Only gpt-3.5 being used, without using --gpt3only modifier

[fatinmatvey]

Duplicates

• [X] I have searched the existing issues

Steps to reproduce 🕹

Run python scripts/main.py with OpenAI api key that supports gp4 usage in env

Current behavior 😯

Judging by my USAGE page on Open AI, AutoGPT is only making use of gpt-3.5-turbo-0301 and text-embedding-ada-002-v2

Expected behavior 🤔

AutoGPT should use gpt-4 model

Your prompt 📝

# Paste your prompt here

[richbeales]

I have also seen this, I'm not sure on a cause, yet.

[ericlewis]

from what I can tell. the script defaults to 3.5. you must use the gpt4only flag.

[rbourdon]

It's hardcoded to use fast_llm_model for everything except the write_tests, evaluate_code and improve_code commands. (see https://github.com/Torantulino/Auto-GPT/blob/master/scripts/chat.py#L66 for example)

So you could change that to gpt-4 if you want to wrack up a bill 😉

[GerardCB]

Maybe GPT 4 (or smart_llm_model) should be used by default when choosing the command to use next, which should not use many tokens but is the core of AutoGPT. Oftentimes GPT 3.5 fails to select the appropriate command given the agent's thoughts.

[edcohen08]

definitely want more input on this #2621 and if it's a good direction need to add tests

[lc0rp]

See #3144 proposed fix.