النماذج والأسعار

Picks the model for each task: Opus 5.5 for hard work, cheaper models for simple turns.

Anthropic's first Claude 5.5 model. Fable 5.1-level performance on most work at 40% lower cost than Opus 5, with a 1M context window; thinking is always on and forced tool_choice is not supported.

Anthropic's most capable model for demanding reasoning and long-horizon agentic work, with a 1M context window. Successor to Fable 5 with cheaper cache reads; forced tool_choice is not supported.

OpenAI's most capable model, built for the hardest end-to-end work: complex reasoning, coding, computer use, research, and document creation. 1M+ context window, async tool calling, mid-turn steering, and reasoning effort from low through max.

DeepSeek's smallest new-architecture model: natively multimodal with 1M context window, faster and cheaper than V4 Pro while exceeding it on agentic, coding and reasoning tasks. Supports image input, tool calling and thinking mode.

Z-AI's flagship reasoning model on a 1M-token context. Reasoning is always on, with low, high, and max effort levels, and strong coding and tool-calling.

Anthropic's most agentic Sonnet model, with performance approaching Opus 4.8 at a fraction of the cost.

Anthropic's Fable 5 model, with a 1M context window. Shares Opus 4.8's reasoning efforts and modalities.

Anthropic's Fable 5 model, with a 1M context window. Shares Opus 4.8's reasoning efforts and modalities.

Anthropic's most agentic Sonnet model, with performance approaching Opus 4.8 at a fraction of the cost.

Anthropic's model for complex agentic coding and enterprise work. Near-Fable 5 intelligence at half the price, with a 1M context window.

Anthropic's model for complex agentic coding and enterprise work. Near-Fable 5 intelligence at half the price, with a 1M context window.

Anthropic's model for complex agentic coding and enterprise work. Near-Fable 5 intelligence at half the price, with a 1M context window.

Google's Flash model with near-Pro intelligence at Flash-tier cost and speed: strong coding, parallel agentic execution, thinking, and search grounding.

OpenAI's GPT-5.6 Sol is a frontier model with a 1M+ context window, reasoning, and broad tool support including web search, file search, image generation, code interpreter, hosted shell, computer use, and MCP.

OpenAI's cost-optimized variant of GPT-5.6 with a 1M+ context window, reasoning, and broad tool support at a fraction of the flagship cost.

OpenAI's most affordable and fastest variant of GPT-5.6 with a 1M+ context window, optimized for high-volume, latency-sensitive applications at minimal cost.

OpenAI's GPT-5.5 is a frontier model with a 1M+ context window, reasoning, and broad tool support including web search, file search, image generation, code interpreter, hosted shell, computer use, and MCP. Knowledge cutoff December 1, 2025.

OpenAI's GPT-5.4 is OpenAI's latest frontier model with a 1M+ context window, improved reasoning with xhigh effort support, and enhanced capabilities for coding, agentic tasks, and computer use.

OpenAI's efficient, cost-optimized variant of GPT-5.4 with strong reasoning and multimodal capabilities at a fraction of the cost.

DeepSeek's fast, cost-efficient general-purpose model with 1M context window, updated July 2026 with substantially enhanced agentic capabilities. Supports tool calling and thinking mode.

Tencent Hunyuan's 295B/21B MoE model for real-world business scenarios: native 256K context, three reasoning modes, strong coding, long-form comprehension, multi-turn dialogue, and agentic task execution.

DeepSeek's fast, cost-efficient general-purpose model with 1M context window, updated July 2026 with substantially enhanced agentic capabilities. Supports tool calling and thinking mode.

Xiaomi's MiMo-V2.5 native omnimodal model with strong agentic capabilities, supporting text, image, video, and audio understanding within a unified architecture. Built on the MiMo-V2-Flash backbone with dedicated vision and audio encoders.

Z-AI's latest flagship model for long-horizon tasks. Delivers a substantial leap in long-horizon capability over GLM-5.1 on a solid 1M-token context, with stronger coding and flexible thinking effort.

DeepSeek's flagship reasoning model with 1M context window. Supports tool calling, extended thinking, and high-quality generation.

MiniMax's natively multimodal flagship with a 1M-token context window powered by MiniMax Sparse Attention (MSA). Supports image and PDF input, toggleable thinking, strong agentic tool calling, and long-context reasoning.

