The Million-Token Standard
Six frontier releases in six weeks, and the spec sheet quietly converged: a million tokens of context is the new floor. What that changes about building with these models.
July 2026 compressed a year of releases into six weeks: Gemini 3.6 Flash on the 21st, Claude Opus 5 on the 24th, DeepSeek-V4-Flash on the 31st, with Grok 4.5, GPT-5.6 Sol and Qwen 3.8 Max landing around them [1]. Strip the branding and one spec repeats: context windows at or above one million tokens.
A million tokens is not a bigger clipboard. It is a different architecture for applications. Retrieval pipelines built to work around 128K limits — chunking, embedding, re-ranking — become optional for a large class of tasks: an entire codebase, a quarter of company documents, or a full legal case file now fits in one call. The engineering question flips from 'how do we squeeze context in' to 'what is worth the $5 it costs to send it.'
The spread is in the hands, not the head
Benchmark spreads on reasoning are narrowing; the visible differences are agentic. OSWorld 2.0 (operating a real desktop) and SWE-bench Pro (real repository fixes) show double-digit gaps between labs [2], where math and knowledge benchmarks show points. For builders, model choice in late 2026 is less about intelligence and more about which vendor's hands you trust inside your systems — and at $25 per million output tokens, how often you can afford to let them work.