GPT 4.1: The New OpenAI Workhorse

Michelle Pokrass said the release includes GPT 4.1, GPT 4.1 Mini, and GPT 4.1 Nano, with the goal of being “great for developers.” She highlighted improvements in instruction following and coding, and said they shipped their first one-million-context models.

Michelle Pokrass said GPT 4.1 is a big improvement over the 4o line, but is smaller/cheaper than GPT 4.5 and therefore doesn’t match GPT 4.5 on certain intelligence evals. She added that for most developers, 4.1 can replace much of their 4.5 usage.

Michelle Pokrass said Nano is a new pre-train, Mini is also a new pre-train, and the larger 4.1 is a new mid-train, but emphasized that “a significant amount of the gains come from new post-training techniques.” She framed this as squeezing more performance from post-training rather than only scaling pre-training.

Josh McGrath said “needle in a haystack” style retrieval is often easy for models, and that the harder part of long context is complex context use and reasoning. He described two open-sourced evaluations: one focused on reasoning about ordering and another on walking through graphs.

Josh McGrath explained Graph Walks as encoding a graph via edge lists in context and asking the model to perform an operation, then evaluating it against the true executed result. He said early models could loop or get stuck when unsure how to use context, and that the benchmark is an idealized lower bound for multi-hop reasoning across many documents with implicit connections.