AIscreenR 0.5.1

Major improvements

All gpt-4 and gpt-3.5 models (e.g., gpt-4, gpt-4o, gpt-4o-mini, gpt-4.1) are deprecated and have been removed from model_prizes. tabscreen_gpt() and tabscreen_gpt.tools() now stop if one of these models is used, also when custom_model = TRUE.
The top_p argument has been removed from tabscreen_gpt(), tabscreen_gpt.tools(), and approximate_price_gpt(), and the top_p variable is no longer part of their output. The argument is unchanged for the other providers (Groq, Mistral, Gemini, and Ollama).
The default model is now "gpt-6-luna" in tabscreen_gpt(), tabscreen_gpt.tools(), approximate_price_gpt(), and rate_limits_per_minute().
reasoning_effort and verbosity are now added to the request body for all gpt-5 and newer models.
Refactoring the price calculations. The reasoning token multipliers have been re-measured per model and reasoning effort, and prices are now looked up with a join on model_prizes.

Adding the confidence argument to all tabscreen functions. When confidence = TRUE, the model also reports how confident it is in each decision (0-100), returned in the new confidence variable. When reps > 1, the aggregated data also contains mean_confidence. The new gpt_confidence argument in report() includes the confidence in the report.

Adding solve_or_guess() and rank_one_diagnostic() to evaluate screening performance without requiring a gold-standard label set. solve_or_guess() estimates each rater’s (system or human) solving probability from patterns of agreement alone, and rank_one_diagnostic() checks whether that model’s core assumption holds for a given fit.

read_ris_to_dataframe() now repairs invalid UTF-8 characters to avoid data loss.

Adding tabscreen_mistral() and get_api_key_mistral() functions to screen titles and abstracts using Mistral’s API models.
Adding tabscreen_gemini() and get_api_key_gemini() function to screen titles and abstracts using Gemini’s API models.
Adding tabscreen_claude() and get_api_key_anthropic() function to screen titles and abstracts using Anthropics’s API models.

Migrating from chat/completions endpoint to responses for all OpenAI functions. This includes:
- tabscreen_gpt()
- screen_errors.gpt()
- rate_limits_per_minute()

Updated the prize function with up-to-date prizes for the GPT-5 models.
Adding the n_screened and n_missing variables to the key results in screen_analyzer().

Updating the default inclusion threshold and documentation hereof when conducting replicate screenings to be aligned with the finding from Vembye et al. (2025).
Updating the handling of coding missing abstracts in the vignette now when using read_ris_to_dataframe().
Better error messages for unknown GPT models when using newer or fine-tuned models.

Fixed bug in report() when rendering large amounts of text.
Fixed bug in tabscreen_gpt() when using multiple reps and gpt-5 models.
Fixed bug in screen_analyzer() when working with multiple prompts, models, and reps.
Correcting path in generating-disagreement-reports article.
Set max_tries in rate_limits_per_minute() to avoid message from httr2.

Adding tabscreen_groq() function to screen titles and abstracts using Groq AI.
Adding tabscreen_ollama() function to screen titles and abstracts using local ollama models.
Adding functions to read and write RIS files: read_ris_to_dataframe() and save_dataframe_to_ris().
Adding function to generate disagreement reports: generate_disagreement_report().
Making new refinements to the tabscreen_* functions. Making it possible to steer the model’s (over)inclusion behavior via the overinclusive = TRUE argument in tabscreen_* functions.

Adding articles for fine-tuning OpenAI models, generating disagreement reports, generating fine-tuning data and reading/writing RIS files.
Adding article for comparing performance of reasoning models (including gpt-5 models) with gpt-4o-mini.

Adding create_fine_tune_data() and write_fine_tune_data() to generate data for fine tuning OpenAI’s models.

A typo in the vignette has been corrected.
The vignette now draws on functions from synthesisr instead of revtools to handle RIS files.
tabscreen_gpt() now treats the study ID variable as a factor to keep original order of the dataset with titles and abstracts.