GitHub Copilot moved to usage-based billing on June 1, restructuring every plan around “AI Credits” priced at $0.01 each and mapped to the underlying model’s API token rates for input, output, and cached tokens (GitHub Blog). Code completions remain included without drawing credits, but agent sessions, chat turns, and Copilot Extensions calls are now metered against a monthly pool, with Pro+ receiving $39 in monthly credits and Pro receiving $10 (GitHub Blog). The change drew immediate pushback from developers, with GitHub’s announcement thread accumulating roughly 900 downvotes and 400 comments within hours, largely from teams that had benchmarked throughput under the previous unlimited-completion model (TechCrunch).
In modeling, WindBorne released WeatherMesh 6, a transformer-based atmospheric model that the company reports matches the European Centre for Medium-Range Weather Forecasts on several key variables, with five-day forecast accuracy comparable to what ECMWF’s one-day output previously achieved, running on consumer hardware and fed by roughly 400 active sensor balloons across 15 global sites (TechCrunch). The architecture applies the same next-token-prediction transformer pattern used in language models to atmospheric state prediction, a data point for the performance of domain-specific transformers trained on structured, high-quality data (TechCrunch). Separately, OpenAI’s Codex “Pro 2x” promotional capacity period ended May 31, returning $100/month Pro-plan users to standard rate limits and affecting any capacity assumptions built around the higher promotional throughput (OpenAI Developers).