35+ LLM models, one API

DeepSeek V4 Flash 0423 is the earlier, text-only Flash entry in the V4 family. Its sparse mixture-of-experts design targets coding and conversation workloads with a long context window; it is distinct from the later 0731 revision.

DeepSeek V4 Pro 0423 is the larger text-reasoning model in the original V4 pair. Its publisher describes a 1.6T-parameter MoE model with 49B activated, designed for demanding reasoning and software work rather than native image understanding.

DeepSeek V4.1 Flash adds native image understanding to the Flash line. Its Causal Encoder-Decoder architecture separates input processing from output generation, with a design aimed at input-heavy coding and computer-use workflows.

DeepSeek V4 Flash 0731 is the publisher’s official V4 Flash release, following the earlier preview. It remains a text-only model; its post-training revision focuses on coding, reasoning and agent tasks.

Qwen3.8 Max is the flagship entry in the Qwen3.8 line: a large mixture-of-experts model with a 1M-token window that accepts text, image and video input. It shares the Qwen3.8 generation with the Flash model but sits a tier above it on price.

Qwen3.8 Flash is Alibaba’s multimodal reasoning entry for text, images and video in the OpenRouter catalog. It brings document and chart understanding into the same model selection as coding assistance and long-context analysis.

MiniMax M2.7 is a text model centered on agentic productivity: debugging, root-cause analysis and multi-step professional work. Its published examples involve external tools; a completion alone does not execute that workflow.

MiniMax M3 is the multimodal successor to M2.7: text, image and video input, text output, and a 1M-token window aimed at long-horizon agentic and coding work. Unlike the free M2.7 entry it is billed per token.

GLM 5.1 is the earlier entry in the GLM 5 line, positioned for long-horizon coding tasks that run without step-by-step supervision. It has a smaller context window than the 5.2 and 5.3 entries.

GLM 5.2 is Z.ai’s text-reasoning model for long-horizon engineering. Its publisher highlights a million-token context and IndexShare, which reuses sparse-attention indexing work for long inputs.

GLM 5.3 builds on the GLM 5.2 base model with revised post-training for complex coding and long-running tasks. OpenRouter documents always-on reasoning for this entry, a meaningful distinction when budgeting output.

GLM 5.3 Flash introduces native multimodality to the GLM 5 series. Unlike the text-only GLM 5.3 entry, its new base architecture combines sparse and linear attention for visual and long-context workloads.

Kimi K3 is Moonshot AI’s open-weight multimodal model for coding, knowledge work and long-horizon agents. Its published design emphasizes iterating against repositories, images, logs and runtime feedback rather than generating a single isolated answer.

Kimi K2.6 is a native multimodal model focused on coding, UI generation and agent orchestration. The reference catalog lists text and image inputs, with a smaller context window than the K3 entry.

Claude Opus 4.6 is an earlier Opus-generation model built for coding and long-running professional work, with a 1M-token window and text, image and file input. It is priced identically to the newer Opus entries in this catalog.

Claude Opus 4.7 is the Opus-generation model built for long-running, asynchronous agents. It keeps the 1M-token window and text, image and file input of 4.6, and exposes reasoning effort as a selectable parameter.

Claude Opus 5 is positioned by OpenRouter as Anthropic’s model for demanding reasoning, software engineering and extended agent work. The reference entry lists text, image and file input with text output.

Claude Sonnet 5 is Anthropic’s Sonnet-class model for coding and professional work. The reviewed reference describes adaptive thinking and an updated tokenizer, so migration involves more than changing a display name.

Claude Haiku 4.5 is an earlier small, fast Claude model positioned for high-volume work where cost and latency matter more than frontier reasoning. It has a 200,000-token window, the shortest of the Claude entries here.

Claude Haiku 5.5 is the current small, fast Claude model, aimed at high-volume and cost-sensitive work such as summarization, subagents and browser automation. It carries a 1M-token window at the lowest input price in the Claude lineup here.

Gemini 3.1 Pro Preview is Google’s frontier reasoning entry, accepting text, image, file, audio and video input over a 1,048,576-token window. The Preview label reflects the publisher’s release stage, not a difference in how the API is called.

Gemini 3.5 Flash is a multimodal Flash-tier entry from Google, accepting text, image, video, file and audio input over a 1,048,576-token window. The publisher positions it for coding proficiency and parallel agentic execution.

Gemini 3.7 Flash is a multimodal entry built for fast agentic workflows, coding and multi-step reasoning, with a 1,048,576-token window over text, image, video, file and audio input.

Gemini 3.8 Flash is the newest Flash-tier entry from Google, with a 1,048,576-token window and text, image, video, file and audio input. The publisher describes it as improving on 3.7 Flash in software engineering and multi-step reasoning.

GPT-5.6 Sol is the flagship entry of the GPT-5.6 family in OpenRouter’s catalog. Its stated focus is complex reasoning, command-line coding and multi-step problem solving, with text, image and file inputs listed in the reference.

GPT-5.6 Luna is the GPT-5.6 entry aimed at high-volume, latency-sensitive chat, classification and lightweight agent tasks. It is a separate cost-efficiency candidate, not a claim of Sol-equivalent reasoning at a lower price.

GPT-6 Luna is the fast, cost-efficient member of the GPT-6 series, positioned below GPT-6 Sol. It accepts file, image and text input over a 1,050,000-token window and is aimed at high-volume, latency-sensitive work.

Grok 4.6 is an xAI model covering coding, knowledge work and STEM tasks, with a 500,000-token window over text, image and file input. It has since been succeeded by Grok 4.7 in the same lineup.

Grok 4.7 is the flagship xAI entry for coding, agentic tasks and knowledge work, with a 500,000-token window and text, image and file input. The publisher highlights long-running software engineering and self-verification.

Tencent Hy3 is a text-only MoE model with 295B total and 21B active parameters in its publisher card. The reference describes direct-answer and reasoning modes, with an emphasis on multi-turn constraints and tool-driven work.

Xiaomi MiMo-V2.5 is an omnimodal model whose reference entry lists text, image, audio and video inputs. Its output is text: broad perception support should not be confused with speech or image generation.

Tencent Hy4 preview scales the Hy family to a 770B-parameter backbone with 49B active parameters and a longer context than Hy3. It is a text-only preview aimed at coding agents and sustained tool-use workflows.

MiMo-V2.6-Pro is Xiaomi’s flagship open-weight entry, described by the publisher as a model above one trillion parameters. It accepts text, image, video and audio input over a 1,050,000-token window and returns text.

Step 3.7 Flash combines a sparse language backbone with a vision encoder. Its publisher emphasizes perception plus tool orchestration; the OpenRouter entry lists image and video input and selectable reasoning effort.

MiMo-V2.6-Flash is Xiaomi’s efficiency-oriented multimodal entry for coding and agent workflows. OpenRouter links the Flash-RL checkpoint card, which describes reinforcement-learning work across varied task environments.