Moonshot AI's K3 flagship model for long-horizon coding and end-to-end knowledge work, with a 1M-token context window.

StepFun AI's flagship high-efficiency multimodal reasoning model on a sparse Mixture-of-Experts architecture (198B total, ~11B active) with a 256K context window, selectable reasoning effort, native image understanding, and multi-step function calling.

Google's latest model with the Flash line's focus on latency, efficiency, and cost, in preview version.

Anthropic's latest Sonnet model with hybrid reasoning, matching near-flagship performance at a fraction of the cost.

Google's most cost-efficient and lowest-latency multimodal model in the 2.5 family, optimized for high-volume classification, simple data extraction, and low-latency tasks.

OpenAI's most affordable and fastest variant of GPT-5.6 with a 1M+ context window, optimized for high-volume, latency-sensitive applications at minimal cost.

Google's fast and cost-effective model with a 1M token context window. Best for high-volume, low-latency tasks and agentic use cases.

Xiaomi's MiMo-V2.5-Pro is an open-source Mixture-of-Experts (MoE) language model with 1.02T total parameters and 42B active parameters. It utilizes the hybrid attention architecture and 3-layer Multi-Token Prediction (MTP) introduced in MiMo-V2-Flash.

Google's lightweight and efficient model in the 3.1 family, optimized for low latency and cost, in preview version.

Dense 30.7B multimodal model from Google DeepMind. Multimodal (text + image input), text output.

DeepSeek's latest general-purpose model with improved reasoning, coding, and instruction-following capabilities.

Efficient MoE variant of Gemma 4 from Google DeepMind. Multimodal (text + image input), text output.

xAI's Grok 4.5 with 500K token context window. Supports configurable reasoning, function calling, structured outputs, and vision.

Anthropic's most capable generally available model. Step-change improvement in agentic coding over Opus 4.6, with a new tokenizer and 1M context window.

Alibaba's 2.4-trillion-parameter MoE flagship with a 1M-token context window. Deep-thinking model tuned for coding, professional work, and long-horizon agentic execution, with native visual understanding of documents and video.

Anthropic's fastest model with near-frontier intelligence, ideal for high-throughput applications.

OpenAI's cost-efficient small model. Great for lightweight tasks with fast responses and low cost while maintaining strong capabilities.

MiniMax's latest flagship model with 200K context window and 131K max output. Advanced agentic capabilities with multi-agent collaboration, strong coding, and tool calling.

Google's most cost-efficient Gemini model, built for high-throughput and agentic use cases with thinking, tool calling, and search grounding.

Google's latest Flash-class multimodal model, balancing latency and capability with thinking, function calling, and search grounding.

OpenAI's smaller and faster variant of GPT-5, a more cost-efficient alternative. Great for well-defined tasks and precise prompts.

Moonshot AI's next-gen agentic model built on K2. Long-horizon coding, proactive autonomous execution, and swarm-based task orchestration.

Google's latest model, in preview version.

Google's Flash model for agentic and multimodal work: thinking, search grounding, code execution, and computer use over a 1M-token context.

Anthropic's most recent model. Current leader on agentic coding evaluation, Terminal Bench 2.0, and Humanity's Last Exam.

Alibaba's Qwen3.7 Flash native vision-language model with a 1M-token context window. Hybrid thinking model supporting text and image inputs, tool calling, structured output, and web search, with upgraded multimodal agent and coding capabilities.

Alibaba's Qwen3.6 model with 35B total / 3B active parameters using mixture-of-experts. Strong general-purpose performance.

Moonshot AI's native multimodal agentic model built on K2. Excels at visual coding, reasoning, and self-directed agent swarm with up to 100 sub-agents.

Moonshot AI's K2.7 model specialized for coding. FP4-quantized for fast inference with a 262K context window.

Google's lightweight and efficient model in the 3.1 family, optimized for low latency and cost, in preview version.

OpenAI's most affordable and fastest variant of GPT-5.4, optimized for high-volume, latency-sensitive applications with minimal cost.

Alibaba's Qwen3.7 Plus model with a 1M-token context window. Hybrid thinking model supporting text and image inputs, tool calling, structured output, and web search.

