Leading AI model developers OpenAI and Anthropic PBC today unveiled new large language models specifically optimized for complex reasoning tasks.

OpenAI's gpt-oss-120b and gpt-oss-20b models are released under open-source licenses, while Anthropic introduced an enhanced version of its proprietary Claude Opus 4 model. The upgraded model features improved coding capabilities that the company claims outperform competitors.

Open-Source Performance Breakthrough

OpenAI reports both gpt-oss-120b (117 billion parameters) and gpt-oss-20b (21 billion parameters) outperform comparable open-source models in reasoning tasks. These models support code execution, external system integration, and dynamic time optimization based on task complexity.

"Proprietary API moats are shrinking," noted Dave Vellante, co-founder and principal analyst at theCUBE Research. "Now that enterprises can run and optimize models internally, differentiation shifts to tools, RL loops, guardrails, and most importantly data."

The gpt-oss-20b requires a 16GB GPU, making it compact enough for consumer devices. OpenAI highlights this model as "ideal for on-device applications, local inference, or cost-effective rapid iteration" in their blog announcement.

For higher output quality, the gpt-oss-120b requires an 80GB GPU but delivers performance comparable to OpenAI's cutting-edge o4-mini model. Both models employ Mixture of Experts architecture, activating only relevant neural networks for each query.

Key optimizations include grouped multi-query attention (reducing memory usage) and rotary position embeddings (enhancing long-text processing). Both models support 128,000-token context windows through OpenAI's development process combining scientific/technical training datasets with supervised fine-tuning and reinforcement learning.

Supervised fine-tuning uses annotated datasets for clearer outputs, while reinforcement learning reduces data annotation costs. "Open-weight inference models democratize AI capabilities but shift value creation to enterprise agents, proprietary data, RL efficiency, and business contexts," Vellante explained. "Companies building digital twin capabilities will create most valuable agents while others compete for diminishing API slices."

Claude Opus 4.1 Launch

Anthropic responded to OpenAI's announcements with Claude Opus 4.1 - an upgraded version of their flagship coding model. This update improves benchmark scores from 72.5% to 74.5% on SWE-bench Verified while enhancing research and data analysis capabilities.

Available through Claude AI premium plans, Anthropic APIs, Amazon Bedrock, and Google Cloud Vertex AI, this update marks the first in a series of planned LLM enhancements. The company anticipates releasing "significantly larger" upgrades in coming weeks.