People are talking about continual learning being solved effectively through brute force, by giving LLMs extremely large context windows. How far will this go in 2026?

By top three AI labs, I mean Anthropic, OpenAI, and Google. So far the largest context window available from these companies is GPT 5.4 with a 1.05M token context, with Anthropic and Google both having models with 1M. Previously Google has also offered models with 2M token context (Gemini 1.5, which didn't seem to make good use of the large context window), but for this question I'll consider only models released during 2026.

I will resolve based on the largest context windows which are publicly available, whether through API, chat, agentic coding services, etc. I won't evaluate how effectively models can actually use their advertised context windows, although I hope we get more robust benchmarks for this in general.