GPT

GPT is OpenAI’s family of large language models. The line starts with OpenAI’s June 11, 2018 work on improving language understanding with unsupervised learning, which combined two existing ideas, transformers and unsupervised pre-training: a model is first trained on a large amount of text with language modeling as the training signal, then fine-tuned for specific tasks.¹ GPT-3, described in a paper first posted on May 28, 2020, had 175 billion parameters.² OpenAI’s API catalog still lists older GPT-3.5, GPT-4 and GPT-5 versions, and as of October 13, 2026 its featured models are GPT-6 Astra (“our most capable model for the most demanding work”), GPT-6.1 Sol (“near-Astra performance for complex work at a lower cost”) and GPT-6 Luna (for focused, high-volume tasks).³ Astra has a 1,050,000-token context window, up to 128,000 output tokens and an April 30, 2026 knowledge cutoff, by OpenAI’s own model page.⁴

GPT models hold the two highest scores on ARC-AGI-3’s Semi-Private set on ARC Prize’s Standard harness: GPT-6 Astra at 62.71% and GPT-6.1 Sol at 52.73%, measured by ARC Prize (Highest ARC-AGI-3 Scores, by Model).⁵ On SWE-bench Verified, GPT 5.2 and GPT 5.2 Codex each resolved 72.8% of tasks in the shared mini-SWE-agent v2.0.0 harness, a benchmark OpenAI itself stopped reporting in February 2026, citing flawed tests and contamination (SWE-bench Verified, Pro and What Replaced Them).⁶⁷ Prices run from $0.50 per 1M output tokens for GPT-6 Luna to $50 for GPT-6 Astra, both vendor-stated list prices at the short-context, standard tier (Cheapest LLM APIs per 1M Output Tokens).⁸ On OpenRouter, GPT-6 Luna processed 7.11 trillion tokens in the week to October 12, 2026, ranking seventh, with GPT-6.1 Sol at 2.23 trillion and GPT-5.6 Luna at 1.81 trillion (Most-Used AI Models on OpenRouter by Weekly Tokens).⁹