Models

Embed 5 Fast is the lower-latency, cost-efficient tier of Cohere's Embed 5 family of multimodal embedding models (about 1B parameters). It embeds text, images, and interleaved text and image content into a single vector space and is suited for high-volume query embedding, agent loops, and multilingual semantic search across 100+ languages. It shares its embedding space with Embed 5 Pro, so documents indexed with Pro can be queried with Fast. Output dimensions are selectable from 256 to 2048, and input_type defaults to search_document when not set.

Embed 5 Pro is the higher-accuracy tier of Cohere's Embed 5 family of multimodal embedding models (about 2.4B parameters). It embeds text, images, and interleaved text and image content into a single vector space and is suited for retrieval and RAG over long or domain-specific documents, visual and document search, and multilingual semantic search across 100+ languages. It shares its embedding space with Embed 5 Fast, so a common pattern is to index a corpus with Pro and embed queries with Fast. Output dimensions are selectable from 256 to 2048, and input_type defaults to search_document when not set.

Rerank 5 Fast is the low-latency tier of Cohere's Rerank 5 family of reranking models. It scores and reorders candidate documents against a query and is suited for high-throughput retrieval such as high-volume search, agent tool loops, and first-pass filtering before a more expensive step. Relevance scores are calibrated to a 0-1 range so a fixed threshold behaves consistently across queries. Rerank 5 Pro is the higher-accuracy sibling for complex queries and long or domain-specific documents.

Rerank 5 Pro is the higher-accuracy tier of Cohere's Rerank 5 family of reranking models. It scores and reorders candidate documents against a query and is suited for retrieval pipelines where ranking quality matters most: complex or reasoning-heavy queries, long or domain-specific documents, and RAG systems that feed a small number of passages to a generator. Relevance scores are calibrated to a 0-1 range so a fixed threshold behaves consistently across queries. Rerank 5 Fast is the lower-latency sibling for high-volume search and agent loops.

Clairic is Cloudflare's System One-style extraction model served on Workers AI. Given a text state and optional images plus a schema of typed fields, it extracts, computes, estimates, and forecasts strongly typed values (text spans, numbers, dates, times, durations, lists, estimates, forecasts) in a single pass, returning a confidence score and null probability for every field instead of generating tokens. Served through the Extractions API (/api/alpha/extractions); not available on /api/alpha/decisions.

Parakeet Ultra is Moondream's post-trained version of NVIDIA's Parakeet TDT 0.6B v3 speech-to-text model, with the same architecture, tokenizer, and 0.6B parameters. It lowers word error rate on the Open ASR Leaderboard English sets (5.80% vs 6.26%) and the 25-language FLEURS set (9.55% vs 11.62%), with larger gains on noisy and long-form audio. Supports 25 European languages and returns transcribed text with segment and word timestamps.

Crouton Alpha is a lightweight model built for scalable coding, agentic, and multimodal workflows. It's ideal for executing agentic sub-tasks, tool use, and document analysis in high-volume, repeated work. Developers can combine it with frontier reasoning models to execute the bounded steps that turn plans and information into action. Crouton Alpha is a stealth model. It is developed and operated by a third-party provider who has chosen to remain anonymous during this preview. OpenRouter routes requests to it and is not its developer, owner, or provider. Prompts and completions may be retained by the provider but are not used for training; all other use is governed by the Stealth Model Terms.

GLM-5.3-FlashX-Legal (GLM-Legal) is Z.ai's legal-domain variant of GLM-5.3-FlashX, a native multimodal model that accepts text, image, and video input and produces text. It adds legal-domain mid-training and post-training with reinforcement learning in task environments modeled on real legal workflows, targeting document understanding, legal reasoning, and multi-step task execution such as legal research, contract review, cross-document consistency checks, and drafting or revising legal documents. It keeps the GLM-5.3-FlashX 1M-token context window, mandatory thinking mode, function calling, structured output, and streaming, and is designed to run in agent setups connected to legal data and document-editing tools (MCP services).

