Models

MiniMax H3 Image Highspeed is the faster version of MiniMax's H3 Image model. It takes the same inputs and limits as MiniMax H3 Image: it creates images from text prompts and edits or composes images from 1 to 16 reference images, with output up to 20 megapixels and aspect ratios from 1:3 to 3:1.

MiniMax H3 Image is MiniMax's image generation and editing model. It creates images from text prompts and edits or composes images from 1 to 16 reference images, with output up to 20 megapixels and aspect ratios from 1:3 to 3:1.

ERNIE 5.5 Flash Preview is the latest foundation model in the ERNIE family, with comprehensive improvements in coding, agentic capabilities, knowledge work, and deep research. It has 280 billion total parameters, activates 14 billion parameters per token, and supports a context window of 1 million tokens. Its new architecture combines strong model capabilities with efficient inference, delivering generation speeds of more than 200 tokens per second.

ERNIE-Light-DT-1.0 is based on open-source GLM models and includes research findings from Baidu Search's Lightning large model. It features Baidu Sparse Attention (BSA), a unique sparse attention mechanism developed by Baidu ERNIE. By using feedback-driven history compression, it improves the storage and indexing of historical information. This significantly reduces KV cache storage and computational costs while maintaining model performance. As a result, it allows for longer context windows at a lower cost.

Grok STT 2.0 is a speech-to-text model from SpaceXAI and the successor to Grok STT 1.0. It is suited for transcription workloads that need word-level timestamps, optional speaker diarization, and multichannel audio.

Gemini 4 Carbon is a frontier Gemini 4 model from Google for coding, agentic and complex reasoning work. It takes text, image, audio, video and file input, supports long decode of up to 1M output tokens through continuation, and controls thinking depth with low, medium and high reasoning effort.

Gemini 4 Carbon is a frontier Gemini 4 model from Google for coding, agentic and complex reasoning work. It takes text, image, audio, video and file input, supports long decode of up to 1M output tokens through continuation, and controls thinking depth with low, medium and high reasoning effort.

Nex-N2.6 is an agentic model from Nex AGI, built on GLM-5.3-Flash with about 321B parameters. It is suited for computer use, coding, and long-horizon tasks that involve tools. Compared with Nex-N2.5-Pro, it reduces repetitive reasoning and follows the tools available in the current environment more closely.

Drex Max Finance is a finance-tuned version of Nace.AI's Drex Max, a 27B multimodal decision model. Given a state as text, JSON, or images such as invoices, receipts, and statements, plus typed questions (yes/no, multiple choice, or ordinal score), it returns a calibrated probability for every option in a single forward pass, without generating text. It is built for banking, accounting, payments, insurance, credit, compliance, filings, and markets workloads, and supports contexts up to 262K tokens.

Inkling 2 Small is a multimodal mixture-of-experts reasoning model from Thinking Machines Lab, with 17B active parameters out of 594B total. It is the first model in the second Inkling generation. It was trained with reinforcement learning on long-horizon coding tasks across many agent harnesses, and is suited for agentic coding, autonomous research, agentic knowledge work and retrieval-augmented generation, with English and general multilingual support. It accepts text and image input (no audio), supports tool calling and variable reasoning effort, and has a 1M-token context window.

Claude Fable 5.2 is Anthropic's most capable model for the most demanding technical work, with gains over Fable 5.1 in coding, reasoning, and vision. In prerelease testing it completed otherwise-infeasible software migrations, designed working electronic devices from scratch, and made progress on open problems in physics and mathematics. It communicates more clearly than earlier Fable models and generates output up to 30% faster than Fable 5.1. Adaptive extended thinking is always on, with reasoning effort levels from low to max (default high) and preserved thinking across turns. Anthropic recommends Opus 5.5 as the daily driver for most coding and knowledge work, with Fable 5.2 for the most ambitious, open-ended problems.

Claude Fable 5.2 is Anthropic's most capable model for the most demanding technical work, with gains over Fable 5.1 in coding, reasoning, and vision. In prerelease testing it completed otherwise-infeasible software migrations, designed working electronic devices from scratch, and made progress on open problems in physics and mathematics. It communicates more clearly than earlier Fable models and generates output up to 30% faster than Fable 5.1. Adaptive extended thinking is always on, with reasoning effort levels from low to max (default high) and preserved thinking across turns. Anthropic recommends Opus 5.5 as the daily driver for most coding and knowledge work, with Fable 5.2 for the most ambitious, open-ended problems.

Muse Ember is the smallest and most efficient model from Meta that’s designed for complex reasoning execution and the work agents repeat all day: tool calls, coding agent subtasks, routing, extraction, and long document parsing. Muse Ember accepts text, images, video, audio, and PDF documents and returns text, with a 1M-token context window.

A compact 1.7B-parameter text model with a 256K-token context, post-trained for agents, tool calling, and structured outputs. It reasons before it answers. Prompts and outputs may be retained and used to train Liquid models.

A compact 1.7B-parameter text model with a 256K-token context, post-trained for agents, tool calling, and structured outputs. It reasons before it answers. Prompts and outputs may be retained and used to train Liquid models.

Decisions Fast is Parallel's web-grounded Decisions model. Instead of generating text, it reads the content passed as state (text or JSON), researches the live web, and returns typed answers to named questions: the probability of yes for a yes/no question (noul), a choice with per-option probabilities (choice), or a level on an ordered rubric (score), with per-question refusals and request-level citations to the sources it used. It suits fact checks, entity screening, and classification that depend on current information rather than only on what the caller supplies. A request can carry up to 100 questions about the same state, with up to 32,000 tokens per question including the shared state.

Gemini Omni 1.2 Flash is Google's native multimodal video generation and editing model, the successor to Gemini Omni 1.1 Flash. It generates clips up to 15 seconds long with native audio in 16:9 or 9:16, from 360p drafts up to 4K, and supports text-to-video, image-to-video, first and last frame interpolation, subject and video references with improved reference management, and editing or extending an uploaded video. Multi-turn editing and extension of a prior generation is supported through previous_job_id. Note: video generation is only supported through the Video Generation API. See the guide: https://openrouter.ai/docs/guides/overview/multimodal/video-generation

Mi:dm K 3.0 Pro is a Mixture-of-Experts reasoning model from KT, with 120B total and 12B active parameters. It uses a hybrid architecture that mixes Mamba, MoE and attention layers.

Auto is OpenRouter’s flagship model router, built for general-purpose usage. It adapts to each request, choosing cost-optimal models and adjusting reasoning effort. It balances task difficulty, benchmark results, live prices, and prompt caching. Use it for apps, chat, and general agents. For agent sessions, it follows progress across turns, keeps the same model while work is going well, and adjusts reasoning effort or switches models when the task gets stuck. It respects all your governance settings, including guardrails, ZDR, and budgets.

Auto Code is OpenRouter’s router built specifically for coding agents and harnesses. It uses the same routing engine as the Auto Router, with every request treated as part of an agent session and model choices guided by coding-task pass rate, completion time, and cost per solved task. It follows edits, test results, and recurring errors across turns, keeping the model while work is going well and trying more reasoning effort before switching when trouble persists. Its default coding tier starts from stronger models and weighs failed code changes more heavily. Your model and provider restrictions still apply.