claude model with extended thinking and Computer Use
Anthropic's Haiku-tier model announced on October 15, 2025, and the first Haiku model to support both extended thinking and Computer Use
Why: Haiku 4.5 is the only current Haiku model and is the first to bring Computer Use and extended thinking to the low-cost tier.
Paid
Best for Fast, Low-Cost Agents
Visit
Balanced Sonnet model with major coding and agentic improvements
Anthropic's Sonnet-tier model announced on September 29, 2025, with major coding, instruction-following, and agentic improvements at the same price as Sonnet 4
Why: Sonnet 4.5 brought meaningful coding and agentic upgrades at the same price as Sonnet 4, making it a notable missing mid-tier entry.
Paid
Best for Balanced Coding Agents
Visit
First Claude model with the effort parameter and context compaction
Anthropic's Opus-tier model announced on November 24, 2025, introducing the effort parameter for balancing capability against cost, context compaction, and a deeper memory tool
Why: Opus 4.5 was the first Claude model to ship the effort parameter, an important capability evolution before Opus 4.6 and 4.7.
Enterprise
Best for Cost-Capability Tradeoffs
Visit
Frontier Opus model with higher-resolution vision and xhigh effort
Anthropic's frontier Opus-tier model announced on April 16, 2026, with substantial gains on the hardest coding tasks, higher-resolution vision input, a new xhigh effort level, file-system-based memory...
Why: Opus 4.7 introduced the xhigh effort level and file-system memory recall, making it a notable step between Opus 4.6 and Opus 4.8.
Enterprise
Best for Hard Coding Tasks
Visit
Limited-availability Mythos-class model without Fable 5 safety classifiers
Anthropic's Mythos-class model announced on June 9, 2026, shares the same capabilities as Claude Fable 5 without the safety classifiers
Why: Mythos 5 is a notable limited-availability variant of the Mythos-class tier, distinct from the generally available Fable 5.
Enterprise
Best for Controlled Research
Visit
Cursor's first-generation agentic coding model
Cursor Composer 1 is the first-generation agentic model inside the Cursor IDE
Why: Composer 1 introduced the original agentic editing experience inside Cursor that later evolved into the stronger long-context planning of Composer 2.5.
Paid
Best for Agentic Editing
Visit
High-volume DeepSeek inference with a 1M-token context window
DeepSeek V4-Flash is the efficient sibling of V4-Pro, offering a 1M-token context window and configurable thinking modes at a fraction of the API cost
Why: V4-Flash delivers the same 1M context and thinking modes as V4-Pro at roughly one-third the API cost, making it the practical default for most production workloads.
Freemium
Best for High-Volume APIs
Visit
The open-weight reasoning model that sparked the efficiency revolution
DeepSeek R1 is a 671B-parameter open-weight reasoning model that matches o1-class performance on math, code, and logic benchmarks through reinforcement learning on verifiable tasks
Why: R1 proved that open-weight models can match proprietary reasoning systems at a fraction of the cost, making it a landmark for reproducible AI research.
Freemium
Best for Open Reasoning
Visit
The 128K-context MoE flagship that introduced sparse attention
DeepSeek V3
Why: V3.2 introduced DeepSeek Sparse Attention and unified thinking modes, making it the architectural bridge that enabled the later 1M-context V4 family.
Freemium
Best for Long-Context MoE
Visit
744B-parameter open-weight MoE flagship for agentic planning and execution
GLM-5 is Zhipu AI's (Z
Why: GLM-5 anchors the open-weight GLM line as a commercially-usable Chinese flagship with a permissive license and a strong reasoning profile.
Paid
Best for Open-Weight Frontier
Visit
MIT-licensed MoE flagship for 8-hour autonomous coding sessions
GLM-5
Why: GLM-5.1 is the open-weight coding release that made Z.ai competitive on long-horizon agentic work while remaining MIT-licensed for unrestricted commercial use.
Paid
Best for Long-Horizon Coding
Visit
Optimized GLM-5 variant for fast sequential task execution
GLM-5-Turbo is a tuned variant of the GLM-5 series that prioritizes lower latency and efficient sequential execution
Why: GLM-5-Turbo is the practical speed layer for the GLM-5 family, trading a small amount of peak capability for noticeably faster multi-step agent execution.
