frontier

A moving target rather than a fixed line. “Frontier” labels the capability envelope set by the top models from the labs investing the most in pretrainingⓘpretrainingtrainingThe first and most compute-expensive training phase, where a base model learns general capabilities by predicting the next token on trillions of words of web and book data. Chat about this → Open full entry →: OpenAI’s GPT and o series, Anthropic’s Claude family, Google’s Gemini, xAI’s Grok at the top end. Open-weightⓘopen weightsweightsA model release that publishes the trained parameters under some downloadable license, distinct from "open source" which (per OSAID) also requires data and training-code openness. Chat about this → Open full entry → families approach but rarely lead the frontier; DeepSeekⓘDeepSeekweightsA Chinese open-weight family known for the V3 MoE base model and the R1 reasoning model, both released under permissive licenses and unusually transparent in their training-cost reporting. Chat about this → Open full entry →-V3 was the clearest near-frontier open release through early 2025.

The term carries governance weight. The UK and EU regulatory regimes, US executive orders, and the Frontier Model Forum’s self-regulation all draw a line at “frontier” compute thresholds (often defined as training runs exceeding 10^25 or 10^26 floating-point operations) and apply specific reporting, evaluation, or safety-case obligations there.

For open AI the implication is unsettling. The capability gap between frontier closed models and frontier open weightsⓘopen weightsweightsA model release that publishes the trained parameters under some downloadable license, distinct from "open source" which (per OSAID) also requires data and training-code openness. Chat about this → Open full entry → has been roughly six to twelve months through 2024 to 2026. Whether the gap closes, widens, or holds depends on questions about open compute, open data, and whether frontier labs continue to release weights at all.