The release of OpenAI's new open-source model, GPT-OSS, received significant attention in the AI community.
Sam Altman announced that GPT-OSS performs at the level of GPT-4 mini and can run on a high-end laptop, with a smaller version capable of running on a phone.
Altman expressed pride in the technological achievement and encouraged continued progress.
Steven Adler, former OpenAI safety researcher, praised OpenAI for rigorous safety testing and transparency about which recommendations were adopted or not.
OpenAI was highlighted as the only leading AI company with a full commitment to such comprehensive safety testing.
Aiden Clark demonstrated GPT-OSS controlling a desktop: the model quickly and efficiently reorganized desktop files, operating at roughly 50–60 tokens per second in the demo.
The creator's own testing of the 20B parameter model on a high-end Mac Studio achieved similar speeds, with some performance questions remaining for larger versions.
Flavio Adamo showed the 20B model excelled at simulated physics tests, outperforming much larger models, though it struggled with more complex scenarios and some syntax errors.
Matt Schumer launched "GPTOSS Pro mode," chaining together up to 10 instances of the new model to deliver better answers, demonstrating collaborative multi-agent capabilities.
Together AI became one of the first to support GPT-OSS, offering the 120B model at 15 cents/million input tokens and 60 cents/million output tokens, citing strong speed and affordability.
The Together AI playground enables fast, accessible testing of the model for anyone, with observed speeds up to 123 tokens per second.
Industry opinion (Miles Brundage) affirmed the new model achieves strong performance and solid safety but called for more defined threat models and even smaller model variants for low-end devices.
Open-source release of model weights allows for community-led model quantization and further training, ensuring broad accessibility and flexibility.
Sam Altman hinted at additional major annoucements, raising speculation about further significant releases, possibly GPT-5.
Aaron Levy (Box CEO) commented that most future AI value will be in applications, not the model layer itself, with AI model pricing converging on infrastructure costs.
Rohan Pande (ex-OpenAI) clarified that pretraining the 20B model cost under $500,000, using 2.1 million H100 GPU hours and optimized algorithms to keep costs low.
Both the 20B and larger models use flash attention and other efficiency techniques.
Clarification given that GPT-OSS is not derived from Horizon models, with various users sharing side-by-side UI comparisons of model interfaces.
Theo GG introduced a humorous "snitchbench" benchmark to measure how likely different models are to report hypothetical corporate wrongdoing, with GPT-OSS 20B showing low rates of "snitching," compared to near-universal reporting by some leading commercial models.