Paid
Best for Fast Sequential Tasks
Visit
Strong general-reasoning model with interleaved thinking
GLM-4
Why: GLM-4.7 is the cost-effective sweet spot for long-context reasoning and general-purpose agent work before stepping up to the GLM-5 series.
Paid
Best for General Reasoning
Visit
Mid-range coding and tool-calling model with 200K context
GLM-4
Why: GLM-4.6 gives developers a capable, lower-cost GLM option for coding and tool-calling agents without sacrificing the long context window.
Paid
Best for Coding & Tool Calls
Visit
Cost-efficient reasoning, coding, and agent model
GLM-4
Why: GLM-4.5-Air extends the GLM-4.5 family downward with a low-cost model that still handles coding, reasoning, and agent tasks.
Paid
Best for Budget Reasoning
Visit
Multimodal coding and visual-reasoning agent model
GLM-5V-Turbo is a vision-language variant of the GLM-5 family, built for multimodal coding, visual reasoning, and image-plus-text agent workflows
Why: GLM-5V-Turbo is the GLM family's main vision agent, letting coding and agent workflows reason over images and screenshots in the same long context.
Paid
Best for Multimodal Coding
Visit
Vision-language model for visual reasoning and UI replication
GLM-4
Why: GLM-4.6V is the practical vision tier for turning screenshots and images into working code or structured analysis.
Paid
Best for Visual Reasoning
Visit
Document parsing model for PDF and image OCR
GLM-OCR is a specialized GLM model for extracting structured Markdown from PDFs and images
Why: GLM-OCR fills a clear gap in the GLM family by turning scanned documents and PDFs into structured, usable text with layout awareness.
Paid
Best for Document Parsing
Visit
Free universal GLM model with a 200K context window
GLM-4
Why: GLM-4.7-Flash is the most capable free-tier GLM option, making long-context prototyping accessible without a subscription.
Free
Best for Free General Use
Visit
Free vision model for image understanding and document snapshots
GLM-4V-Flash is a free-tier vision model in the GLM family, offering zero-cost image understanding and document snapshot analysis with a 16K context window
Why: GLM-4V-Flash is the entry-level vision option for GLM, letting users test multimodal document understanding before upgrading to paid vision tiers.
Free
Best for Free Vision
Visit
Google's long-context multimodal flagship with up to 2M tokens
Gemini 1
Freemium
Best for Long Context
Visit
Fast, cost-efficient multimodal model with a 1M context window
Gemini 1
Freemium
Best for Fast Multimodal Tasks
Visit
Google's low-latency agentic model with native tool use
Gemini 2
Freemium
Best for Agentic Apps
Visit
Google's high-performance reasoning model with advanced coding
Gemini 2
Freemium
Best for Complex Reasoning
Visit
Google's open multimodal model for research and developers
Gemma 3 is an open-weights family of multimodal models from Google, ranging from 1B to 27B parameters
Free
Best for Open Multimodal
Visit
OpenAI's balanced GPT-5.6 model for intelligence and cost
GPT-5
Why: Terra is the sensible default for most GPT-5.6 work: it delivers the lion's share of Sol's capability at roughly 40% of the cost and is the default model for ChatGPT Free and Go users.
Freemium
Best for Balanced Cost and Capability
Visit
OpenAI's cost-optimized GPT-5.6 model for high-volume workloads
GPT-5
Why: Luna brings GPT-5.6-scale capabilities to high-volume applications at roughly one-tenth of Sol's cost, with strong enough performance for everyday tasks and broad API availability.
Freemium
Best for Cost-Sensitive Workloads
Visit
xAI's long-context flagship with a 1M-token window
Grok 4
Why: Grok 4.3 is the sweet spot in xAI's lineup for anyone who needs a frontier model with a very large context window at a lower price than Grok 4.5.
Paid
Best for Long-Context Work
Visit
xAI's high-volume, 2M-context workhorse model
Grok 4
Why: Grok 4.1 Fast delivers one of the largest context windows in the family at the lowest price point, making it the default pick for bulk work.
