Models

Specify a model

To use a particular model, pass a "{provider}/{model}" string to the client parameter, which gets passed along to ellmer::chat():

qc <- querychat(penguins, client = "anthropic/claude-sonnet-4-5")
qc$app()  # Launch the app

And, if you’d like to effectively set a new default model, you can use the querychat.client R option or the QUERYCHAT_CLIENT environment variable.

# In your .Rprofile
options(querychat.client = "anthropic/claude-sonnet-4-5")

Note that it can also be useful to pass a full Chat object to the client parameter for more advanced use cases (e.g., custom parameters, tools, etc):

client <- chat_anthropic(model = "claude-sonnet-4-5")
qc <- querychat(penguins, client = client)
qc$app()  # Launch the app

Credentials

Most models require an API key or some other form of authentication. See the reference page for the relevant model provider (e.g., chat_anthropic()) to learn more on how to set up credentials.

library(ellmer)

# Just works if GITHUB_TOKEN is set in your environment
client <- chat_github(model = "gpt-4.1")

In general, most providers will prefer credentials stored as environment variables. Common practice is to use an .Renviron file to manage these variables. For example, for chat_openai(), you might add to your .Renviron file:

OPENAI_API_KEY="your_api_key_here"

Then, you can edit your .Renviron file using:

usethis::edit_r_environ()

Recommended models

In theory, you could use any model that has tool calling support, but we currently recommend (as of November 2025):

• GPT-4.1 (the default)
• Claude 4.5 Sonnet
• Google Gemini 3.0

In our testing, we’ve found that those models strike a good balance between accuracy and latency. That said, smaller/faster models like GPT-4.1-mini or Claude Haiku 4.5 work well for most tables and are worth trying first—they’re significantly cheaper and faster. You can always switch to a larger model if you find the results aren’t meeting your needs. On the other end of the spectrum, reasoning models like o3-mini tend to slow down responses without providing meaningfully better results for this task.

We’ve also seen some decent results with frontier local models (e.g., gpt-oss:20b), but even if you have the compute to run the largest models, they still tend to lag behind the cloud-hosted options in terms of accuracy and speed.