AI models explained: tiers, reasoning, open vs closed
Top, mid and small models, reasoning, and open versus closed: a plain overview with a model table.
A large language model is the engine behind a chat app. Providers sell several models at once. Pick the right one and you save money and time.
Three tiers
Most providers offer the same ladder:
- Top. Best at hard tasks: long analysis, tricky code, careful writing. Slowest and most costly.
- Mid. The everyday choice. Good quality at a fair price and speed.
- Small. Fast and cheap. Fit for sorting, short answers, extraction and bulk work.
Names change, so judge by tier and price, not by name. The price table shows what each tier costs.
Reasoning
A reasoning model works through a problem in steps before it answers. That often helps with maths, logic and planning. It also uses more tokens and takes longer, so it costs more. Use it when the task needs it, not for every question.
Open and closed
- Closed models run only on the provider's servers. You use them through an app or an API.
- Open-weight models publish their weights. You can run them yourself, as a local model, so data need not leave your machine. Check the licence: "open" does not always mean "free for any use".
Reading model claims
Providers publish scores from a benchmark. Treat them with care: a test is not your task. Also look at the knowledge cutoff and the size of the context window.
How to pick
Start small. Run your real task on a small model and read the result. If it falls short, move up one tier and test again. Pay for a top model only where the gain is clear. For work with private data, see safety.
Current models
| Model | Provider | Tier | Context | Reasoning | Image input | Open weights | Good at |
|---|---|---|---|---|---|---|---|
| Claude Fable 5.1 | Anthropic | top | 1M | yes | yes | no | Demanding reasoning and long-horizon agentic work |
| Claude Opus 5.5 | Anthropic | top | 1M | yes | yes | no | Long-running agentic coding and knowledge work |
| Claude Sonnet 5.5 | Anthropic | mid | 1M | yes | yes | no | Best mix of speed and intelligence |
| Claude Haiku 5.5 | Anthropic | small | 1M | yes | yes | no | Fast and cheap: classification, extraction, routing at volume |
| DeepSeek V4 Pro | DeepSeek | top | 1M | yes | no | yes | Strong agentic and coding model, text only, 1M context |
| DeepSeek V4.1 Flash | DeepSeek | small | 1M | yes | yes | yes | Very cheap multimodal model with strong coding and agent results |
| Gemini 4 Argon | Google (Gemini API) | top | 1.05M | yes | yes | no | Flagship reasoning model for software engineering, science and agents |
| Gemini 3.8 Flash | Google (Gemini API) | mid | 1.05M | yes | yes | no | Fast Flash model for long-horizon coding and autonomous agents |
| Gemini 3.1 Flash-Lite | Google (Gemini API) | small | 1.05M | yes | yes | no | Low-cost model for translation, extraction and simple tasks at scale |
| Gemini 3.5 Flash-Lite | Google (Gemini API) | small | 1.05M | yes | yes | no | Cheapest 3.5 model for high-throughput subagent and parsing tasks |
| Mistral Large 4 | Mistral | top | 1M | yes | yes | yes | Open-weight multimodal MoE model, 1.05T parameters, 52B active |
| Mistral Medium 3.5 | Mistral | mid | 256k | yes | yes | yes | Frontier-class multimodal model for agentic and coding work |
| Ministral 3 8B | Mistral | small | 256k | no | yes | yes | Efficient small model with text and vision for edge use |
| Mistral Small 4 | Mistral | small | 256k | yes | yes | yes | Hybrid instruct, reasoning and coding model, 119B parameters, 6.5B active |
| Kimi K3 | Moonshot AI (Kimi) | top | 1.05M | yes | yes | yes | Previous Kimi flagship, 2.8T parameters, open weights, 1M context |
| Kimi K3.1 Pro | Moonshot AI (Kimi) | top | 1.05M | yes | yes | no | Newest Kimi flagship for coding, agents and long-horizon work |
| GPT-6 Astra | OpenAI | top | 1.05M | yes | yes | no | OpenAI's most capable model for demanding reasoning and coding |
| GPT-6.1 Sol | OpenAI | mid | 1.05M | yes | yes | no | Near-Astra quality for complex coding and professional work |
| GPT-6.1 Terra | OpenAI | mid | 1.05M | yes | yes | no | Fast model for visual reasoning, computer use and clear writing |
| GPT-6 Luna | OpenAI | small | 1.05M | yes | yes | no | Most efficient model for focused, high-volume tasks |
| Grok 4.7 | xAI (Grok) | top | 500k | yes | yes | no | xAI's flagship for coding, agentic tool calling, configurable reasoning |
| Grok 4.3 | xAI (Grok) | mid | 1M | yes | yes | no | Fast, reliable model with strong tool calling and instruction following |
| Grok Build 0.1 | xAI (Grok) | small | 256k | yes | yes | no | Coding model for agentic software and engineering workflows |
USD per 1 million tokens, list price, excl. tax. Prices as of 26 October 2026.
Not sure which to use? Read how to choose a model.
All pages in this section
- How to choose an AI model
Match the model to the job: small for volume, mid for daily work, top or reasoning for hard problems and local for private data, with a decision tree and a test.
Read next
- How to choose an AI model
Match the model to the job: small for volume, mid for daily work, top or reasoning for hard problems and local for private data, with a decision tree and a test.
- Reasoning model
A reasoning model spends extra tokens on hidden step-by-step thinking before it writes its answer, which helps on hard problems but costs more.
- Open-weight model
An open-weight model is an AI model whose trained weights are published for download, so you can run and often adapt it yourself.
- Large language model (LLM)
A large language model (LLM) is an AI system trained on huge amounts of text to predict the next token, which lets it write, summarise and answer.