Paid
Best for Volume and Cost Efficiency
Visit
xAI's 2M-context beta model with multi-agent capabilities
Grok 4
Why: Grok 4.20 remains notable as xAI's first multi-agent beta model with a 2M context window, even though newer 4.3/4.5 models now offer flagship alternatives.
Paid
Best for Multi-Agent Beta Work
Visit
xAI's fast, cheap coding specialist model
Grok Code Fast 1 is a lightweight coding-focused model from xAI with a 256K context window
Why: Grok Code Fast 1 is the practical, low-cost coding specialist in xAI's family for developers who want Grok reasoning without flagship pricing.
Paid
Best for Fast Coding Assistance
Visit
Tencent's 389B-parameter open-source MoE language model
Open-source Transformer-based Mixture-of-Experts language model with 389 billion total parameters and 52 billion active parameters
Why: Largest open-source Transformer-based MoE model from Tencent, ideal for researchers and builders who want to self-host a capable long-context LLM.
Free
Best for Open-Source LLM Workloads
Visit
Tencent's Mamba-powered deep-thinking reasoning model
Hybrid Mamba-Transformer MoE reasoning model released March 2025, built on Hunyuan TurboS with 52 billion active parameters and a 256K context window
Why: One of the first ultra-large Mamba-Transformer MoE reasoning models, offering strong benchmark scores and a 256K context window.
Freemium
Best for Reasoning
Visit
Tencent's fast, cost-efficient flagship Hunyuan model
A 200K-context open-weight Hunyuan model optimized for speed while maintaining strong performance on general chat, coding, and agentic tasks
Why: Speed-optimized Hunyuan flagship with a 200K context window and strong price/performance for production APIs.
Freemium
Best for Speed
Visit
Tencent's general-purpose instruction-tuned Hunyuan 2.0 model
Open-weight instruction-tuned variant of Tencent's Hunyuan 2
Why: Versatile instruction-tuned Hunyuan model balancing capability and context for a wide range of tasks.
Freemium
Best for General-Purpose Chat
Visit
The deep-thinking variant of Hunyuan 2.0
Open-weight reasoning variant of Hunyuan 2
Why: Hunyuan 2.0's reasoning mode for tasks that benefit from longer thought chains.
Freemium
Best for Reasoning
Visit
Tencent's efficient small-scale MoE instruct model
A compact open-weight MoE instruction model with a 131K context window, listed as a cost-efficient everyday Hunyuan option via OpenRouter and Tencent Cloud
Why: Smallest listed Hunyuan instruct model, making it attractive for budget-conscious long-context deployments.
Freemium
Best for Cost-Efficient Inference
Visit
Tencent's latest open-source MoE flagship with tool use
Open-weight preview of Hunyuan 3 (Hy3), a 295B-parameter MoE model with 21B active parameters and a 262K context window
Why: Tencent's strongest open-source Hunyuan model to date, with competitive coding and agentic benchmarks.
Freemium
Best for Coding and Agents
Visit
Moonshot's open-weight multimodal generalist with agent swarms
Kimi K2
Why: Kimi K2.5 was Moonshot's first widely available open-weight multimodal generalist and remains a notable reference point for the K2 family before K2.6 and K3 arrived.
Freemium
Best for Open Multimodal Agents
Visit
Moonshot's open-weight multimodal successor with long-context coding stability
Kimi K2
Why: Kimi K2.6 improves on K2.5 with stronger long-context coding and is a practical open-weight alternative for teams that want multimodal agents without the cost of closed frontier models.
Freemium
Best for Long-Context Coding
Visit
Faster inference variant of Kimi's coding specialist
Kimi K2
Why: Kimi K2.7 Code Highspeed is the latency-optimized version of an already strong coding model, making it a good pick for interactive coding agents and live pair-programming workflows.
Freemium
Best for Fast Coding
Visit
Meta's open-weight flagship with native multimodal reasoning
Llama 4 Maverick is Meta's flagship open-weight model, released in April 2025 as part of the Llama 4 family
Why: Maverick is the top open-weight model Meta actually ships today, with strong multimodal reasoning and a practical API ecosystem, making it the default choice for open Llama deployments.
