Getting started

Pick a model

Bells Up AI gives you access to the leading AI model families, the same models used by Am Law 100 firms and Fortune 500 legal departments, with your data kept private.

Select a provider from the top bar, then choose a specific model version from the dropdown.

Within a family, the versions run from faster and more efficient to more capable. Start with the lighter option that can complete the task reliably, and move up only when the work calls for it. With Anthropic Claude, that ladder runs Haiku → Sonnet → Opus; with OpenAI GPT, Luna → Terra → Sol. The other families follow the same small-to-large pattern. Reserving the most capable models for the work that needs them keeps responses fast and costs down, and, as our environmental impact paper explains, uses less energy per query.

Each family name below links to its privacy disclosure. Our privacy reference page sets out who receives and keeps what you type.

Model family Overview
Claude (Anthropic) Follows complex instructions and analyzes long documents. Anthropic emphasizes model safety and alignment. Claude Sonnet 5 is the default Claude version.
GPT (OpenAI) The models behind ChatGPT. They handle step-by-step reasoning and produce structured output.
Gemini (Google) Large context windows suit reviewing and summarizing long documents. Gemini’s writing style is direct and structured.
Grok (xAI) Grok models are fast and inexpensive. xAI markets them as having an “edgy” and “less woke” personality, and they may refuse fewer requests on sensitive topics.
Llama (Meta) Widely deployed by developers and vendors building AI tools, including tools that can be self-hosted. No longer a leader on reasoning or drafting.
DeepSeek Open-weight models from a China-headquartered developer, noted for efficient training and inference.
Mistral A European-headquartered developer. Mistral models are common in developer tools and enterprise workflows.
Qwen (Alibaba) Open-weight models from Alibaba, a China-headquartered company. Often used for coding and multilingual work.
Kimi (Moonshot AI) Open-weight models from a China-headquartered developer, strong on long-document analysis and coding. Kimi K3 runs on Fireworks’ United States servers; Kimi K2.5 and K2 Thinking run on Amazon Bedrock. Moonshot never receives your content.
GLM (Z.AI) Open-weight models from a China-headquartered developer, strong in coding and outperforming many Western models on Chinese-language tasks. GLM 5.3 and GLM 5.3 Flash run on Fireworks’ United States servers; GLM 5, GLM 4.7, and GLM 4.7 Flash run on Amazon Bedrock. Z.AI never receives your content.

Where to start

On the free plan, start with Claude Haiku 4.5 for careful work and GPT-5.6 Luna for fast drafting. On a paid plan, Claude Sonnet 5 is the step up for work that needs more depth.

For more on what each family brings to the table, see Why we offer these models. Your firm’s administrator can disable a family for everyone at the firm. Any user can hide a family from their own account.