Every model, explained and priced.
Filter by what you’re building, search by capability or tag, and compare cost, speed, and benchmarks — on one OpenAI-compatible key.
What are you building?
Pick a job to narrow the catalog and surface leaders.
Fast, low-cost Nova for high-volume text and light multimodal tasks.
Smallest Nova — text-only ultra-cheap option for classification and routing.
Bedrock Nova Pro — multimodal workhorse with 300K context.
Anthropic's next-generation model for complex knowledge work and coding, built for sustained autonomous operation across multi-day tasks — plans across stages, delegates to sub-agents, and self-verifies its work.
Fastest and most compact Claude model. Ideal for classification, extraction, and high-volume tasks where cost and latency matter most.
Most capable Claude model. Sets the bar on advanced reasoning, research synthesis, and nuanced long-form content.
The latest Opus 4.x model — advanced reasoning and research synthesis, with fast mode available for premium-priced low-latency responses.
Balanced intelligence and speed. Handles complex coding, analysis, and multi-step reasoning without the cost of Opus.
Balanced Claude 5.x model with the full 1M-token context window at standard pricing. Handles complex coding, analysis, and multi-step reasoning at Sonnet-tier cost.
Mistral’s code-specialized model — fill-in-the-middle and multi-language coding.
Cohere Command R — efficient RAG and tool use at budget rates.
Cohere’s RAG-oriented flagship — strong retrieval-augmented generation and tool use.
OpenAI image generation — priced per image upstream.
DeepSeek’s chat alias — V3-class coding and general chat at low cost.
Open-source reasoning model rivalling o1. Publishes its chain of thought — useful for auditability and research workflows.
Mixture-of-experts model with strong coding and math benchmarks at a fraction of frontier model prices.
Bedrock-hosted DeepSeek V3.2 — coding and math at competitive MoE rates.
DeepSeek's flagship (Pro tier) — 1M-token context with strong reasoning and coding at a fraction of frontier-tier cost.
Mistral Devstral 2 — large coding-focused model on Bedrock.
Shared community capacity for Dots Studio Dots 3 Note Preview.
Prior Gemini Flash — cheap long-context option.
Prior Gemini Pro generation — 1M context for long documents.
Low-latency workhorse with a 1M context window and multimodal inputs.
Best-value Gemini model. Handles massive context at low cost with strong performance on coding and document analysis.
Gemini 2.5 Flash with native audio — preview long-context multimodal.
Google's highest-capability model with native multimodal reasoning across text, image, video, and audio — all within a 1M-token window.
Previous Flash tier — prefer 3.6 Flash for new workloads.
High-throughput Flash-Lite for extraction, classification, and subagent loops.
July 2026 Flash workhorse — stronger coding/agentic loops than 3.5 at lower output cost, 1M context.
Open Gemma 3 12B instruct — compact open-weight generalist on Bedrock.
Open Gemma 3 27B instruct — strong open-weight generalist on Bedrock.
Smallest Gemma 3 instruct — ultra-cheap classification and light chat.
Shared community capacity for Google Gemma 4 26B A4B IT (multimodal).
Shared community capacity for Google Gemma 4 31B IT. Rate-limited free pool.
Prior GLM-4.7 generation — capable bilingual generalist.
Fast GLM-4.7 tier for high-throughput bilingual workloads.
Zhipu GLM-5 — strong Chinese/English bilingual reasoning and coding.
Lightweight 4.1 variant with the same 1M-token window at a fraction of the cost. Ideal for agentic pipelines that need long context cheaply.
Cheapest 4.1-family model — high-throughput extraction and subagent loops with 1M context.
Flagship multimodal model optimised for real-time dialogue. Natively processes text, images, and structured data.
GPT-4o with native audio input/output (preview).
Affordable GPT-4o sibling that handles everyday tasks with excellent speed. Strong default for high-volume deployments.
OpenAI's most capable model — advanced coding, research, analysis, software operation, document workflows, and long-running agentic tasks with less orchestration.
Successor to GPT-5.5 (Sol tier) — OpenAI's current flagship for advanced coding, research, analysis, and long-running agentic tasks.
Groq-hosted open-weight flagship replacing Llama 3.3 70B. Strong reasoning at low latency.
Open GPT-OSS 120B via Bedrock — large open-weight generalist.
Groq-hosted open-weight model replacing Llama 3.1 8B. Fast, cheap, and ideal for high-volume prototyping.
Open GPT-OSS 20B via Bedrock — compact open-weight option.
Safety-tuned GPT-OSS 120B on Bedrock for moderation-style workloads.
Safety-tuned GPT-OSS 20B on Bedrock — compact moderation helper.
Prior Grok flagship — strong reasoning and coding with a 131K context window.
