LLM API pricing, tracked from launch.

Ling-2.6-flash is a 104B parameter model from inclusionAI with 7.4B active parameters, designed for fast inference and efficient token usage in agent applications.

Llama Guard 3 8B is Meta's content safety classifier designed to detect and evaluate harmful content in user prompts and model outputs.

A 12B parameter model with a 128k token context window built by Mistral in collaboration with NVIDIA, designed to balance efficiency with capability across multiple languages.

An 8B parameter model based on Llama 3, created through merging multiple models to balance creative and logical capabilities.

One of the highest performing and most popular fine-tunes of Llama 2 13B, with rich descriptions and roleplay. #merge

Mistral Small 3 is a 24-billion-parameter language model designed for low-latency inference across common AI tasks.

Meta's 8-billion-parameter instruction-tuned language model from the Llama 3.1 family, with a 131k token context window.

Granite 4.1 8B is an 8-billion-parameter decoder-only language model from IBM with a 131K-token context window, designed for enterprise applications.

A 7-billion-parameter multimodal model that processes images, videos, and text to generate text outputs.

The smallest model in the Ministral 3 family, Ministral 3 3B is a powerful, efficient tiny language model with vision capabilities.

Gemma 3 4B is Google's 4-billion-parameter open-weight text model with a 131K token context window.

Qwen2.5 7B Instruct is a 7-billion parameter open-weight language model from Alibaba optimized for instruction-following tasks with a 32K token context window.

Nex-N2-Mini is an open-source mixture-of-experts model from Nex AGI designed for agentic tasks, accepting text and image inputs with a 262K token context window.

A 3B parameter model from IBM's Granite 4 family, designed to handle long context windows efficiently.

A 24-billion parameter mixture-of-experts model from LiquidAI that activates only 2 billion parameters per token for efficient inference.

A 4-billion-parameter lightweight open-weight text model from Google with a 32K token context window.

Laguna XS 2.1 is a lightweight coding model from Poolside with a 262k token context window.

Qwen3.7 Flash is a lightweight vision-language model from Alibaba designed for fast inference on multimodal tasks including text and image inputs.

An open-weight 20-billion-parameter model from OpenAI, released after the larger gpt-oss-120b.

A 14-billion-parameter model from Microsoft Research designed for complex reasoning with efficient performance on limited hardware.

A lightweight text-only model from Amazon's Nova family designed for low-latency, cost-efficient inference.

Meta's 8-billion-parameter instruction-tuned language model from the Llama 3 series, with an 8,192-token context window.

Qwen3.5-9B is a 9-billion-parameter text model from Alibaba with a 262K token context window.

An 8-billion-parameter dense model trained by EssentialAI for programming, math, and scientific reasoning tasks.

A balanced model in the Ministral 3 family, Ministral 3 8B is a powerful, efficient tiny language model with vision capabilities.

A 26B sparse mixture-of-experts language model with 3B parameters active per token and 131k token context window, designed for efficient reasoning over extended documents.

Google's 12-billion-parameter open-weight text model with a 131K token context window, released in 2025 between the Gemma 3 4B and Gemma 3 27B.

Command R7B (12-2024) is Cohere's 7B-parameter language model with a 128K context window, optimized for retrieval-augmented generation and agentic workflows.

An open-weight 120B parameter model from OpenAI released before the gpt-oss-20b, larger and less compute-efficient than its successor.

Llama Guard 4 12B is Meta's content safety classifier, built to detect and evaluate harmful content in both user prompts and model outputs.

A 30.5B-parameter mixture-of-experts model from Alibaba's Qwen3 that activates 3.3B parameters per token, designed for instruction following and multilingual tasks.

A 30-billion-parameter mixture-of-experts model from NVIDIA designed for efficient inference and specialized task performance.

A 14-billion parameter language model from Mistral offering frontier-level performance in a compact form factor.

A 7B multimodal vision-language model from ByteDance designed to understand and interact with graphical user interfaces across desktop, web, mobile, and game environments.

Reka Flash 3 is a 21-billion-parameter instruction-tuned model from Reka designed for general-purpose tasks including chat, coding, and function calling.

