Frontier models: beyond the announcements

OpenAI launched the GPT-5.6 family on 9 July 2026 in Sol, Terra, and Luna variants and later announced price changes. Moonshot AI described Kimi K3 as a 2.8-trillion-parameter mixture-of-experts model with native vision and a 1-million-token context window, with full weights released later in July. Google's own July roundup lists Gemini 3.6 Flash, Gemini 3.5 Flash-Lite, and Gemini 3.5 Flash Cyber; it does not support the earlier claim that a product named Gemini 3.5 Pro launched on 17 July.

Compare release types, not slogans

An API model family, an open-weight release, and a provider's monthly product roundup create different choices. API services shift infrastructure responsibility to the provider. Open weights can increase deployment control, but require licence review, hardware planning, evaluation, monitoring, and safety controls.

Evidence boundaries

Performance tables published by model vendors are vendor claims until reproduced under a buyer's workload. Context-window size and parameter count do not by themselves predict task quality, latency, or total cost.

Enterprise checklist

Benchmark representative tasks, record exact model versions and reasoning settings, measure complete-task cost, test failure paths, and preserve a provider-switching route. Re-run the evaluation after price or model updates; do not infer production value from one leaderboard score.