Free
Best for Open Multimodal Reasoning
Visit
Long-context, efficient open multimodal model for edge and single-GPU use
Llama 4 Scout is Meta's efficient Llama 4 variant, released in April 2025
Why: Scout is notable for its extreme 10M-token context window and efficient single-GPU deployment, making it the standout open model for very long documents and memory-heavy applications.
Free
Best for Long Context
Visit
Efficient 70B open model matching 405B quality
Llama 3
Why: Llama 3.3 is the practical sweet spot in the Llama family: it gives users near-frontier open-model quality in a 70B package that is far cheaper to host and fine-tune than the 405B model.
Free
Best for Efficient Open LLMs
Visit
The first frontier-scale open-weight language model
Llama 3
Why: Llama 3.1 405B remains a landmark open release: it proved open weights could compete with proprietary frontier models and still serves as a high-quality baseline for research and synthetic-data generation.
Free
Best for Frontier Open Research
Visit
Microsoft's everyday AI assistant across web, PC, and mobile
Microsoft Copilot is the free consumer AI assistant formerly known as Bing Chat
Why: The free, broadly available Microsoft AI assistant that brings search, chat, and image generation into one cross-platform experience.
Freemium
Best for Everyday AI
Visit
AI assistant embedded across Word, Excel, PowerPoint, Outlook, and Teams
Microsoft 365 Copilot is an enterprise AI assistant that integrates with Microsoft 365 apps and organizational data through Microsoft Graph
Why: The enterprise-grade AI assistant that grounds responses in your Microsoft 365 data and works directly inside Office apps.
Enterprise
Best for Enterprise Productivity
Visit
Recursive self-improvement language model for real-world engineering
MiniMax M2
Why: MiniMax M2.7 is the current production language model below M3 and is explicitly listed as beginning recursive self-improvement, making it a notable addition to the family.
Freemium
Best for Engineering Tasks
Visit
Same M2.7 performance with significantly faster inference
MiniMax M2
Why: The Highspeed variant is a current, actively promoted option for developers who need M2.7 capability with lower latency.
Freemium
Best for Low-Latency Coding
Visit
Mistral's flagship open-weight multimodal frontier model
A 675B-parameter sparse mixture-of-experts model with 41B active parameters and a 262K context window, released under Apache 2
Why: Mistral Large 3 is one of the most capable permissive open-weight models available, offering frontier performance with the deployment flexibility of Apache 2.0 licensing.
Freemium
Best for Open-Weight Frontier
Visit
Mistral's mid-tier workhorse for reasoning, coding, and instruction
A mid-tier model that balances performance and cost, optimized for instruction following, reasoning, and coding
Why: Mistral Medium 3.5 delivers strong performance at a lower cost than the flagship, making it the sensible default for most business and development workloads.
Freemium
Best for Everyday Workloads
Visit
Unified open-source small model for chat, reasoning, vision, and coding
A 119B-parameter MoE model with 6B active parameters and a 256K context window, released under Apache 2
Why: Small 4 packs flagship-class reasoning, vision, and coding into a single open-source model that is efficient enough for high-throughput and local deployments.
Freemium
Best for Efficient Open Multimodal
Visit
Mistral's edge family of small, dense open-source models
A family of 3B, 8B, and 14B parameter dense models released under Apache 2
Why: Ministral 3 brings Mistral's open-weight lineage to edge devices, offering a strong 14B reasoning option and smaller variants for local and on-device use.
Freemium
Best for Edge Deployment
Visit
Compact 30B open-weight model with configurable reasoning for agents
A 30B total / 3B active parameter hybrid Mamba-2 + Transformer MoE language model built for efficient on-device and edge agentic tasks
Why: The smallest open-weight member of the Nemotron 3 family, giving teams frontier-style reasoning and tool-use without data-center hardware.
Free
Best for Efficient Agents
Visit
120B open-weight hybrid MoE for efficient multi-agent reasoning
A 120B total / 12B active parameter hybrid Mamba-Transformer MoE language model with LatentMoE, multi-token prediction, and native NVFP4 pretraining
Why: Fills the gap between Nano and Ultra with a strong efficiency-to-accuracy ratio for agentic orchestration and latency-sensitive serving.