OpenAI's balanced small model with a massive 1M token context window. Offers great performance at low cost, beating GPT-4o in many benchmarks.

Anthropic's latest Sonnet model featuring exceptional performance on coding, analysis, and instruction following.

Z-AI's next-generation flagship model for agentic engineering. Stronger coding capabilities and state-of-the-art performance on SWE-Bench Pro.

ZhipuAI's most capable reasoning model with 128K context window. Supports tool calling, web search, and thinking.

Alibaba's flagship Qwen3.7 agent model with a 1M-token context window. Hybrid thinking model tuned for coding, agentic workflows, and long-horizon autonomous execution. Supports tool calling and reasoning.

xAI's Grok 4.3 with 1M token context window. Supports configurable reasoning, function calling, structured outputs, and vision.

ZhipuAI's advanced reasoning model with 128K context window. Supports tool calling, web search, and thinking.

Amazon's fastest text-only model optimized for speed and cost. Ideal for text-based tasks requiring low latency with 128K context window.

OpenAI's cost-optimized variant of GPT-5.6 with a 1M+ context window, reasoning, and broad tool support at a fraction of the flagship cost.

OpenAI's latest and greatest, GPT 5.2 is OpenAI's flagship model for coding and agentic tasks across industries.

Meta's efficient 8B parameter model optimized for multilingual dialogue. Fast inference with great performance for everyday tasks.

OpenAI's most cost-effective and efficient model in the GPT-5 series, optimized for speed and affordability. Ideal for straightforward tasks and high-volume applications where latency and cost are critical.

OpenAI's GPT-5.3-Codex is the most capable agentic coding model, combining frontier coding performance with reasoning capabilities. Features mid-task steering and 25% faster inference than GPT-5.2-Codex.

MiniMax's latest flagship model with 200K context window, ~60 tokens/second. Supports function calling and thinking.

Alibaba's Qwen3.6 Plus model with a 1M-token context window. Hybrid thinking model supporting text and image inputs, tool calling, structured output, and web search.

Meta's powerful 70B parameter model quantized to FP8 for fast inference. Excellent for complex reasoning and multilingual tasks.

Google's largest Gemma 3 model with 27B parameters. Strong performance across reasoning, coding, and multilingual tasks with vision support.

Alibaba's Qwen3.5 9B open-source model with a 256K-token context window. Hybrid thinking model supporting text and image inputs, tool calling, and structured output.

Alibaba's Qwen3.6 27B dense model with reasoning and vision support. Strong general-purpose performance at a small footprint.

Alibaba's latest Qwen3 coding model. Cutting-edge code generation and understanding capabilities.

Meta's Llama 4 Maverick with 17B active parameters using mixture-of-experts. Excels at coding, reasoning, and multilingual tasks.

Alibaba's Qwen3 dense 32B model. Strong performance across reasoning, math, and coding tasks.

Meta's latest Llama 4 model with 17B active parameters using mixture-of-experts architecture.

Alibaba's Qwen3.5 35B-A3B open-source model with a 256K-token context window. Hybrid thinking model supporting text and image inputs, tool calling, structured output, and web search.

Google's lightweight open model well-suited for text generation and image understanding. Supports LoRA fine-tuning with 128K context.

Alibaba's Qwen3.5 native vision-language model with 122B total / 10B active parameters using a hybrid linear-attention sparse mixture-of-experts architecture.

Mistral's efficient 24B model optimized for simple tasks with low latency and 128K context. Great for classification, customer support, text generation, and multimodal tasks.

Alibaba's Qwen3 vision-language model with 235B total / 22B active parameters. Supports text and image inputs.

OpenAI's previous flagship model for coding and agentic tasks with configurable reasoning and non-reasoning effort.

Anthropic's former flagship model combining maximum intelligence with practical performance.

OpenAI's former advanced reasoning model with enhanced problem-solving capabilities, deeper understanding, and improved accuracy across complex tasks.

ZhipuAI's fast and efficient model variant. Optimized for speed while maintaining strong language capabilities.

Alibaba's Qwen3.5 397B-A17B open-source model with a 256K-token context window. Hybrid thinking model supporting text and image inputs, tool calling, structured output, and web search.

Meta's 70B parameter model with 128K context. Strong performance on reasoning, coding, and multilingual tasks.

