Key points
- A handful of frontier labs build the most capable models; each offers tiers from large and powerful to small and fast.
- Open-weight models can run privately on your own infrastructure; closed models are used through the vendor's services.
- For most organizations, the right choice depends more on your existing tools, data terms, and use case than on benchmark scores.
Who builds the leading models
These are the model families you'll hear about most often. Specific version names change frequently — follow our AI news feed for the latest releases.
| Lab | Model family | Where you'll meet it |
|---|---|---|
| OpenAI | GPT models, including reasoning models | ChatGPT, the OpenAI API, Microsoft Azure |
| Anthropic | Claude (Opus, Sonnet, and Haiku tiers) | Claude apps, the Claude API, Amazon Bedrock, Google Vertex AI |
| Google / DeepMind | Gemini (plus open Gemma models) | Gemini app, Google Workspace, Vertex AI, Android |
| Microsoft | Copilot (built on partner and in-house models) | Microsoft 365, Windows, GitHub, Microsoft Foundry |
| Meta | Llama (open-weight) | Meta apps; self-hosted and on major clouds |
| xAI | Grok | Grok app, X, the xAI API |
| Mistral AI | Open-weight and commercial models | Le Chat, the Mistral API, major clouds; strong in Europe |
| Others | DeepSeek, Qwen (Alibaba), Amazon Nova, and more | Open-weight downloads and cloud model catalogs |
Model tiers: big, medium, small
Most labs offer several sizes. Large models are the most capable and the most expensive; mid-sized models balance quality and cost and handle most business work well; small, fast models are cheap and quick, ideal for high-volume tasks like classification or routing. Many products now pick a model automatically, or let you toggle a slower "reasoning" mode for hard problems.
Open vs. closed models
| Closed (API / app) | Open-weight (self-hosted) | |
|---|---|---|
| Capability | Typically the frontier | Strong and improving fast |
| Data control | Governed by vendor contract | Can stay entirely on your infrastructure |
| Effort | Low — sign up and go | Higher — you run, secure, and update it |
| Cost model | Per seat or per use | Hardware / cloud compute |
| Good for | Most everyday and business use | Highly sensitive data, offline needs, customization |
Where a model comes from matters for compliance. Using a hosted service sends data to that provider under its terms and jurisdiction; running open weights yourself does not. Review both the model license and the hosting arrangement before use.
How to choose (without chasing benchmarks)
- Start with the tools you already have. If you live in Microsoft 365 or Google Workspace, their built-in AI is often the simplest safe starting point.
- Match the task. Writing and summarizing work well on most models; complex analysis benefits from reasoning modes; high-volume automation favors small models.
- Check the data terms for your specific plan — training, retention, residency, and certifications.
- Test on your own examples. Twenty real tasks from your team beat any public leaderboard.
- Stay portable. Standards like MCP and a thin integration layer make switching models later far easier.
Our platform picker gives a quick starting point, and our platform guide lists the questions to ask any vendor.
Keeping up with the pace of change
Frontier labs ship major updates several times a year, and smaller improvements constantly. You don't need to react to every release. A sensible rhythm: review your AI tools quarterly, follow a trusted summary like the Weekly AI Brief, and re-test only when a release clearly addresses a need you have.