Models

rerank-3-lite is a reranker optimized for both latency and quality and a drop-in upgrade to rerank-2.5-lite, improving on it by 0.94% NDCG@10 on average across domain evaluations and by 1.86% on long-document evaluations, with code retrieval gains of 2.77% atop voyage-3-large and 2.59% atop voyage-4-large. It matches the retrieval quality of rerank-2.5, and across 93 retrieval datasets it outperforms Cohere Rerank v4.0 Pro by 1.44% and Qwen3-Reranker-8B by 2.61%. The model supports a combined context length of 32K tokens per query-document pair, including up to 8K tokens for the query, enabling more accurate retrieval over longer documents. Additionally, rerank-3-lite supports instruction following, allowing users to guide relevance scoring through natural language prompts. Learn more about rerank-3-lite here: blog.voyageai.com/2026/09/30/rerank-3

rerank-3 is a reranker optimized for quality and a drop-in upgrade to rerank-2.5, improving on it by 0.80% NDCG@10 on average across domain evaluations and by 3.35% on long-document evaluations, with code retrieval gains of 2.01% atop voyage-3-large and 1.96% atop voyage-4-large. Across 93 retrieval datasets it outperforms Cohere Rerank v4.0 Pro by 2.14% and Qwen3-Reranker-8B by 3.31%. The model supports a combined context length of 32K tokens per query-document pair, including up to 8K tokens for the query, enabling more accurate retrieval over longer documents. Additionally, rerank-3 supports instruction following, allowing users to guide relevance scoring through natural language prompts. Learn more about rerank-3 here: blog.voyageai.com/2026/09/30/rerank-3

GPT-6.1 Sol Pro is the same underlying model as GPT-6.1 Sol, served with reasoning.mode set to pro for higher-quality responses on complex tasks. Cost note: pro mode spends far more reasoning tokens per request, so a typical request costs several times more than the same request on GPT-6.1 Sol and takes much longer to complete. It is intended for hard, high-stakes problems where the extra accuracy justifies the cost. For everyday coding, agentic, and chat workloads, use GPT-6.1 Sol instead. Learn more in OpenAI's docs: https://developers.openai.com/api/docs/guides/reasoning#reasoning-mode

GPT-6.1 Sol Pro is the same underlying model as GPT-6.1 Sol, served with reasoning.mode set to pro for higher-quality responses on complex tasks. Cost note: pro mode spends far more reasoning tokens per request, so a typical request costs several times more than the same request on GPT-6.1 Sol and takes much longer to complete. It is intended for hard, high-stakes problems where the extra accuracy justifies the cost. For everyday coding, agentic, and chat workloads, use GPT-6.1 Sol instead. Learn more in OpenAI's docs: https://developers.openai.com/api/docs/guides/reasoning#reasoning-mode

GPT-6.1 Sol is an upgrade to GPT-6 Sol from OpenAI, positioned below the flagship GPT-6 Astra in the GPT-6 series. It is suited for agentic coding, computer use, document-heavy professional work, and multi-step business workflow automation, and approaches Astra-level results on these tasks at a much lower cost. Compared with GPT-6 Sol, it makes fewer factual errors and is more reliable at respecting explicit restrictions and user intent during agentic tasks.

GPT-6.1 Sol is an upgrade to GPT-6 Sol from OpenAI, positioned below the flagship GPT-6 Astra in the GPT-6 series. It is suited for agentic coding, computer use, document-heavy professional work, and multi-step business workflow automation, and approaches Astra-level results on these tasks at a much lower cost. Compared with GPT-6 Sol, it makes fewer factual errors and is more reliable at respecting explicit restrictions and user intent during agentic tasks.

Claude Sonnet 5.5 is Anthropic's Sonnet-class model for well-scoped everyday work, succeeding Claude Sonnet 5 as a direct upgrade. It is especially strong at building features, fixing bugs, and producing polished documents, slides, and spreadsheets, and it writes and communicates more clearly than its predecessor. Thinking is always on, so effort is the main lever for trading off depth, latency, and cost, and lower effort settings keep it responsive for everyday agentic loops.

Claude Sonnet 5.5 is Anthropic's Sonnet-class model for well-scoped everyday work, succeeding Claude Sonnet 5 as a direct upgrade. It is especially strong at building features, fixing bugs, and producing polished documents, slides, and spreadsheets, and it writes and communicates more clearly than its predecessor. Thinking is always on, so effort is the main lever for trading off depth, latency, and cost, and lower effort settings keep it responsive for everyday agentic loops.