Free
Best for Multi-Agent Efficiency
Visit
NVIDIA-aligned 253B Llama 3.1 for helpfulness and instruction following
A 253B-parameter variant of Llama 3
Why: NVIDIA's largest aligned Llama collaboration, offering a strong open-weight alternative for teams already standardizing on Llama architectures.
Free
Best for Aligned Llama Performance
Visit
NVIDIA-aligned 49B Llama 3.1 for balanced performance
A 49B-parameter variant of Llama 3
Why: A mid-size aligned Llama model that balances capability and deployment cost for teams using NVIDIA tooling.
Free
Best for Balanced Llama Deployment
Visit
NVIDIA-aligned 8B Llama 3.1 for efficient inference
An 8B-parameter variant of Llama 3
Why: A compact, NVIDIA-aligned Llama model for teams that need HelpSteer-tuned instruction following on limited hardware.
Free
Best for Efficient Aligned Llama
Visit
Lightweight, real-time search model for cited answers at low cost
Perplexity's base Sonar model pairs live web search with a compact LLM to deliver fast, citation-backed answers
Why: Sonar is the affordable, fast entry point to Perplexity's live-search API and is widely used as the default model for citation-backed Q&A.
Paid
Best for Everyday Search
Visit
Advanced search model with deeper reasoning and richer citations
Sonar Pro uses a more capable model and expanded search context to answer complex questions with detailed, source-backed responses
Why: Sonar Pro adds the depth and reliability needed for serious research while remaining accessible through Perplexity's API and Pro app tier.
Paid
Best for Deep Research
Visit
Chain-of-thought reasoning model for multi-step logical analysis
Sonar Reasoning Pro exposes explicit chain-of-thought reasoning to solve complex, multi-step problems with transparent intermediate steps
Why: Sonar Reasoning Pro is Perplexity's option for users who need transparent, step-by-step reasoning rather than just a final answer.
Paid
Best for Complex Reasoning
Visit
Autonomous research agent that performs multi-source deep dives
Sonar Deep Research conducts autonomous, multi-step research across many web sources and synthesizes a comprehensive, cited report
Why: Sonar Deep Research is the most agentic member of the Sonar family, capable of running lengthy, autonomous research workflows and producing publishable reports.
Enterprise
Best for Autonomous Research
Visit
Anonymous 1M-context reasoning model available free through OpenRouter
Ox Alpha is a stealth AI model that appeared on OpenRouter and OpenCode on 20 August 2026
Why: The combination of a one-million-token context window, multimodal inputs, free pricing during the preview, and rapid adoption by coding-agent builders makes it worth tracking even before its creator is known. Within a day of launch, coding agents had pushed billions of tokens through it, suggesting real production interest rather than curiosity traffic. Preliminary independent DeepSWE testing also places it ahead of Claude Fable 5 and GPT-5.6 Sol on a small task subset.
Free
Best for Anonymous Preview
Visit
Google's coding workhorse, three weeks after 3.6 Flash
Gemini 3
Why: It is the cheapest route to a current-generation Google coding model. Introductory pricing runs at half the standard Flash rate until the end of 2026, and the published jump over 3.6 Flash is large enough to matter on exactly the agentic and web-development work Flash-tier models are usually bought for.
Freemium
Best for Coding Value
Visit
Microsoft's unified AI model family from Build 2026
Microsoft announced a family of MAI-branded models at Build 2026, including MAI-Thinking-1 for reasoning, MAI-Image-2
Why: The MAI family gives Microsoft a cohesive, enterprise-ready AI stack. For organizations already using Microsoft services, these models reduce friction by running inside familiar tools rather than requiring separate platforms.
Enterprise
Best for Microsoft Ecosystem
Visit
Open-source MoE LLM with strong Chinese NLP and multimodal capabilities
Baidu ERNIE 4
Why: Leading Chinese LLM with strong multilingual capabilities, open-source availability, and cost-efficient MoE architecture.
Freemium
Best for Chinese
Visit
Advanced multilingual LLM with enhanced reasoning and long-context support
GLM-4
Why: Advanced Chinese LLM with strong multilingual capabilities, efficient inference, and comprehensive deployment options.
