Mainnet Models

Newest OpenAI flagship — first fully retrained base since GPT-4.5. 1M context, 128K output, native agent + computer use

Most capable and efficient frontier model with 1M context, native computer use, and thinking mode

Premium GPT-5.4 with maximum compute for the hardest problems

High intelligence with medium speed. Multimodal with vision, function calling, and structured outputs

Frontier model with 400K context and adaptive reasoning

Strongest mini model for coding, computer use, and subagents with GPT-5.4 capabilities

Cost-optimized reasoning and chat

Fastest and most affordable GPT-5.4 model for high-throughput tasks

Uses more compute for consistently better answers

Industry-leading agentic coding model. 400K context, reasoning, tool use, and complex execution

Latest GPT-4 generation model

Fast and affordable GPT-4.1 model

Ultra-fast and cost-effective GPT-4.1

Multimodal model with vision and audio

Fast and affordable GPT-4o model

Advanced reasoning model for complex tasks

Latest reasoning model with improved performance

Efficient reasoning model for STEM tasks

Latest generation efficient reasoning model

Fastest and most efficient Claude, near-frontier intelligence

Newest Sonnet — near-Opus coding/agentic quality at Sonnet cost. 1M context, 128k output, adaptive thinking, vision

Best balance of intelligence, speed, and cost

Sonnet 4.5 — strong coding and agentic performance, vision

Latest Anthropic flagship with enhanced reasoning and creativity

Powerful Claude Opus for complex reasoning and agentic coding. 1M context, 128k output, adaptive thinking

Anthropic's most capable model — Mythos-class tier above Opus, for the most demanding reasoning and long-horizon agentic work. 1M context, 128K output, always-on thinking

Most capable Claude for complex reasoning and agentic coding. 1M context, 128k output, adaptive thinking

Latest Gemini with improved thinking, token efficiency, and agentic capabilities. Optimized for software engineering (requires new SDK)

Frontier-class performance with Pro-level intelligence at Flash speed and pricing. Includes thinking mode (requires new SDK)

Latest-generation Flash with built-in thinking mode — frontier-class quality at Flash speed and pricing

State-of-the-art for reasoning, coding, and mathematics

Fast and efficient Gemini model with vision support

Ultra-fast and lightweight Gemini 3.1 model with thinking mode for high-throughput tasks

Most economical Gemini model - ultra-fast and lightweight (requires new SDK)

DeepSeek V4 flagship — 1.6T MoE / 49B active, 1M context. Strongest open-weight reasoner. Thinking mode default.

Paid V4 Flash in non-thinking mode (1.6T-class quality at $0.20 in / $0.40 out). Same model as the free nvidia/deepseek-v4-flash but on a paid endpoint with higher reliability and 5MB request bodies.

Paid V4 Flash in thinking mode for reasoning tasks. Same upstream as deepseek/deepseek-chat but with thinking enabled by default.

Moonshot's flagship multi-modal reasoning model — 256K context, image + VIDEO input, returns reasoning_content. Served via the OpenRouter credit pool (slug moonshotai/kimi-k2.7-code), failing over to direct Moonshot.

Z.AI's newest flagship — 1M-token context, top open-source on long-horizon coding. Verified live on Z.AI.

Z.AI flagship — #1 open source on SWE-Bench Pro, 8-hour autonomous execution. 200K context

Z.AI's foundation model with 200K context. Strong reasoning and agentic capabilities

Optimized GLM-5 variant with faster inference

xAI's Grok 4.3 reasoning model. 1M context, vision-capable, tuned for agentic workflows and instruction-following.

xAI's fast agentic coding model, trained for interactive software-engineering workflows. 256K context, text + image input.

MiniMax's flagship reasoning model with recursive self-improvement. Great value for complex tasks (~60 tps)

MiniMax's M3 flagship — 1M context, strong reasoning + coding. Served via OpenRouter.

NVIDIA's multimodal reasoning Nemotron Nano Omni hosted free by NVIDIA. 31B / 3.2B active MoE. Accepts text, images, video, audio. ChartQA 90.3, DocVQA 95.6, MMMU 70.8 — the only vision-capable free model in our catalog

Mistral's flagship 675B model hosted free by NVIDIA. Largest Mistral model ever released

Meta's Llama 4 Maverick MoE (17B x 128 experts) hosted free by NVIDIA

Qwen3-Next 80B (3B active MoE) hosted free by NVIDIA. 262K context, strong reasoning + coding, fast.

Qwen3.5 122B MoE (10B active) hosted free by NVIDIA. Balanced reasoning + coding, 131K context.

Mistral × NVIDIA Nemotron instruction model hosted free by NVIDIA. Fast (~0.2s), strong instruction following.

StepFun Step 3.7 Flash hosted free by NVIDIA. Fast lightweight reasoning, 131K context.

ByteDance Seed-OSS 36B instruct hosted free by NVIDIA. Strong open-source coder, 131K context.

NVIDIA Nemotron Nano 9B v2 hosted free by NVIDIA. Compact + fast (~0.7s), good for high-volume light tasks.

NVIDIA Nemotron Nano 12B v2 Vision-Language hosted free by NVIDIA. Accepts images; compact + fast.