OpenAI and Anthropic PBC, two of the leading artificial intelligence model providers, today both introduced new large language models optimized for reasoning tasks.

OpenAI’s new algorithms, gpt-oss-120b and gpt-oss-20b, are available under an open-source license. Anthropic, for its part, released an upgraded version of its proprietary Claude Opus 4 large language model. The update improves upon the LLM’s coding capabilities, which the company claims already outperformed the competition.

Open-source performance

OpenAI says that gpt-oss-120b and gpt-oss-20b outperform comparably sized open models across multiple reasoning tasks. The former algorithm features 117 billion parameters, while the latter includes 21 billion. They can both run code, interact with external systems such as databases and optimize the amount of time they spend on a task based on its complexity.

“Proprietary API moats shrink; enterprises can now run and refine models in-house,” commented Dave Vellante, co-founder and chief analyst at theCUBE Research. “Differentiation in our view now rises to tools, RL loops, guardrails and — most importantly — data.”

Running gpt-oss-20b requires a single graphics card with 16 gigabytes of memory. That means the model is compact enough to run on certain consumer devices. The model is “ideal for on-device use cases, local inference, or rapid iteration without costly infrastructure,” OpenAI researchers wrote in a blog post today.

The company’s other new model, gpt-oss-120b, trades off some hardware efficiency for increased output quality. It can run on a single graphics card with 80 gigabytes of memory. The algorithm provides comparable performance to o4-mini, one of OpenAI’s newest and most advanced proprietary reasoning models.

Both gpt-oss-120b and gpt-oss-20b are based on a mixture-of-experts architecture. A mixture-of-experts model comprises multiple neural networks that are each optimized for a narrow set of tasks. When it receives a prompt, the model activates only the neural network that is best equipped to generate an answer.

OpenAI’s new models feature two performance-optimization features called grouped multi-query attention and rotary positional embeddings. The former technology reduces the memory usage of the algorithms’ attention mechanism, which they use to interpret user prompts. Rotary positional embeddings, in turn, make language models better at processing lengthy input. Both models support a context window of 128,000 tokens.

OpenAI developed gpt-oss-120b and gpt-oss-20b through a multistep process. First, it trained them on a dataset that mostly comprised English-language text about science and technology topics. OpenaI then carried out two more training runs that used supervised fine-tuning and reinforcement learning, respectively.

Supervised fine-tuning is carried out with training datasets that contain annotations explaining their contents. Reinforcement learning, in turn, doesn’t use annotations. The latter technique can be more cost-efficient because it reduces the amount of time that developers must spend organizing their datasets. 

“Irrespective of OpenAI’s intentions, open-weight reasoning models democratize frontier model capability but push the value conversation up the stack into enterprise agents, proprietary data, RL feedback efficacy and business context,” Vellante stated. “In our view, enterprises that build a digital-twin capability will program the most valuable agents; everyone else will fight for thinner slices of an ever-cheaper API.”

Claude Opus 4.1

Against the backdrop of OpenAI’s latest product update, rival Anthropic debuted a new proprietary LLM called Claude 4.1 Opus. It’s an upgraded version of the company’s flagship Claude 4 Opus reasoning model. Anthropic described the latter LLM as the “world’s best coding model” when it launched in May.

Claude Opus 4 scored 72.5% on SWE-bench Verified, a benchmark for measuring LLMs’ coding capabilities. The new Claude Opus 4.1 model achieved 74.5%. Additionally, Anthropic has improved the LLM’s research and data analysis capabilities.

Claude Opus 4.1 is available today in the paid versions of the company’s Claude AI assistant, as well as via its application programming interface, Amazon Bedrock and Google Cloud’s Vertex AI service. The update is the first in a planned series of enhancements to Anthropic’s LLM lineup. The company expects to release the other upgrades, which it describes as a “substantially larger,” in the coming weeks.

Photo: Focal Foto/Flickr