The AI Model Reference

Anthropic's next-generation model for knowledge work and coding. It is built for days-long projects, deep analysis, and work that needs fewer check-ins. This is the top of the ladder the rest of this guide is measured against.

Anthropic's powerful model for complex tasks and deep research. It handles documents, slides, spreadsheets, and analysis that needs real reasoning. One step below Fable, and the model most people mean when they say Claude is smart.

The balanced workhorse. Sonnet is the middle of the ladder, the model that does most everyday work well without the cost or latency of the frontier tiers. It is the default choice for most people most of the time.

The fast, light tier. Haiku is built for speed and cost, the model you reach for when the task is routine and you want an answer now. It is the fundamental rung of the ladder.

OpenAI's flagship model for complex reasoning and coding. Sol is the top of the GPT-5.6 family, priced at the top end, and it is the model OpenAI points you to when you are not sure where to start.

The balanced member of the GPT-5.6 family. Terra trades a little of Sol's peak capability for a lower price, and it is the model OpenAI recommends when you want to balance intelligence and cost.

The cost-sensitive member of the family. Luna is built for high-volume workloads where the price per token matters more than peak intelligence, and it is dramatically cheaper than Sol.

Google's most intelligent Flash model, engineered for long-horizon software engineering, autonomous agents, and complex enterprise workflows. The Flash line is Google's fast tier, and 3.8 is the most capable of it.

Google's advanced intelligence model for complex problem-solving and agentic work. Pro is the top of the Gemini line, the counterpart to Fable and Sol.

Google DeepMind's open family, built to deliver frontier-level performance at each size. The smaller E2B and E4B variants run on laptops and phones, while the 12B, 26B and 31B models are workstation-class. It is the open model that most closely tracks the closed Gemini line.

The smaller of Meta's two Llama 4 models, a mixture-of-experts design with 109B total parameters but only 17B active at any moment. That sparsity is what lets a model this capable run on hardware that could never hold 109B dense parameters.

Meta's largest open model, 400B parameters with 17B active. It is the most capable model Meta has released, and it is the default answer when someone asks what the strongest open model is.

Mistral's flagship, a 675B mixture-of-experts model released under Apache 2.0. It is open-weight but too large to run on consumer hardware, so it runs on Ollama's cloud. The European answer to the American and Chinese frontier models.

Mistral's small model, 24B parameters, built for function calling and instruction following. It is the model you run locally when you want Mistral quality without the datacenter.

DeepSeek's flagship, a 1.6-trillion-parameter mixture-of-experts model with three reasoning modes (none, thinking, max). It is the Chinese frontier model that reset the market's price expectations, and it runs on Ollama's cloud.

The efficient member of the V4 series, 284B total with 13B active, built for reasoning across a 1M-token window at a fraction of the Pro cost.

The open reasoning model that started the current wave. R1 is MIT-licensed and comes in distilled sizes from 1.5B up to the full 671B, so it runs on everything from a laptop to a cluster. Its reasoning patterns were distilled into the smaller Qwen and Llama models.

Alibaba's open multimodal family, spanning 0.8B to 122B parameters and 201 languages. It is the most widely pulled open model family on Ollama, and the 122B variant is the strongest of the line.

The follow-up to Qwen 3.5, focused on agentic coding and thinking preservation. It is built for developers who want a stable, productive coding experience from an open model.

Moonshot's flagship, a 2.8-trillion-parameter model and the world's first open 3T-class model. It is natively multimodal and built for long-horizon coding and knowledge work, and it runs on Ollama's cloud.

xAI's flagship, the model behind the Grok assistant on X. It is closed and cloud-only, with a 500K context window and a knowledge cutoff of February 2026. xAI positions it as the most intelligent and fastest model they have built.

Nvidia's open 30B mixture-of-experts model with only 3B active parameters, built for always-on agents. The sparsity gives it four times the throughput of similar-sized models, which is what you want for an agent that runs all day.

Z.ai's flagship, the most capable open-weights model for coding, with major gains on long-horizon agentic tasks. It is 753B parameters with a 1M context window, and it approaches Claude Opus on coding benchmarks. Z.ai is the consumer brand of Zhipu AI.

Microsoft's small open model, 14B parameters, built for reasoning and logic in memory-constrained environments. It is the model you run when the machine is modest and the task is focused.

Cohere's 111B open model, built for enterprise RAG and tool use. It is trained specifically for the final step of retrieval-augmented generation, which makes it a strong choice for business applications that answer from a document base.

IBM's enterprise-ready open family, released under Apache 2.0. Granite is built for governance, risk and compliance use, with tool use, thinking mode and structured JSON output. It is the model enterprises pick when the licence and the audit trail matter as much as the capability.