ZhipuAI's reasoning model with 128K context window. Supports tool calling, web search, and thinking.

OpenAI's multimodal model optimized for speed and efficiency. Capable of processing text, images, and audio with high accuracy at reduced latency and cost compared to GPT-4 Turbo.

Alibaba's latest Qwen3 model with 80B total / 3B active parameters using mixture-of-experts. Strong general-purpose performance.

Alibaba's Qwen3.6 Flash model with a 1M-token context window. Hybrid thinking model supporting text and image inputs, tool calling, structured output, and web search.

OpenAI's cost-efficient small model. Great for lightweight tasks with fast responses and low cost while maintaining strong capabilities.

DeepSeek R1 May 2025 update. Reasoning model with strong performance on math, code, and complex reasoning tasks.

Alibaba's Qwen3.5 27B open-source model with a 256K-token context window. Thinking model supporting text and image inputs, tool calling, and structured output.

Alibaba's Qwen3 vision-language model with 30B total / 3B active parameters. Supports text and image inputs.

Alibaba's Qwen3 coding-focused model with 30B total / 3B active parameters using mixture-of-experts architecture.

Alibaba's Qwen3 14B model with strong reasoning and instruction-following. A capable mid-sized model for coding and complex tasks.

ZhipuAI's efficient reasoning model with 128K context window. Supports tool calling, web search, and thinking.

OpenAI's GPT-5.6 Sol is a frontier model with a 1M+ context window, reasoning, and broad tool support including web search, file search, image generation, code interpreter, hosted shell, computer use, and MCP.

Z-AI's first natively multimodal GLM-5 model: image and PDF input on a 1M-token context at flash pricing. Reasoning is always on (low/high/max), with function calling and implicit prompt caching.

xAI's frontier model for coding, agentic tasks, and knowledge work, with a 500K token context window. Supports configurable reasoning (including xhigh), function calling, structured outputs, web search, image and PDF input.

Anthropic's previous flagship model with maximum intelligence. Legacy model - consider using Opus 4.5 for new projects.

Mistral's specialized model for code generation. Optimized for coding tasks including code completion, generation, and explanation.

Command A is Cohere's most performant model to date, excelling at tool use, agents, retrieval augmented generation (RAG), and multilingual use cases.

DeepSeek's reasoning model trained with reinforcement learning. Excels at math, code, and complex reasoning tasks.

DeepSeek's reasoning model distilled from R1 based on Qwen2.5. Outperforms OpenAI o1-mini across various benchmarks with state-of-the-art dense model results.

DeepSeek V3 (March 2024 release) — general-purpose text model with strong reasoning, coding, and instruction-following capabilities.

DeepSeek V3.1 hybrid model with optional thinking mode, function calling, and a 128K context window.

DeepSeek's fast, cost-efficient general-purpose model with 1M context window. Supports tool calling and thinking mode.

DeepSeek's experimental multimodal model. Matches DeepSeek V4 Flash on text — agentic use, reasoning and world knowledge — and adds image understanding, at V4 Flash pricing. Supports tool calling and thinking mode.

DeepSeek's flagship reasoning model with 1M context window, updated to the DeepSeek-V4-Pro-0813 build served by DeepSeek's direct API. Supports tool calling, extended thinking, and high-quality generation.

Google's most capable model for complex reasoning tasks. Features a 1M token context window and strong performance across benchmarks.

Google's Flash model for agentic and multimodal work: thinking, search grounding, code execution, and computer use over a 1M-token context.

Google's lightweight 4B parameter Gemma 3 model. Efficient for simple tasks with vision capabilities.

Compact per-layer-embedding variant of Gemma 4 from Google DeepMind. Text-only input, text output, with toggleable reasoning.

ZhipuAI's efficient vision-language model with 128K context window. Supports image inputs and thinking.

ZhipuAI's vision-language model with 128K context window. Supports image inputs and thinking.

OpenAI's improved version of GPT-4, featuring enhanced reasoning capabilities, better contextual understanding, and increased accuracy across a wide range of tasks.

OpenAI's most powerful GPT-5.4 variant with a 1M+ context window, designed for the most demanding reasoning, coding, and agentic tasks with maximum capability.

