AI Radar

Noteworthy AI conversations and writing, summarized from publisher notes.

OpenAI GPT 4.1: New Developer-Focused Workhorse Models

· Latent Space

OpenAI's Michelle Pokrass (now leading post-training research) and Josh McGrath join Latent Space to discuss the launch of GPT 4.1, GPT 4.1 Mini, and the new GPT 4.1 Nano. The models are positioned as developer-focused workhorses with three core improvements: instruction following, coding, and a first-of-its-kind 1 million token context window. The team explains the confusing 4.5-to-4.1 naming jump by noting that 4.1 is significantly smaller and cheaper than 4.5, doesn't beat it on all benchmarks, and most developers can replace much of their 4.5 usage with 4.1. Josh McGrath details the long-context work, including two open-source evaluations (graph traversal and MRCR multi-needle retrieval) designed to test reasoning across the full context rather than simple needle-in-haystack tasks. On coding, GPT 4.1 scores 55 on SWE-bench (beating o1), with the team emphasizing post-training as the primary driver of gains rather than just pre-training scale. The conversation also covers vision improvements (especially in 4.1 Mini), preference fine-tuning, prompting best practices (XML for inputs, instruction placement), pricing updates (prompt caching discount raised from 50% to 75%), and the strategic rationale behind deprecating 4.5 to reclaim GPU compute.

AI Radar wrote this summary from the publisher’s notes with AI, not from a transcript. How it works.