OpenAI, led by CEO Sam Altman and VP of Research Aidan Clark, is developing its first open language model since GPT-2. The new “reasoning model” will be released with a highly permissive license in early summer this year, according to TechCrunch.
The company is responding to growing pressure from rivals like DeepSeek and Meta who offer open AI models. OpenAI will thoroughly test the model before making it available to the global developer community.
What Is an “Open” AI Model?
An open model means developers can see and modify the actual code inside the AI system – like sharing a recipe instead of just selling the finished cake. This is different from OpenAI’s recent approach where they kept their models private and only allowed access through limited interfaces.
Since 2019, OpenAI has kept its powerful models like GPT-3 and GPT-4 locked away, only letting people use them through restricted access points. The new model will change this by being “open-weight,” meaning developers can download and customize it.
Why OpenAI Is Changing Direction
Sam Altman recently admitted OpenAI has been “on the wrong side of history” regarding open-sourcing technology. This shift comes after competitors like DeepSeek released their R1 model, which cost just $5.6 million to train but performs impressively.
Key Features of the New Model
- Text-to-text design focused on reasoning abilities
- Ability to run on high-end consumer computers (not just data centers)
- Option to toggle reasoning capabilities on/off
- Highly permissive license allowing modifications
- Comprehensive documentation showing test results
Benchmark Targets vs. Competition
| Test Name | What It Measures | DeepSeek R1 Score | OpenAI Target |
|---|---|---|---|
| ARC-AGI | General reasoning | Below 87.5% | Above 87.5% |
| AIME | Math problems | 79.8% | Higher than 79.8% |
| GPQA-Diamond | Complex reasoning | 71.5% | Better than 71.5% |
Safety remains a top priority for OpenAI, with plans for extensive “red-team evaluation” – where experts try to find problems before release. The company will publish a detailed “model card” showing all test results.
This strategic shift marks a significant change in OpenAI’s approach to AI development. By embracing openness, the company hopes to balance innovation with accessibility, potentially transforming how researchers and developers build upon AI technology while maintaining their competitive edge in the rapidly evolving AI landscape.