OpenAI's safety-focused open-weight model with 120B parameters. Designed for content moderation and safe AI applications.

xAI's Grok 4.20 reasoning model with 2M token context window. Supports reasoning, function calling, structured outputs, and vision.

xAI's multi-agent model with 2M token context window. Optimized for multi-agent workflows with reasoning capabilities. No client-side or custom tools: Client-side tools (function calling) and custom tools are not currently supported by the multi-agent model variant.

xAI's Grok 4.20 non-reasoning model with 2M token context window. Optimized for speed without reasoning overhead, supports function calling, structured outputs, and vision.

xAI's fast coding model trained specifically for agentic coding. Supports reasoning, function calling, structured outputs, and image input. 256K context window.

IBM's ultra-efficient micro model. Small but mighty - perfect for simple tasks requiring minimal latency and cost.

AI21's Jamba 1.5 Large model with hybrid SSM-Transformer architecture. Excels at long-context understanding and generation.

AI21's efficient Jamba 1.5 Mini model with hybrid SSM-Transformer architecture. Fast and cost-effective for everyday tasks.

Moonshot AI's trillion-parameter MoE reasoning model (32B activated). Excels at multi-step reasoning with 200+ sequential tool calls. Supports function calling and extended thinking.

Meta's original Llama 3 70B model optimized for dialogue. Strong general-purpose performance across a wide range of tasks.

Meta's efficient Llama 3 8B model optimized for dialogue. Fast inference suitable for lightweight tasks.

Meta's compact and efficient 1 billion parameter model designed for on-device and edge deployment. Optimized for instruction following with strong performance despite its small size, ideal for resource-constrained environments.

Meta's compact 3B parameter model optimized for edge deployment and multilingual dialogue. Great balance of speed and capability.

Mistral's reasoning-focused small model with vision capabilities. Optimized for step-by-step reasoning tasks.

MiniMax's flagship agentic language model with 200K context window. Supports function calling and reasoning.

MiniMax's lightweight MoE model optimized for coding, agentic workflows, and modern application development. 10B activated parameters with strong multilingual code generation.

MiniMax M2.1 optimized for speed at ~100 tokens/second with 200K context window.

MiniMax M2.5 optimized for speed at ~100 tokens/second with 200K context window.

MiniMax M2.7 optimized for speed with 200K context window and 131K max output.

Mistral's mid-range 14B parameter model with vision support. Enhanced reasoning over smaller Ministrals.

Mistral's ultra-efficient 3B parameter model with vision support. Designed for edge and low-latency applications.

Mistral's efficient 8B parameter model with vision support. Good balance of capability and speed for moderate tasks.

Mistral's flagship 675B parameter model with state-of-the-art reasoning, coding, and multilingual capabilities with vision support.

Mistral's frontier-class multimodal model released May 2025.

Multimodal model from Mistral, released August 2025. Improved tone and performance.

Mistral's efficient 24B model optimized for simple tasks with low latency and 128K context. Great for classification, customer support, text generation, and multimodal tasks.

Meta's Muse Spark reasoning model with a 1M-token context window. Supports image input, tool calling, web search, and structured output.

Meta's updated Muse Spark reasoning checkpoint with a 1M-token context window. Supports image input, tool calling, web search, and structured output.

Meta's Muse Spark 1.2 checkpoint on the contributor tier: heavily discounted pricing in exchange for permission to train future Meta models on your prompts and completions.

Meta's latest Muse Spark checkpoint, tuned for agentic multi-step tool workflows and improved coding over 1.2, with a 1M-token context window. Supports image and PDF input, tool calling, web search, and structured output.

Meta's Muse Spark 1.3 checkpoint on the contributor tier: heavily discounted pricing in exchange for permission to train future Meta models on your prompts and completions.

NVIDIA's hybrid MoE model with leading accuracy for multi-agent applications and specialized agentic AI systems.

NVIDIA's efficient hybrid MoE model combining Mamba-2 and Transformer layers, with 30B total and 3.5B active parameters. Reasoning can be toggled on or off, and it handles English, code and several European and Asian languages.

NVIDIA's compact multimodal Nemotron 3 Nano with 30B total / 3B active MoE parameters and built-in reasoning. Strong at agentic and analytical tasks.

