Google Gemini 1.5 Pro (1M-token context)

Adopt with guardrails

Why

The 1M-token jump reset the industry's context baseline (from ~200K) and is the direct technical ancestor of every long-context claim every later model (including pragma.vision's own model choices) makes — but the specific artifact is retired and its own technical report shows the capability degrades on realistic multi-fact tasks, so the register should credit the breakthrough while flagging both obsolescence and the underlying reliability ceiling.

Next

Treat 1M+ token context windows as a now-standard, low-friction capability for long-document/codebase/video ingestion — but always target current-generation models (Gemini 2.5 Pro/3.x or later; never the retired gemini-1.5-pro/-flash aliases, dead since 2025-09-29). Add explicit multi-fact retrieval verification (not raw single-needle recall%) before shipping any long-context flow that replaces chunked RAG, since accuracy on multi-needle extraction still sits at 60-70% even on this lineage's most-tested model.