Modele AI

Latest Fable — long-running agentic coding, knowledge work, and research with 1M context and 75% cheaper cache reads

Previous Fable — demanding reasoning, complex agentic coding, and long-horizon work with 1M context

Latest Opus — multi-step codebase changes, code review, financial analysis, and long-horizon agentic work with 1M context

Previous Opus — prefer Claude Opus 5.5 for hardest reasoning and coding

Previous Opus — enhanced coding, agentic workflows, and long-horizon reasoning with 1M context

Previous Opus — enhanced SWE, vision, and long-horizon agentic reasoning with 1M context

Hybrid reasoning model with 1M context, top-tier coding and agentic performance

Previous premium model with maximum intelligence

Latest Sonnet — 30%+ faster with stronger coding and agentic performance at the same list price as Sonnet 5

Previous Sonnet — prefer Claude Sonnet 5.5 for speed and cost per task

Previous Sonnet — strong speed and intelligence balance

Previous smart model for complex agents and coding

Latest Haiku — fastest Claude with 1M context; list price ~10× lower input vs Haiku 4.5 for prompts ≤100K tokens

Previous Haiku — prefer Claude Haiku 5.5 for speed and cost

Anthropic flagship Opus via OpenRouter — multi-step codebase work, code review, and long-horizon agentic tasks

Latest Anthropic Sonnet via OpenRouter — fast coding, agents, and everyday intelligence with 1M context

OpenAI flagship for complex reasoning, coding, computer use, and agentic work

Upgraded GPT-6 Sol — near-Astra coding and agentic work with cheaper cached input; prefer over GPT-6 Sol for new projects

Upgraded agentic execution model — stronger tool loops and visual reasoning vs GPT-5.6 Terra at ~50% lower token cost

Cost-efficient GPT-6 high-end model — prefer GPT-6.1 Sol for agentic coding and long-horizon software engineering

Fast, cost-efficient GPT-6 model for chat, classification, high-volume agents, and subagent loops

GPT-5.6 alias — prefer GPT-6 Astra for frontier work; routes to GPT-5.6 Sol

Prior frontier GPT-5.6 model — prefer GPT-6 Astra for hardest reasoning and coding

GPT-5.6 agentic execution — prefer GPT-6.1 Terra for new agent and tool workflows

Cost-efficient GPT-5.6 model for high-volume workloads and subagent loops

Prior frontier model — prefer GPT-5.6 Sol for new projects

Highly capable GPT model for coding and agentic tasks — prefer GPT-5.6 Terra at same price

Fast GPT-5.4 variant for coding, computer use, and high-volume subagent workloads

Most powerful GPT model with maximum compute for complex reasoning

The best model for coding and agentic tasks across industries

Intelligent reasoning model with configurable reasoning effort

Previous intelligent reasoning model for coding and agentic tasks

A faster, cost-efficient version of GPT-5

Fastest, most cost-efficient version of GPT-5

Smarter and more precise responses

Smartest non-reasoning model

Smaller, faster version of GPT-4.1

Reasoning model for complex tasks

Version of o3 with more compute for better responses

Fast, cost-efficient reasoning model

GPT-6 Luna with reasoning.mode pro — higher-quality responses on complex tasks while staying cost-efficient

GPT-6 Sol with reasoning.mode pro — stronger reasoning on demanding coding and agentic tasks

Google DeepMind frontier — long-horizon software engineering, enterprise knowledge work, and agentic workflows (intro $2/$10 per 1M tokens through Nov 2026)

Most advanced reasoning model with complex problem-solving

Cheapest frontier-class model — half the cost of Gemini 3 Flash with strong tool calling

Default Live API model for low-latency voice agents — fluid dialogue without reasoning delays

High-reasoning Live API model for complex multi-step voice interactions with background tool use

Most intelligent Flash model — long-horizon software engineering, autonomous agents, and enterprise workflows at Flash speed (intro pricing through Dec 2026)

Newest Flash workhorse — best-in-class coding, agents, and multimodal at half the original 3.6 Flash cost (intro pricing through Dec 2026)

Stable Flash — strong agentic and multimodal performance with lower output cost than 3.5 Flash

Fastest 3.5-class model — high-throughput chat, tools, and vision at lower cost than full Flash

Frontier intelligence optimized for agentic workflows, coding, and video at higher speed

Frontier intelligence with superior search and grounding

State-of-the-art thinking model for complex problems

Best price-performance for large scale processing

Fastest flash model for cost-efficiency

Latest frontier Grok — coding, knowledge work, long-running agents, self-verification — xhigh reasoning, tool calling, web/X search

Previous frontier Grok — long-running agents, coding, interactive apps, and visual work — xhigh reasoning, tool calling, web/X search