NVIDIA's frontier-scale hybrid Latent-MoE model (550B total, 55B active) for the most demanding agentic, reasoning, and long-context workloads across code, math, and science. Trained with an NVFP4 recipe.

NVIDIA's hybrid Mamba-Transformer MoE model (31B total, 3B active) with a multi-token prediction speculative decoding head for low-latency serving. Tuned for high-throughput reasoning and agentic workloads.

Amazon's second-generation lightweight multimodal model. Processes text and images with improved performance over Nova Lite.

Amazon's multimodal model for processing images, video, and text. Can analyze multiple images with 300K context.

Amazon's highly capable multimodal model balancing accuracy, speed, and cost. Processes text and images with 300K context.

OpenAI's reasoning model designed to think before responding. Uses chain-of-thought reasoning for complex tasks in science, coding, and math.

OpenAI's reasoning model that excels at math, science, coding, and visual reasoning tasks. Uses chain-of-thought reasoning before responding.

OpenAI's small reasoning model, optimized for cost-efficient chain-of-thought reasoning in coding, math, and science tasks.

OpenAI's latest small o-series reasoning model, optimized for fast, effective reasoning across coding, math, and visual tasks.

Writer's enterprise-grade model optimized for business content generation, analysis, and transformation.

Writer's most capable enterprise model with enhanced reasoning, analysis, and content generation capabilities.

Alibaba's fastest, most cost-efficient model with a 1M-token context window. Hybrid thinking model supporting text and PDF inputs and structured output.

Alibaba's balanced mid-tier model with a 1M-token context window. Hybrid thinking model supporting text and PDF inputs and structured output.

Alibaba's Qwen3 235B mixture-of-experts model (22B active, 128 experts) with a native 262K context. Operates in non-thinking mode only and emits no reasoning blocks.

Alibaba's Qwen3 model with groundbreaking advancements in reasoning and instruction-following. Excellent for complex tasks and coding.

Alibaba's open-weights Qwen3 coding model with 480B total / 35B active parameters using mixture-of-experts architecture.

Alibaba's Qwen3 Coder Flash model with a 1M-token context window. Inherits the coding agent capabilities of Qwen3 Coder Plus, tuned for repository-level understanding and multi-turn tool interaction, supporting text and PDF inputs.

Alibaba's Qwen3 Coder Plus model with a 1M-token context window. Hybrid thinking model tuned for coding, supporting text and PDF inputs.

Alibaba's flagship Qwen3 Max model with a 256K-token context window. Non-thinking instruct model tuned for agent programming and tool invocation, supporting structured output and web search.

Alibaba's flagship Qwen3 Max with a 256K-token context window, tuned for extended reasoning on hard math, coding and agentic tasks. Text-only, with tool invocation and JSON object output.

Alibaba's small-scale Qwen3 vision-language model with a 256K-token context window. Integrates thinking and non-thinking modes for fast image and document understanding, and supports structured output.

Alibaba's balanced Qwen3 vision-language model with a 256K-token context window. Supports text, image, and PDF inputs and structured output.

Alibaba's Qwen3.5 Flash model with a 1M-token context window. Hybrid thinking model supporting text and image inputs, tool calling, structured output, and web search.

Alibaba's Qwen3.5 Plus model with a 1M-token context window. Hybrid thinking model supporting text and image inputs, tool calling, structured output, and web search.

Alibaba's Qwen3.8 27B dense model with reasoning, vision, and function calling on a 256K context.

Alibaba's QwQ reasoning model optimized for analytical tasks. Strong performance on math, coding, and logical reasoning benchmarks.

OpenAI's most affordable and fastest GPT-6 model with a 1M+ context window, optimized for high-volume, latency-sensitive applications at minimal cost.

OpenAI's GPT-6 Sol is the balanced frontier model of the GPT-6 family: strong reasoning, coding, and agentic tool use at mid-tier cost, with a 1M+ context window, reasoning effort from none through max, web search, and native image and PDF input.

xAI's flagship model for coding, agentic tool calling, and knowledge work, with a 500K token context window and minimal hallucinations. Supports configurable reasoning (low through xhigh), function calling, structured outputs, web search, and image input.

Alibaba's Qwen3.8 Flash native multimodal model with a 1M-token context window. Combines powerful reasoning and generation with remarkable speed, shining in coding assistance, agentic workflows, and visual understanding.