Laguna XS.2 is a lightweight coding model from Poolside with a 262k token context window.

Laguna S 2.1 is a 118B parameter coding agent model from Poolside with a 1M token context window, designed for extended code understanding and generation tasks.

A 1-billion-parameter instruction-tuned language model from Meta, released in September 2024 as part of the Llama 3.2 family.

Hy3 preview is an early-release version of Tencent's mixture-of-experts reasoning model, giving early access to features and capability development before the full Hy3 release.

Amazon's lightweight multimodal model that processes text, images, and video inputs with a 300K token context window.

Qwen3.5-Flash is a lightweight text and vision model from Alibaba with a 1 million token context window, positioned as a faster alternative to the mid-tier Qwen3.5 Plus and Qwen3.6 Plus.

A 30.5B parameter mixture-of-experts model from Alibaba with 128 experts (8 active per token), designed for code generation and repository-scale code understanding.

Ultra-cheap frontier-class model (284B total, 13B active MoE). Launched alongside V4 Pro.

Xiaomi's native omnimodal model that processes text, images, and video in a single forward pass.

Qwen3-32B is a 32.8B parameter language model from Alibaba that supports both standard generation and extended thinking mode for complex reasoning tasks.

A 32B parameter model distilled from DeepSeek R1 using Qwen 2.5 as the base architecture, trained to replicate reasoning capabilities of the larger R1 model.

Meta's smaller open-weight MoE model (17B active / 109B total). Supports up to 10M context on some providers.

Fast, cost-effective model for high-volume and latency-sensitive workloads. Launched at $0.15/$0.60, later cut to $0.10/$0.30.

StepFun's sparse Mixture of Experts model that activates 11B of 196B parameters per token, providing a balance between capability and efficiency.

A fast multimodal model from ByteDance Seed that handles text and image inputs with a 256k token context window.

MiMo-V2-Flash is a Mixture-of-Experts language model from Xiaomi with 309B total parameters and 15B active parameters, supporting a 262K token context window.

A 24B multimodal model from Mistral that processes both text and audio inputs, with a 32K token context window.

gpt-oss-safeguard-20b is OpenAI's open-source safety-focused model with a 131k token context window.

Mistral Small 3.2 24B is a 24-billion parameter model from Mistral optimized for instruction following and function calling, with a 128k token context window.

A 3-billion-parameter instruction-tuned language model from Meta with an 80,000-token context window, released in 2024.

An 11-billion-parameter instruction-tuned vision model from Meta that processes both text and images with a 131k token context window.

Phi-4-mini-instruct is a lightweight open model trained on synthetic and curated public data, designed to handle reasoning-heavy tasks efficiently.

Cheapest OpenAI model — designed for classification, routing, and high-throughput extraction.

Gemma 4 31B is Google's 31-billion-parameter open-weight text model with a 262K token context window.

A 120B-parameter open hybrid MoE model from NVIDIA that activates 12B parameters per token, combining Mamba and Transformer architectures for inference efficiency.

A lightweight model from ByteDance designed for low-latency inference at high concurrency, with 256k context window and multimodal understanding.

GLM 4.7 Flash is Z.ai's optimized inference variant of the GLM 4.7 foundation model, designed for faster response times with a 202K-token context window.

DeepSeek V3.2 is a large language model that prioritizes computational efficiency while maintaining reasoning and tool-use capabilities, with a 128K token context window.

A 49B-parameter model optimized for reasoning and agentic workflows, derived from Llama 3.3 and supporting 131K token context.

Hermes 4 70B is a hybrid reasoning model from Nous Research that can switch between fast and deliberate reasoning modes to balance speed and accuracy on a per-query basis.

GPT-5 Nano is OpenAI's smallest general-purpose model in the GPT-5 family, with a 400k token context window.

Google's lightweight multimodal model with a 1M token context window, released in 2025 as a faster alternative to the Gemini 2.5 Flash line.

Meta's 70-billion-parameter instruction-tuned language model with a 131k token context window, released after Llama 3.1 70B Instruct.

UnslopNemo v4.1 is the latest addition from the creator of Rocinante, designed for adventure writing and role-play scenarios.