Previous frontier Grok for coding, agentic tasks, and knowledge work — tool calling, web/X search, code execution

Fast, intelligent Grok — 1M context, 3 reasoning levels, top agentic tool calling

Latest Grok beta optimized for multi-agent orchestration

Latest Grok beta with extended reasoning

Fast Grok without reasoning overhead

Smaller, faster Grok with reasoning

Oct 17, 2026 V4.1 Flash snapshot (CED MoE) — native vision, coding & agentic workloads, 1M context, compressed KV cache

1M context, thinking + non-thinking modes, tool calls

Flagship model — 1M context, thinking + non-thinking modes

Kunya Tab v1

Kunya

Szybki

Fast Tab completion router — routes to the configured flash model with curation telemetry.

Latest M-series Flash — 1M context, multimodal (image & video), agentic coding, tunable thinking depth via reasoning_effort

M-series flagship — 1M context, agentic coding, tool use, vision + video input, interleaved thinking

Recursive self-improvement — SOTA in software engineering, tool calling, and office productivity

M2.7 at ~100 tps — same performance, faster and more agile

Peak performance and ultimate value — master the complex

M2.5 at ~100 tps — same performance, faster and more agile

Polyglot programming mastery with precision code refactoring

Agentic capabilities with function calling and advanced reasoning

ByteDance flagship — 76.5% SWE-Bench, 98.3% AIME 2025, hour-long video understanding

Fastest reasoning diffusion LLM — parallel token generation (~140 tps), tunable reasoning, tool calling, and JSON schema output. Built for search agents, voice pipelines, and coding subagents.

Meta's flagship Llama 4 model

Efficient Llama 4 model

Meta's powerful open source model

Meta's powerful open source model

Balanced Mistral model

Latest large Mistral model

Frontier multimodal Mistral — 512K context, 256K output, vision, tool calling, and structured JSON

Creative writing focused model

Large Qwen model with 235B parameters

Vision-language Qwen model

Most powerful Qwen model

Z.ai reasoning flagship — complex software engineering and long-horizon agent tasks with 1M context (reasoning always on)

Native multimodal flash model — efficient coding and long-horizon agent tasks with 1M context, image & video input, tool calling

High-speed GLM 5.3 serving tier — 1.5–2× throughput vs GLM 5.3 with the same reasoning, tools, and 1M context

High-speed GLM 5.3 Flash serving tier — up to ~200 tok/s for coding, vision, and agent tasks with 1M context

Enhanced long-horizon coding and autonomous agent tasks

Z-AI flagship model with strong reasoning and tool use

Fast inference model optimized for agentic workflows and tool use

Latest GLM model

Powerful GLM model

Lightweight GLM model

Large-scale reasoning model — long-horizon agent workflows, project-level software engineering, 256K context, tools + structured outputs. Free on OpenRouter (rate-limited preview).

Cost-efficient long-context model — office productivity, document-intensive work, coding, and agentic workflows with 524K context

Native omnimodal model — Pro-level agentic performance at half the cost, with image & video understanding and 1M context

Xiaomi's 1T-parameter flagship — agentic workflows, tool calling, and advanced reasoning with 1M context

Xiaomi's fast AI model

309B MoE open foundation model — 1M context, native multimodal, strong coding and agentic workflows at low cost

1T+ parameter flagship — 1M context, native multimodal, top-tier agentic coding, visual, and research workloads

Same 1T MiMo v2.6 Pro checkpoint with ~10× output speed — 1M context multimodal agentic workflows

550B MoE frontier reasoning model (55B active) — long-horizon agentic workflows, coding agents, and deep research with 1M context. Free on OpenRouter. Web search via Kunya smart search (native tool calling not supported).

Nvidia's compact free MoE model for fast chat and lightweight tasks

30B MoE (3B active) — high-throughput agentic workloads, tool calling, 1M context. Free on OpenRouter (rate-limited). Newer/fast tier vs Nemotron 3 Nano.

Multimodal visual reasoning over images and video — dense scenes, frame tracking, and physical-world grounding for agentic visual workflows

Open multimodal MoE (12B active) — image & video understanding, tool calling, 1M context for vision-heavy agents

Free tier of Nemotron 3.5 Super VL — multimodal image & video chat with 1M context (rate-limited on OpenRouter)

Moonshot pro-tier flagship with 1M context, native vision and video input, tool calling, and reasoning for complex agentic and knowledge work

2.8T MoE flagship for complex coding, knowledge work, and long-horizon agentic workflows with 1M context and native vision

Coding-focused MoE model for long-horizon programming, agentic task decomposition, and multimodal reasoning

State-of-the-art visual coding and agentic tool-calling with multimodal reasoning

