How Close Are Open-Source Models to GPT-5-Class Performance? The 2026 State of Play

Open-source models are closing in on GPT-5-class performance, but the degree of similarity varies depending on the task. For tasks like retrieval, embeddings, and narrow focused work, open models often outperform proprietary options. However, when it comes to complex reasoning, multimodal capabilities, and long-running agent tasks, proprietary models still lead the way.

The term "GPT-5-class" is fluid, as OpenAI has released several iterations (5.1, 5.2, 5.4, 5.5, 5.6) since the launch of GPT-5 in August 2025. It is more accurate to compare open models to the current top-tier proprietary models rather than a specific version number.

An Artificial Analysis Intelligence Index, combining nine rigorous evaluations, shows that the best open-weight model trails behind the best proprietary model by about six points. However, this average hides considerable differences.

Open models excel in retrieval, narrow tasks, and specific narrow tasks such as OCR. They are particularly cost-effective for tasks like embeddings, where Qwen3-Embedding-0.6B costs $0.011 per million tokens, significantly less than hosted alternatives. For narrow generation tasks, open models like Qwen3.8-27B, which runs on a single 24 GB GPU, present an economical alternative to expensive API usage.

Yet, when a task demands the highest reasoning tier and the open sources fail evaluations, proprietary models are the better choice. The decision between open models and expensive APIs depends on utilization rates, weighing the per-token cost against the GPU cost and operational overhead of self-hosting.