TBC mischief 1 is the world's first neuron-derived video model from The Biological Computing Co. It supports image to video and is a fast and affordable way to bring ideas to life. Ideal for ideation, previsualization, moodboarding and variant testing.

Joey 1 is a 2B decision model from Base Labs, Baseten's research team, built on Qwen3.5-2B. Given a state (text or JSON) and typed questions (yes/no, multiple choice, or ordinal score), it returns a calibrated probability for every option in a single forward pass, with no text generation. It is TypeSafe System One compatible and suits high-volume routing, classification, and guardrail checks where latency and cost matter most. Its larger siblings are Elephant 1 and Mammoth 1.

Riverflow 2.6 Pro is the flagship model of Sourceful's Riverflow 2.6 lineup, a unified text-to-image and image-to-image family that treats generation as a production workflow: an integrated reasoning model plans multi-step edits and judges candidates before accepting a result, combining that reasoning with a mix of closed and open image diffusion models for accuracy and steerability. Reasoning effort is controllable via the reasoning parameter (low/medium/high/xhigh); higher levels do more editing passes and apply a stricter internal judge. Riverflow 2.6 adds custom output dimensions: pass size as WIDTHxHEIGHT (up to 16,384 px per edge, 16.8 megapixels total, 16:1 aspect ratio) instead of a preset, or keep using 1K/2K/4K with an aspect ratio (now including 5:4 and 4:5). It accepts up to 10 input images for editing. Pricing is dynamic: cost is finalized per job at completion based on billable processing, so it scales with reasoning effort, output size, and editing complexity rather than a fixed per-image rate. Additional features (via image_config): - Custom font rendering via font_inputs (max 2) to match brand lettering, spacing, and weight - Custom scoring via scoring_prompt and scoring_rubric, so the reasoning model evaluates and steers each candidate against the criteria you care about - Background control via background_mode (original, transparent, solid) and background_hex_color See the image generation docs for details: https://openrouter.ai/docs/features/multimodal/image-generation Note: Sourceful imposes a 4.5MB request size limit, therefore it is highly recommended to pass image URLs instead of Base64 data.

Nemotron 3.5 Transcribe is an open-weight speech recognition model from NVIDIA, a 30B mixture-of-experts with about 4B active parameters. It transcribes recordings of up to one hour in 40 languages, with speaker attribution, speech translation, and prompt-guided vocabulary and style.

AssemblyAI TTS 1 is a text-to-speech model from AssemblyAI. It is suited for voice agents, narration, and multilingual applications, and draws on a public library of several thousand voices across languages. Voices are selected by slug or ID, and the spoken language follows the chosen voice or can be set explicitly through the language provider option.

Particle-1 is a token-efficient agentic coding model from Particle AI, post-trained on DeepSeek V4.1 Flash. It is suited for software engineering work such as terminal workflows, codebase changes, and data engineering tasks, and is trained to complete them with substantially fewer reasoning and tool-call tokens while keeping accuracy close to its base model.

d1.1 is Liquid AI's structured decision model, served as a System One endpoint and the successor to d1. Send a state along with typed questions, and it returns a choice, a score, or a yes/no answer, each with a probability taken directly from the model rather than written out as text. It uses the same /v1/systemone schema as other OpenRouter Decisions models, so it suits routing, classification, and policy checks that need a fast, scored answer instead of prose.

This model always redirects to the latest model in the d1-ultra family.

This model always redirects to the latest model in the d1 family.

Amazon Titan Multimodal Embeddings G1 embeds text, images, or text and image together into a shared vector space for multimodal search and recommendation. Images are sent as base64 (up to 25 MB), text input is limited to 128 tokens, and vectors are 1,024 dimensions by default with 384 and 256 selectable via the dimensions parameter.

Amazon Titan Embeddings G1 - Text is the first-generation Titan text embedding model. It accepts up to 8,192 tokens per input and returns fixed 1,536-dimension vectors; the dimensions parameter is not supported.

Amazon Titan Text Embeddings V2 is a text embedding model for retrieval, search, and clustering. It accepts up to 8,192 tokens per input and returns normalized vectors of 1,024 dimensions by default, with 512 and 256 also selectable via the dimensions parameter.