Browse model providers, compare their offerings and follow changes to the APIs you build on.
OpenAI develops GPT models and provides hosted APIs for reasoning, tool use and media generation.
- GPT-6.2 inputs and limits. Astra and Sol accept text and images and return text. Each has a 1,050,000-token context window, with maximums of 922,000 input tokens and 128,000 output tokens. Source ↗
- Tool calling. GPT-6.2 Astra and Sol support both Responses and Chat Completions, with tool calling available through Responses. Source ↗
Monitored
Anthropic develops Claude language models and serves them through its API and cloud partners.
- Model families. Claude Fable 5.2 and the Opus, Sonnet and Haiku 5.5 models accept text and images, return text, support tool use and have one-million-token context windows. Source ↗
- Direct API pricing. Anthropic launched Claude Fable 5.2 on November 9, 2026 at a base rate of $8 per million input tokens and $40 per million output tokens. AWS, Google Cloud and Azure availability follows separate platform arrangements. Source ↗
Monitored
Google develops hosted Gemini models and publishes Gemma weights for deployment on independent infrastructure.
- Direct Gemini API access. Gemini API accounts start on a rate-limited free tier for select models. Paid access requires a linked Cloud Billing account. Source ↗
- Gemini 4 Neon limits. Neon accepts text, images, video, audio and PDFs and returns text. Default synchronous output is capped at 65,536 tokens, and long decode shares the 1,048,576-token context budget. Source ↗
Scheduled · awaiting first completed check
DeepSeek develops downloadable and hosted models for reasoning, tool use and multimodal workloads.
- Model IDs and inputs. The direct API serves V4.1-Flash-1017 as deepseek-flash and V4-Pro-0813 as deepseek-v4-pro. Both support one-million-token context, thinking and tool calls. Flash also accepts images. Source ↗
- Direct API billing. DeepSeek rates vary by cache hit, cache miss and output tokens and by peak or off-peak period. Source ↗
Monitored
Tencent develops Hunyuan and Hy models for language and multimodal applications.
- Hy4 Flash inputs. Hy4 Flash accepts text and images, returns text and supports a one-million-token context window. Source ↗
- Deployable checkpoints. Tencent publishes instruction-tuned, FP8 instruction-tuned and pretrained base checkpoints. The model card documents local Transformers and OpenAI-compatible vLLM serving. Source ↗
Monitored
Z.ai develops GLM models for coding, reasoning, tool use and document processing.
- GLM-5.3-Flash reasoning. GLM-5.3-Flash uses hybrid sparse and linear attention, supports a one-million-token context and runs with thinking enabled. Source ↗
- Direct API and plan billing. New Standard keys draw from the pay-as-you-go balance or bundles. Coding Plan keys draw from the associated plan quota. Some older Compatible keys still follow endpoint and account rules. Source ↗
Monitored
Alibaba publishes Qwen model weights and serves Qwen reasoning, vision and audio models through Model Studio.
- Qwen4-Max inputs and tools. Qwen4-Max accepts text, images and video and returns text. It supports parallel tool calls and JSON Schema output across a one-million-token context. Source ↗
- Direct API access. Model Studio pay-as-you-go API keys are region-specific and separate from Token Plan and Coding Plan keys. Source ↗
Monitored
NVIDIA publishes Nemotron text, vision, speech and retrieval models and offers GPU inference through NIM.
- Nemotron distribution. Nemotron 3.5 Super VL and Lightning are available as downloadable checkpoints, NVIDIA NIM endpoints and separately routed OpenRouter endpoints. Source ↗
- Local vision model. Nemotron 3.5 Bolt 74B-A9B 24GB uses native 2-bit quantization to fit 24 GB of memory and ships under the OpenMDW 1.1 license. Source ↗
Monitored
MiniMax develops language, image, video and audio models and provides a hosted developer API.
- Language models. MiniMax M3 and M3.1-Flash support multimodal input and a one-million-token context. Flash exposes adjustable thinking depth. Source ↗
- Direct API access. Direct API calls require a MiniMax account and an API key, with pay-as-you-go billing or a separate M Plan subscription. Source ↗
Monitored
Meta develops the Llama and Muse model families and serves Muse models through the Meta Model API.
- Llama weights. Llama 4 uses a multimodal mixture-of-experts architecture. Its downloadable weights use the Llama 4 license and acceptable use policy. Source ↗
- Hosted Muse access. Muse Wonder accepts text, images, video, audio and PDFs, returns text, and has a 1,048,576-token context. The standard API is pay-as-you-go, while Wonder Full is exclusive to Muse Code Power Usage and Maximum Usage plans. Source ↗
Monitored
The xAI developer platform, now branded SpaceXSI, serves Grok language, reasoning and media models.
- Grok 4.7. Grok 4.7 accepts text and images, returns text, and supports function calls, structured output and batch requests across a 500,000-token context. Source ↗
- Direct access. Requests to api.x.ai require a developer console key. The quickstart documents OpenAI SDK compatibility and a dedicated xAI SDK. Source ↗
Monitored
Mistral AI provides hosted language, vision and audio models alongside downloadable weights for selected models.
- Mistral Small 4. Mistral Small 4 accepts text and images, has a 256,000-token context and supports either no reasoning or high reasoning. Source ↗
- Direct API access. Direct requests require a Mistral Studio API key and a plan with API access. Source ↗
Monitored
inclusionAI publishes Ling and Ring language models for self-hosted instruction following and reasoning.
- Ling 3.1 Flash. Ling-3.1-flash uses hybrid attention and a mixture of experts with 560 billion total and 25 billion active parameters, with a one-million-token context. Source ↗
- Smaller checkpoints. Ling-mini-2.0 offers MIT-licensed base and post-trained checkpoints. Its 32,000-token context extends to 128,000 tokens with YaRN. Source ↗
Monitored
Moonshot AI publishes Kimi models for long-context reasoning, multimodal input and tool use through its API and downloadable weights.
- K3.1 Pro inputs. K3.1 Pro supports a one-million-token context, image and video input, and function calling. The API accepts images as base64 data or uploaded file IDs. Source ↗
- Direct API billing. Kimi bills its pay-as-you-go API by token separately from Kimi membership and Kimi Code. Cache writes and cached inputs have distinct rates. Source ↗
Monitored
StepFun develops Step models for reasoning and multimodal input and offers both hosted access and downloadable weights.
- Hosted preview limits. The hosted step-5-preview model accepts text, images and video, with a one-million-token context and up to 64,000 output tokens. Source ↗
- Direct API access. The hosted preview requires a StepFun API key, bills input and output tokens, and supports JSON Schema output and application-defined tools. Source ↗
Monitored
Xiaomi publishes MiMo models for reasoning, coding and multimodal applications.
- MiMo V2.6 Pro MOPD inputs. This checkpoint supports a one-million-token context and accepts text, images, video and audio. Source ↗
- Self-hosting. The MiMo V2.6 Pro MOPD weights are MIT-licensed and Xiaomi documents serving them with SGLang and vLLM. Source ↗
Monitored