Qwen2.5 72B Instruct is a 72-billion-parameter instruction-tuned language model from Alibaba with a 32K token context window.

Meta's 70-billion-parameter instruction-tuned language model with a 131k token context window, released in July 2024.

Laguna M.1 is a mid-sized coding model from Poolside with a 262k token context window.

An experimental large language model from DeepSeek featuring a sparse attention mechanism designed to reduce computational overhead while maintaining reasoning capability.

A 32B multimodal model from Alibaba that processes text, images, and video with a 131k-token context window.

Open-weight reasoning model that stunned the industry at launch. Repriced with V3.2-Exp in September 2025; now routes to V4 Flash thinking mode and is scheduled for deprecation July 2026.

General-purpose chat model. Launched on a promotional $0.14/$0.28 (Dec 2024–Feb 2025), then settled at $0.27/$1.10; received a 50%+ output price cut with the V3.2-Exp update in September 2025.

Gemma 4 26B A4B is Google's 26-billion-parameter open-weight text model with a 262K token context window.

Google's 27-billion-parameter open-weight text model with a 110K token context window.

An 8B multimodal vision-language model from Alibaba that processes text, images, and video with a 131k-token context window.

Qwen3 8B is a 8.2B parameter language model from Alibaba that supports both standard inference and extended thinking mode for complex reasoning tasks.

Olmo 3 32B Think is a 32-billion-parameter reasoning model from AllenAI designed for complex logical inference and extended problem-solving.

Uncensored and creative writing model based on Mistral Small 3.2 24B with good recall, prompt adherence, and intelligence.

A 30B dense model from Alibaba's Qwen3 series with a 40K token context window, designed for reasoning and multilingual tasks.

Rocinante 12B is a 12-billion-parameter model from TheDrummer designed for narrative generation and creative writing with a 32K token context window.

Hy3 is a 295B-parameter mixture-of-experts model from Tencent with 21B active parameters that supports configurable reasoning effort for complex problem-solving.

A multilingual mixture-of-experts model with 235B total parameters and 22B active per forward pass, instruction-tuned for general-purpose text generation.

A 24B parameter model from Mistral with multimodal capabilities and a 128K token context window.

A 13-billion-parameter instruction-tuned language model from Tencent with a 131k-token context window.

Meta's open-weight MoE flagship (17B active / 400B+ total). Hosted pricing varies widely; figures are a representative hosted rate.

A Mixture-of-Experts language model from Upstage with 102B parameters total and 12B active per forward pass, operating on a 128K token context window.

A 30B multimodal vision-language model from Alibaba that processes text, images, and video with a 262k-token context window.

GPT-4o-mini with web search capability, allowing it to retrieve and reference current information from the internet during inference.

A 24B parameter model from Mistral trained for Middle East and South Asia regions with a 32k token context window.

GPT-4o-mini is OpenAI's lightweight multimodal model that processes text and image inputs, released in July 2024 with a 128K token context window.

Cohere's scalable RAG and tool-use workhorse. The August 2024 refresh cut input pricing 70% to $0.15/$0.60.

A code-focused language model from Kwaipilot with a 256k token context window.

A 176-billion-parameter mixture-of-experts model from Microsoft that combines eight 22-billion experts with a large context window for handling extended documents and conversations.

Ring-2.6-1T is a 1 trillion parameter mixture-of-experts model from inclusionAI with 63B active parameters, featuring a 262K token context window optimized for agentic workflows.

Ling-2.6-1T is a trillion-parameter instruct model from inclusionAI designed for fast execution and efficient operation at scale.

Google's 27-billion-parameter open-weight text model with an 8K token context window.

A recreation trial of the original MythoMax-L2-B13 but with updated models. #merge

Hermes 3 70B Instruct is a 70-billion parameter generalist model with a 131K token context window, trained to follow instructions and support agentic workflows.

Meta's 70-billion-parameter instruction-tuned language model released in April 2024, with an 8,192-token context window.

Mercury 2 is a reasoning model from Inception that uses diffusion-based token generation to produce and refine multiple tokens in parallel rather than sequentially.

