Choose the right AI model

Pick the right AI model for each Agent, chat, and Workflow step using Cassidy’s recommended models and the Capability and Usage consumption ratings.

Cassidy gives you access to leading AI models from OpenAI, Anthropic, and Google. You can choose a model for each Agent, each chat, and each Workflow step.

If you’re not sure which model to use, choose Default. It works well for most tasks, and Cassidy keeps it up to date as new models come out.

Where you choose models

Agents

Open your Agent and choose a model in the AI Model field on the Setup tab. Everyone who chats with the Agent uses that model.

Chat

Use the picker in the message box to switch between models and Agents at any time during a conversation. Hover over a model in the picker to see its Capability and Usage consumption ratings.

Workflows

In a Generate Text action, choose a model in the AI Model field. Each step can use a different model, so routine steps can use less while harder steps use a smarter model.

Start with recommended models

The Recommended Models at the top of the model selector are Cassidy’s picks for common kinds of work. Cassidy updates them as new models are released and tested, so you don’t need to track model releases yourself.

Recommended pick	Best for
Default	Everyday Agent chats and Workflow steps — the best starting point for most work
Fastest	High-volume, routine tasks like classification, extraction, formatting, and summaries
Best Writer	Emails, marketing copy, and other writing that should sound natural
Smartest	Complex work, connectors, artifacts, and anything where accuracy matters most

Compare capability and usage

Hover over any model in the model selector or the chat picker to see two ratings:

• Capability: how smart the model is.
• Usage consumption: how much usage the model takes to do the same work.

Hover over the Capability or Usage consumption label for a short explanation, and select More info for the full details. For a recommended pick such as Smartest, the ratings describe the model it uses right now, shown as Current Model.

Capability

Capability shows how smart a model is at reasoning, analysis, and problem solving, based on public benchmarks and our own testing. The smartest model you can choose scores 100, so a model that scores 80 is about 80% as capable. Hover over the sparkles to see a model’s capability score.

Sparkles	Capability score	What it means
5	90 to 100	One of the smartest models you can choose
4	70 to 89	Close to the smartest models
3	50 to 69	Good for everyday work
2	30 to 49	Best for simpler tasks
1	Under 30	Simple tasks only

The smartest model you can choose sets the top of the scale. When a smarter model comes out, other models’ scores go down and they can lose a sparkle without getting any worse.

Usage consumption

Usage consumption shows how much of your usage a model takes to do the same work. Every model is given the same set of benchmark tasks, and we measure how much each one uses. That includes the model’s price and how much it writes and thinks along the way, so a model with a low price can still use a lot if it thinks for a long time.

Hover over the rating, such as Moderate, to see where a model falls on the full scale. Select How is this calculated? for the details.

Every model is measured against our Default model (currently GPT-6.1 Sol), which is always Moderate. For example, if a task uses 10 credits on Default, the same task would use roughly:

Usage consumption	Same task, if Default uses 10 credits	In other words
Low	Under 5 credits	Less than half as much
Moderate	5 to 30 credits	Between half and 3 times as much
High	30 to 100 credits	3 to 10 times as much
Very high	Over 100 credits	More than 10 times as much

Each step up the scale is a bigger jump than the one before. The bar in the model selector gives each level the same amount of space so the whole range fits, but moving from High to Very high adds far more usage than moving from Low to Moderate. Some Very high models use 20 times as much as Default or more.

Models that haven’t been benchmarked yet show Not tested yet for both ratings.

Where the ratings come from

The ratings combine public benchmarks, such as the Artificial Analysis Intelligence Index, with our own testing. We adjust the results for how Cassidy runs each model and update them as new models are added.

Choose a thinking level

Many models come in more than one thinking level. The thinking level controls how long the model works before it answers.

Thinking level	When to use it
Instant	Fast replies when speed matters more than extra deliberation
Thinking	Better accuracy for most Agent work
High Thinking	The hardest problems — slower, with the highest usage

Each thinking level has its own Capability and Usage consumption ratings. Turning thinking up usually makes a model smarter and raises its Usage consumption, because the model writes more before it answers.

A few thinking levels haven’t been benchmarked yet. When the closest benchmarked level is a good match, those options show its ratings, and the How is this calculated? window says which level the ratings are for. When it isn’t, for example when only the highest thinking level was benchmarked, we estimate the ratings from the levels that were. An asterisk on the capability score, such as 61/100*, marks an estimate.

You can think of it like this: the model is how smart the brain is, and the thinking level is how long that brain gets to work.

Model providers and strengths

Each provider has distinct strengths. Use this as a starting point, then try models on your own tasks.

GPT (OpenAI)

A strong all-rounder for Agents, everyday questions, and automation. Reliable for finding information and answering a wide range of business questions.

Claude (Anthropic)

Preferred for writing that needs a human touch, like emails, sales outreach, and marketing copy.

Gemini (Google)

Strong at combining information from many sources and working with documents, images, and other files. A good fit for deep research.

Open Source

Leading open models such as DeepSeek, Kimi, and GLM, for teams that want options beyond the major AI labs. Learn more about Open Source models.

If your organization uses Open Source models, compare them using the same Capability and Usage consumption ratings as every other model.

Tips for choosing a model

• Start with Default unless you already know you need Fastest, Best Writer, or Smartest.
• Choose the lowest Usage consumption that still does the job well.
• Use Fastest or another Low usage model for high-volume Workflow steps like classification, extraction, and formatting.
• Move up to Smartest, or turn thinking up, when accuracy matters more than speed or usage.
• Mix models within one Workflow: a Low usage model for routine steps, and a smarter model only where it makes a difference.
• Try important Agents and Workflows with a couple of models before you settle on one.

Keep in mind

• Ratings come from general benchmarks. They don’t measure how well a model follows your Agent’s instructions or writes in your company’s voice.
• Your actual credit usage depends on your prompts, Knowledge Base results, tools, response length, and thinking level.
• The benchmarks lean toward reasoning and multi-step work, so models that think a lot can look more expensive than they are on simple questions.
• Ratings are refreshed when new models are added, so they can lag behind recent changes from model providers.

To track and manage your credits, see Manage credits.

Next steps

Build and configure an Agent

Create a custom Agent and select the right model for its role.

Prompt engineering guide

Write effective prompts that get the most out of any model.