Solar Decide is Upstage's structured decision model, served as a System One endpoint on Solar Mini 4. Send a state along with typed questions, and it returns a choice, a score, or a yes/no answer, each with a calibrated probability taken directly from the model rather than written out as text. Because it generates no prose, each decision takes a single forward pass and output tokens are free. With a 512K context window, an entire document can serve as the state. Solar Decide uses the same /v1/systemone schema as Jev, bringing Solar Mini 4's strong Korean understanding to routing, classification, and policy checks.

Span-01 is a behavior scoring model from Respan. It reads a conversation span and returns, for each plain-language behavior you define, the probability that the behavior is present. It is suited for evaluation, guardrails, and monitoring of LLM and agent outputs at scale. It is the higher-accuracy tier of the family. Span-01 Lite is the free, lighter tier.

Span-01 Lite is the free, lighter tier of Span-01, a behavior scoring model from Respan. It returns, for each plain-language behavior you define, the probability that the behavior is present in a conversation span, and is suited for high-volume evaluation and monitoring where cost matters more than peak accuracy.

Span-01 Lite is the free, lighter tier of Span-01, a behavior scoring model from Respan. It returns, for each plain-language behavior you define, the probability that the behavior is present in a conversation span, and is suited for high-volume evaluation and monitoring where cost matters more than peak accuracy.

Seed Audio 1.0 is ByteDance Seed's non-streaming audio generation model. It produces speech and other audio from a natural-language text prompt that can describe the desired voice, tone, and sound effects, optionally guided by a Seed speaker ID or a reference audio clip for voice cloning. Output is limited to 120 seconds per request and is billed per second of generated audio. Suited to audiobooks, voiceovers, games, and similar workloads.

Jev Router picks the best model and reasoning effort for each request, balancing quality, speed, and cost. It runs on Jev, TypeSafe's first System One model, and adapts as your conversation evolves. One endpoint gives you the whole model ecosystem, leveraging the ecosystem's token and spend share.

Kev 4B is a small open-weight decision model from Jared Palmer, built as a LoRA adapter and pointer head on Qwen3.5-4B-Base and served over the same /v1/systemone contract as TypeSafe's Jev. Send a state and typed questions (yes/no, multiple choice, or score) and it returns a calibrated probability per question in one forward pass, with no generated text. It is the recommended checkpoint in the Kev family (0.8B, 4B, 9B) and is suited for routing, classification, and policy checks that want a compact, Apache-2.0 alternative to Jev. Code, model cards, and eval suites: https://github.com/jaredpalmer/kev

Perceptron Mk1.5 is Perceptron's embodied reasoning model for physical agents. It accepts text, image, video, and audio input, and answers with text plus optional structured annotations: points, boxes, polygons, tracks, and clips. It supports graded reasoning through the standard reasoning controls, function tool calling, and structured outputs via JSON Schema. Structured annotations are emitted inline with text only when requested via the annotation_format parameter ("point", "box", or "polygon" for spatial localization on images, "clip" for temporal segments in video). Video soundtracks are analyzed only when explicitly enabled per request.

Gemini 3.5 Transcribe is a speech-to-text model from Google. It is suited for synchronous transcription that needs word-level timestamps or speaker diarization, with support for up to eight speakers. Audio can be up to one hour, or 30 minutes when timestamps or diarization are enabled.

Transcribe 1 Pro is a speech-to-text model from Fish Audio tuned for interviews, meetings, and podcasts. It labels speakers with inline speaker markers, preserves emotion and vocal-event cues such as [laughter], detects language automatically, and can return timestamped word-level segments.

Ember-1 is a specialized reasoning model from Fireworks Research, built on Kimi K3. It is designed to make every token go further: it produces shorter reasoning traces, using roughly 40% fewer tokens than the base model while maintaining comparable quality across Fireworks' evaluations. It is suited for coding, knowledge work, and agentic workflows where reasoning cost and latency matter.

Ming Image 0.1 Design Layer is an image-to-image model from inclusionAI that decomposes a flattened design image into separate RGBA layers, such as a background layer and foreground elements, and returns one image per layer. It requires exactly one reference image and a prompt describing the layer plan. Output format can be requested as PNG or WebP. Output dimensions follow the input image, so explicit sizes and aspect ratios are rejected instead of silently reshaped.