Euryale L3.3 70B is a model focused on creative roleplay from [Sao10k](https://ko-fi.com/sao10k). It is the successor of [Euryale L3 70B v2.2](/models/sao10k/l3-euryale-70b).

An attempt to recreate Claude-style verbosity, but don't expect the same level of coherence or memory. Meant for use in roleplay/narrative situations.

A reasoning-focused chat model from Alibaba's Qwen3-Next line that outputs structured thinking traces for multi-step problem solving.

Qwen Plus 0728, based on the Qwen3 foundation model, is a 1 million context hybrid reasoning model with a balanced performance, speed, and cost combination.

Qwen-Plus, based on the Qwen2.5 foundation model, is a 131K context model with a balanced performance, speed, and cost combination.

Coder Large is a 32B parameter model from Arcee AI built on Qwen 2.5-Instruct and trained on open-source code repositories and synthetic bug-fix examples to specialize in code generation and analysis.

Skyfall 36B v2 is an enhanced iteration of Mistral Small 2501, specifically fine-tuned for improved creativity, nuanced writing, role-playing, and coherent storytelling.

A 70B parameter model distilled from DeepSeek R1 using Llama 3.3 as a base, designed to replicate advanced reasoning capabilities at reduced computational cost.

An open-source reasoning model from Arcee AI designed for complex problem-solving and agentic tasks, with extended thinking capabilities.

GLM 4.5 Air is Z.ai's lightweight inference variant of the GLM 4.5 foundation model, maintaining a 131K-token context window with optimizations for faster response times.

Euryale L3.1 70B v2.2 is a model focused on creative roleplay from [Sao10k](https://ko-fi.com/sao10k). It is the successor of [Euryale L3 70B v2.1](/models/sao10k/l3-euryale-70b).

Open-weight frontier model. DeepSeek made a 75% promotional discount permanent in May 2026, making it one of the cheapest frontier models.

MiMo-V2.5-Pro is Xiaomi's flagship model with a 1M-token context window designed for complex reasoning and extended reasoning over long inputs.

MiniMax M2.5 is a large language model designed for productivity tasks, trained on complex real-world digital working environments with particular strength in code understanding.

Qwen3 Coder Next is a code-specialized model from Alibaba's Qwen family with a 262K token context window.

GLM 4.6V is Z.ai's vision-language model that processes text, image, and video inputs for multimodal reasoning tasks.

Mistral's cutting-edge language model for coding released end of July 2025. Codestral specializes in low-latency, high-frequency tasks such as fill-in-the-middle (FIM), code correction and test generation. [Blog Post](https://mistral.ai/news/codestral-25-08)

An instruction-tuned 24B model based on Mistral that minimizes refusals on sensitive or controversial topics.

Qwen3-14B is a 14.8B parameter language model from Alibaba's Qwen3 series with a 40K token context window, designed to handle complex reasoning and dialogue tasks with a configurable thinking mode.

DeepSeek V3.1 is a 671-billion-parameter hybrid reasoning model with 37 billion active parameters that supports both thinking and non-thinking inference modes via prompt templates.

Qwen3 Coder Flash is a lightweight code generation model from Alibaba designed for fast coding tasks with a 1M-token context window.

Qwen3.6 35B A3B is a 35-billion-parameter multimodal model from Alibaba combining linear attention with sparse mixture-of-experts routing for efficient processing of text, images, and video.

MiniMax M2.7 is a compact language model designed for agentic tasks and real-world automation with a 196k token context window.

A 35-billion-parameter multimodal model from Alibaba combining linear attention with sparse mixture-of-experts routing for efficient processing of text and images across a 262K token context window.

DeepSeek V3.1 Terminus is an updated version of DeepSeek V3.1 that refines language consistency and agent task performance.

A Mixture-of-Experts code generation model from Alibaba optimized for agentic coding tasks and long-context reasoning.

Qwen2.5-VL is proficient in recognizing common objects such as flowers, birds, fish, and insects. It is also highly capable of analyzing texts, charts, icons, graphics, and layouts within images.

Perplexity's Sonar is a lightweight model designed for question-answering tasks with citation support and customizable source integration.