196B MoE reasoning model — activates 11B per token, extremely fast

196B MoE multimodal model — native image & video understanding, selectable reasoning depth

StepFun flagship agentic MoE (27B active) — sparse experts, image & video input, tool use, 1M context

Alibaba's most capable model — 2.4T MoE (95B active), 1M context, vision, agentic coding, office productivity, and long-horizon autonomous execution

Higher-throughput Qwen 3.8 Max SKU — same 2.4T MoE capability with faster serving, vision, tools, 1M context

Next-gen multimodal reasoning MoE — coding, agentic workflows, visual and long-video analysis, document and codebase understanding, 1M context

First agentic omni-modal Qwen — native audio, image & video input, tool calling, 1M context for multimodal agents

TypeSafe meta-router — picks model and reasoning effort per request for quality, speed, and cost balance

Fast vision-language reasoning model — multimodal agents, visual coding, search, and computer interaction with 1M context. Web search via Kunya tools + smart-search fallback when native tool calls stall.

Flagship Qwen3.7 — 1M context, agentic coding, office productivity, long-horizon execution, prompt caching

Cost-effective Qwen3.7 VL model — 1M context, vision + tools + agent workflows, GUI and mobile navigation

1T parameter frontier model built for agentic multi-step reasoning

Omni-modal frontier model with vision, hearing, reasoning, and action

Flagship uncensored reasoning model from Nous Research — hybrid think/respond mode, low refusal rates, strong at math, code, and structured output

Efficient uncensored reasoning model from Nous Research — hybrid think/respond mode, low refusal rates, strong at math, code, and structured output

Perplexity agentic search — multi-step research with citations (compare vs Brave / Gemini grounding)

295B MoE reasoning model (21B active) — agentic workflows, 256K context, configurable chain-of-thought for coding, analysis, and tool use

295B MoE reasoning model (21B active) — agentic workflows, 256K context, configurable chain-of-thought. Free on Kunya until July 21.

Poolside flagship coding agent — tool calling, reasoning, 256K context, up to 32K output. Free on Kunya until July 28.

Meta frontier reasoning model for long-horizon agentic work — plans, delegates to sub-agents, parallel tool use, and error recovery across text, image, video, audio, and PDF input with 1M context. Strong on multi-file engineering, extended sessions, and browser or desktop automation workflows.

Meta multimodal reasoning model for long-running agentic, multi-agent, and coding workflows — text, image, video, PDF, and audio input with 1M context, tool calling, structured output, and configurable reasoning. Available globally on OpenRouter (unlike Muse Spark 1.1). Audio understanding is still limited upstream.

Sakana Namazu

Sakana

Premium

Japanese-specialized reasoning model based on Kimi K2.6 — instruction following, business writing, math, coding, research, and agent workflows. Built-in web search and code execution (billed separately by provider). Sakana may use inputs for training by default (opt-out in Sakana Console).

Fugu Ultra

Sakana

Premium

Sakana multi-agent orchestration for complex reasoning, coding, and agentic workflows with 1M context. Configurable reasoning effort, tool calling, and built-in web search. Orchestration tokens billed as standard I/O. Not available in EU/EEA via Sakana.

Multimodal foundation model — text, image & video input, 1M context, agentic coding and tool use. Free OpenRouter preview (rate-limited); use minimax-m3.1-flash for direct paid API.

Reasoning model built for coding, sustained agentic work, and production workloads — 1M context, 131K output, image & video input, tool calling. Free on Kunya until August 27.

Versatile multimodal model with low latency for agent and vision tasks

Z.ai GLM 5.2 via EvoLink — long-horizon coding agents, repo Q&A, tool-using engineering workflows, 1M context with thinking + prompt caching

Kunya V1

Kunya

Premium

Intelligently routed model — Opus-level quality at budget cost. Routes to the best model for each request.

Kunya

Kunya

Szybki

Kunya picks the model for you — included with the free trial and Kunya Basic.

Intelligent routing optimized for creative writing — Opus planning, Sonnet prose, Terra agentic, DeepSeek Nebius on budget tiers.

Cost-optimized router — DeepSeek Nebius prose tiers, Terra agentic, Haiku 4.5 fallback, Gemini Flash Lite vision.

Fastest, lowest-cost Nano Banana — native 1K, ~4s generation, editing & 14 aspect ratios

High-efficiency image generation optimized for speed and volume, up to 4K with thinking

GA workhorse — improved quality, text rendering, multi-turn consistency, and Search grounding at Flash speed (1K/2K/4K)

Professional asset production with advanced reasoning and 4K output

Fast everyday OpenAI image generation and editing — higher quality than GPT Image 2 at lower latency

Most capable OpenAI image model — premium quality and precise edits for production creative work

Prior OpenAI image generation — flexible sizes; Batch API 50% discount still available