Reasoning

Qwen3 Next 80B A3B Thinking, an open-weight Alibaba model available via OpenRouter.

Qwen3 VL 8B Thinking, an open-weight Alibaba model available via OpenRouter.

Qwen3 VL 30B A3B Thinking, an open-weight Alibaba model available via OpenRouter.

Reasoning-focused model with chain-of-thought capabilities rivaling o1.

Qwen3 VL 235B A22B Thinking, an open-weight Alibaba model available via OpenRouter.

Updated R1 with improved reasoning accuracy and reduced hallucination.

MiniMax M1 open-weight LLM with a 1M-token context window.

Kimi K2 Thinking, an open-weight Moonshot AI model available via OpenRouter.

Moonshot AI's open-weight flagship: natively multimodal 1T MoE model (32B active) with text, image, and video input, thinking and non-thinking modes, and 256K context. State-of-the-art open-source coding and agentic performance (SWE-Bench Verified 80.2); Agent Swarm scales to 300 parallel sub-agents.

Coding-focused agentic model built on Kimi K2.6 with ~30% lower thinking-token usage. 1T MoE (32B active), native INT4 quantization, 256K context, text/image/video input. Forces thinking and preserve_thinking modes; tuned for long-horizon software engineering and agent workflows.

Qwen 3 Max with extended reasoning and chain-of-thought capabilities.

Lightweight reasoning model balancing chain-of-thought rigor with speed and cost efficiency.

Reasoning-focused model with step-by-step deliberation for complex math, coding, and science tasks.

The high-tier variant of Sakana AI's Fugu orchestration model. Reportedly outperforms Opus 4.8, GPT-5.5 and Gemini 3.1 Pro on many benchmarks (73.7 on SWE-bench Pro), standing shoulder-to-shoulder with Claude Fable 5. Built on the TRINITY and Conductor (ICLR 2026) research.

Most capable reasoning model in OpenAI's lineup with extended thinking for maximum reliability.

Sakana AI's multi-agent orchestration model — a trained "conductor" LM that routes and synthesizes responses across a swappable pool of underlying frontier models, exposed as a single API.