GPT-4o mini: advancing cost-efficient intelligence
Introducing our most cost-efficient small model
Via The Neuron
ChatGPT-4o mini just launched, and it’s perfect for fast requests.
|
| The newest version of ChatGPT is out. It’s called “ChatGPT 4o mini” and it’s apparently the most “capable and cost-efficient model” available today (try it yourself here). |
| Today’s launch comes at the end of a huge week for small models. Nowadays, everyone from Microsoft,Google, Mistral, Anthropic, to Cohere are all providing their own mini versions. |
|
|
| Why so many small models? To put it simply: smaller models = lower costs. Big companies need smaller models to use and offer AI at scale. Pilot projects and experiments are fine at today’s prices, but actual deployments? Wayyyy too expensive. |
- GPT-4o mini is 60% cheaper than 3.5 turbo to run (and it’s smarter, too).
- Sam Altman said that the best model from 2022 was way worse and “cost 100x more” than GPT-4o mini.
|
| Some speculated that this new launch might be OpenAI’s attempt to bury cheaper competitors like Claude Haiku or Gemini Flash. |
| Here’s why this is really important: Large models like GPT-4 cost a ton to build ($100M+), and so the people who make them charge more for their best models (that’s why they limit GPT-4o usage on the free plan). |
| Meanwhile, smaller models are trained on less data, are cheaper to run, and work fine for specific tasks.Some examples: |
- Mr. Cooper (a mortgage company) and TD Bank are testing midsize models for call centers to analyze voice data and improve efficiency.
- WPP (a marketing company) is using Google Flash for tasks like analyzing shopping habits and writing product descriptions.
|
| That said, we’ll still need big models for some more complex tasks, like discovering the formula behind the world’s most gripping mystery novels by analyzing every book written by Agatha Christie. |
|
|
| Our take? Use GPT-4o mini when speed matters more than quality. Here are a few tasks we tried it on: |
- Coded a Chrome extension to “show me pictures of cats” (<8 seconds)
- Wrote a children’s story about a tiny goldendoodle named Goldie (<5 seconds).
- Summarized an article and a whole email thread in crazy record time (<4 seconds).
- Counted the frequency of “yes”, “no” and “mid” in a spreadsheet dataset (<3 seconds).
|
| However, when we asked it to run the same spreadsheet test on 100 rows of data, it only counted ~300 total. So avoid GPT-4o mini for complicated tasks involving large datasets. |
0 Responses
Stay in touch with the conversation, subscribe to the RSS feed for comments on this post.