ChatGPT users may now try a lightweight version of deep research, the generative AI option that produces relatively slow, thorough answers. Lightweight OpenAI o4-mini became available for ChatGPT Plus, Team, Pro, and free tiers as of April 24. It will become available to Enterprise and Education tier users next week, following the same usage rules as for Team users.
This marks the first time deep research answers from the AI have been available for users at the free tier.
Lightweight o4-mini will kick in after reaching a usage limit, depending on tier
Unlike the full-size version, the lightweight o4-mini does not appear in the model picker; instead, it activates once the user reaches the usage limit of the free version. For Team and Plus users, the rate limit is 25 monthly tasks, while for Pro users the rate limit is 250 tasks. Free users can get up to five answers from the lightweight o4-mini.
Responses from lightweight o4-mini will “will typically be shorter while maintaining the depth and quality you’ve come to expect,” OpenAI wrote on X (formerly Twitter) on Thursday.
The company showed an accuracy score based on the BrowseComp Agentic Browsing benchmark of 45.6% when using lightweight o4-mini compared to 51.5% when using deep research.
Tier scheme benefits OpenAI by moving users to lower-cost plans
The lightweight version of deep research is “significantly cheaper to serve,” OpenAI wrote. For the company, balancing increased usage with more powerful offerings is another part of the calculation Sam Altman referred to this week regarding users adding tokens to their prompts by being polite.
Reasoning models are the new normal in generative AI
Reasoning models show their work and can “think” through complex mathematics and science problems. Adding reasoning models to chatbots has become a common practice among major AI companies. Anthropic offers Claude 3.7 Sonnet and its variations. The Gemini 2.0 family from Google and Grok 3 family from xAI both include reasoning. DeepSeek made a name for itself partially because of the reasoning performance of its R1 model.