Freemium
Best for Multilingual
Visit
Z.ai's post-trained coding and agentic model on the GLM-5.2 base
GLM-5
Why: The Terminal-Bench 3.0 score moved from 4.6% to 28.3%, DeepSWE v1.1 from 46.2% to 66.9%, and CyberGym from 77.2% to 84.5% — all on the same base as GLM-5.2. That is a real signal about post-training returns, even if most numbers are vendor-run and weights are not yet released. It is also reported as one of the fastest models in its class, at roughly 115 tokens per second.
Paid
Best for Post-Training Gains
Visit
Meta's closed-weight agentic model, and its first paid model API
Muse Spark 1
Why: This is the release where Meta stopped giving models away. Muse Spark is closed, metered and sold through Meta's own API, and it lands at rank 15 on the independent index while undercutting comparable models on price. Worth tracking for that reason alone if your stack assumed Meta meant open weights.
Freemium
Best Value for Agentic Multimodal Work
Visit
NVIDIA's 550B open-weights reasoning model, built for inference speed
Nemotron 3 Ultra is the largest member of NVIDIA's Nemotron 3 family, released 4 June 2026 at Computex
Why: The architecture is optimised for throughput rather than peak benchmark score, and it shows: over 400 output tokens per second at 550B parameters. It is also the most openly documented release at this scale, because publishing the datasets and post-training recipes lets you actually reproduce and extend the model instead of just running it.
Free
Best for Open-Weight Throughput
Visit
Autonomous AI agent for complex multi-step workflows and research automation
Manus is an autonomous AI agent developed by Butterfly Effect Pte
Why: Pioneering autonomous AI agent platform with proven real-world task execution capabilities, now backed by Meta's resources.
Enterprise
Best for Automation
Visit
Google's faster, sharper agentic-coding upgrade to 3.5 Flash
Gemini 3
Why: Gemini 3.6 Flash is the clearest upgrade path for teams already running high-volume agentic and coding workloads on Flash-tier pricing. It delivers a real benchmark jump over 3.5 Flash without moving up to Ultra-tier cost.
Freemium
Best for Fast Agentic Coding
Visit
Alibaba's 2.4-trillion-parameter flagship, currently in preview
Qwen 3
Why: Qwen 3.8-Max is Alibaba's answer to the current wave of massive open-weight-adjacent models from Chinese labs, and the discounted preview pricing makes it worth evaluating early even before the full release details land.
Paid
Best for Long-Horizon Agentic Work (Preview)
Visit
753B open-weight MoE coding model with a 1M-token context, MIT licensed
GLM-5
Why: The strongest open-weight coding model published to date: 62.1 on SWE-bench Pro against GPT-5.5's 58.6, and 81.0 on Terminal-Bench 2.1, at roughly a sixth of GPT-5.5's API price. The MIT licence carries no regional restrictions, so the weights can genuinely be self-hosted commercially, which is the reason to choose it over a closed model of similar strength.
Freemium
Best Open-Weight Coder
Visit
Thinking Machines' 975B Apache-2.0 model that takes text, images and audio natively
Inkling is the first open-weights model from Thinking Machines Lab, released 15 July 2026 under Apache 2
Why: It is the first roughly trillion-parameter open-weights model that takes audio and images natively rather than through a bolted-on encoder, and Apache 2.0 means the weights can be used commercially without asking anyone. Thinking Machines is candid that this is not the strongest model available but a base worth customising, which is a more honest pitch than most open releases make.
Free
Best Open-Weight Multimodal Base
Visit
HOW TO CHOOSE
What actually decides between large language models:
- Benchmark score, then your own eval: Published intelligence scores are a starting filter, not an answer. Every serious deployment ends up scored against its own task set, because the gap between models on a benchmark rarely matches the gap on your work.
- Price per million tokens: The spread between frontier and mid-tier models is often 10-30x. A model that is 4% better and 20x more expensive is the wrong default for anything running at volume.
- Context window, honestly measured: A stated 1M-token window does not mean the model uses all of it well. Retrieval accuracy usually degrades long before the limit, so test at the length you actually work at.
- Open weights or API: Open-weight models can run on your own hardware, be fine-tuned, and cannot be deprecated out from under you. API models are faster to adopt and usually stronger at the frontier.
- Deprecation risk: Hosted models get retired. If a model is load-bearing for you, check the provider's deprecation policy before you build on it.