A code-focused LLM from Alibaba designed for generating and reasoning about code across multiple programming languages.

Hermes 3 405B Instruct is a large generalist language model from Nous with a 131k token context window, designed for multi-turn conversations and agentic tasks.

Nex-N2-Pro is a mixture-of-experts model from Nex AGI that processes text and image inputs with a 262K token context window.

MiniMax M2 is a 10B activated parameter model with a 196K token context window, designed for coding and agentic workflows.

A 106B-parameter mixture-of-experts model from Prime Intellect with 12B active parameters, post-trained with supervised fine-tuning and reinforcement learning on a 131k-token context window.

Qwen3-Next-80B-A3B-Instruct is an instruction-tuned 80B parameter model from Alibaba's Qwen3 series with a 262k token context window, designed for chat and task completion without extended reasoning traces.

MiniMax-01 is a multimodal model combining text generation and image understanding with 456 billion parameters and a 1 million token context window.

DeepSeek V3 is a 685-billion-parameter mixture-of-experts model designed for multi-turn conversations and reasoning tasks.

A lightweight text and vision model from Alibaba's Qwen family with a 1 million token context window.

StepFun's multimodal Mixture-of-Experts model that processes text, images, and video with a 196B parameter backbone while activating approximately 11B parameters per inference.

MiniMax M3 is a multimodal foundation model that processes text, image, and video inputs to produce text output, with a 524K token context window.

A code-focused language model from Kwaipilot with a 256K token context window.

MiniMax M2-her is a dialogue-focused model designed for character-driven conversations and roleplay scenarios with a 65K token context window.

MiniMax M2.1 is a 10-billion-parameter lightweight language model designed for coding, agentic workflows, and application development.

Morph's fastest model for applying code edits and transformations, optimized to process code modifications at high throughput.

Virtuoso Large is a 72B general-purpose language model from Arcee AI with a 131k token context window.

A sparse mixture-of-experts model from Meituan with 48B active parameters designed to handle very long contexts up to 1M tokens.

A lightweight variant of GPT-5.4 with a 400k token context window, positioned as a smaller, faster alternative within OpenAI's GPT-5.4 series.

Cogito v2.1 671B is a 671-billion parameter mixture-of-experts model trained with self-play reinforcement learning, offering a 128k token context window.

A specialized code-patching model that applies AI-suggested edits directly into source files from multiple LLM providers.

A multimodal mixture-of-experts model from Baidu with 424B total parameters and 47B active per token, trained on both text and image data.

Qwen3.7 Plus is Alibaba's multimodal model supporting text, images, and video inputs with a 1 million token context window.

An 8B multimodal vision-language model from Alibaba that processes text, images, and video with extended reasoning capabilities for structured problem-solving.

Aion-1.0-Mini is a smaller parameter variant of Aion-1.0 from AionLabs, sharing the same 131k-token context window as its full-size sibling.

Aion-3.0-Mini is a multi-model roleplaying and storytelling system from AionLabs built on the DeepSeek family, using collaborative generation from specialized models.

Alibaba's previous flagship. Received a 50% price cut in November 2025 as China's AI price war intensified.

Mistral's Apache-2.0 open-weight frontier model (675B total / 41B active MoE). 75% cheaper than Large 2 at $0.50/$1.50.

A vision-language model from Perceptron designed to process images and videos alongside natural language queries for visual understanding tasks.

Google's lightweight text model designed for efficient execution in multi-agent and agentic workflows with a 1M token context window.

OpenAI's lightweight language model optimized for chat and code tasks with a 16K token context window.

Qwen3.5-27B is a 27-billion-parameter text model from Alibaba with a 262K token context window.

Qwen3.5 Plus is Alibaba's mid-tier language model from February 2026 with a 1 million token context window.

A 30B multimodal vision-language model from Alibaba that processes text, images, and video with extended reasoning capabilities and a 131k-token context window.

A 30B parameter mixture-of-experts model from Alibaba designed for reasoning tasks with extended thinking capabilities and internal reasoning traces.

Fast, cost-effective model with 1M context for high-volume workloads.