Lightweight Grok 3 — fast and cheap for lighter agent loops.
Long-context Grok — 1M tokens with strong reasoning at a lower rate than Grok 4.5.
xAI's current frontier model — strong reasoning and agentic coding with a 500K-token context window.
Shared community capacity for Thinking Machines Inkling.
Shared community capacity for Thinking Machines Inkling Small.
Thinking-mode Kimi for multi-step reasoning and research synthesis.
General Kimi K2.5 — long-context reasoning and coding at Bedrock rates.
Moonshot coding-focused Kimi — 256K context for large repos and agent loops.
Shared community capacity for Poolside Laguna S 2.1.
Shared community capacity for Poolside Laguna XS 2.1 — compact coding model.
Shared community capacity for Liquid LFM 2.5 2.6B. Small and fast.
Shared community capacity for InclusionAI Ling 3.0 Flash Fin — finance-tuned (MoE).
Shared community capacity for InclusionAI Ling 3.0 Flash Sante — health-domain tuned (MoE).
Shared community capacity for InclusionAI Ling 3.0 Flash VL — vision-language (MoE).
Meta Llama 3 70B instruct on Bedrock — classic open-weight workhorse.
Meta Llama 3 8B instruct — small, fast open-weight option on Bedrock.
Meta Llama 3.1 405B instruct via Fireworks.
Meta Llama 3.1 70B instruct via Fireworks.
Groq speculative-decoding Llama 3.3 70B — faster throughput variant.
Mistral Magistral Small — reasoning-oriented small model on Bedrock.
MiniMax M2 — long-context agentic model via Bedrock.
MiniMax M2.1 — long-context agentic model via Bedrock.
MiniMax M2.5 — long-context agentic model via Bedrock.
Ministral 3 14B — mid-size open instruct model on Bedrock.
Tiny Ministral 3 — ultra-cheap classification and light chat on Bedrock.
Compact Ministral 3 8B — small open instruct model on Bedrock.
Classic Mistral 7B instruct on Bedrock.
Flagship Mistral model with top-tier coding, reasoning, and multilingual performance. Native function calling support.
Mistral's current flagship (675B) — top-tier coding, reasoning, and multilingual performance, routed via Bedrock.
Efficient European model for classification, summarisation, and structured extraction. Strong on French and other EU languages.
Mixtral MoE instruct on Bedrock — classic open MoE generalist.
Shared community capacity for NVIDIA Nemotron 3 Nano Omni — multimodal reasoning.
Shared community capacity for NVIDIA Nemotron 3 Super.
Shared community capacity for NVIDIA Nemotron 3 Ultra (MoE). Large free context window.
Shared community capacity for NVIDIA Nemotron 3.5 content-safety classifier.
Shared community capacity for NVIDIA Nemotron 3.5 Lightning.
NVIDIA Nemotron Nano 12B — slightly larger nano tier for general chat.
NVIDIA Nemotron Nano 3 30B — open reasoning-capable mid-size model.
NVIDIA Nemotron Nano 9B — compact open model for light workloads.
NVIDIA Nemotron Super 3 120B — open reasoning and coding at Bedrock rates.
Shared community capacity for Nex AGI Nex N2.5 Mini.
Shared community capacity for Nex AGI Nex N2.5 Pro.
Shared community capacity for Cohere North Mini Code.
OpenAI's most powerful reasoning model. Spends more compute thinking before answering — ideal for math, science, and strategic planning.
Compact reasoning model tuned for STEM tasks. Faster and cheaper than o3 while retaining strong chain-of-thought capabilities.
Next-generation compact reasoning model. Faster than o3-mini with improved accuracy on code, math, and agentic workflows.
Dense Qwen3 32B — solid generalist at low cost on Bedrock.
Qwen3 MoE coder — strong coding at low cost with 256K context.
Latest Qwen3 coding model — strong agentic coding with 256K context.
Qwen3 Next MoE — long-context generalist for analysis and agents.
Large vision-language Qwen3 — document and image understanding at MoE scale.
OpenAI embedding model — higher-dimensional vectors.
OpenAI embedding model — small, cheap vectors for retrieval.
Mistral Voxtral Mini — compact audio-aware model on Bedrock.
Mistral Voxtral Small — larger audio-aware model on Bedrock.
OpenAI speech-to-text — priced per minute upstream.
Three steps to your first call
Getting startedCreate a free account
Email and a password. No card, no sales call. Community models are free from the first minute.
Create an inference key
Name it, scope it to the models you want, and set a spend ceiling. The cap is enforced before each request, not reconciled after.
Change one line
Point your existing OpenAI-compatible client at the Relixr base URL. Same SDK, same request shape, same response shape.