Q3 2023 — Q2 2026 / 36 models + Q3 2026 to date
One bar per quarter for each of three camps: OpenAI's flagship, Anthropic's flagship, and the leading Asian lab's flagship. Values are the maximum advertised context window of the most capable model each camp had shipped by the end of that quarter. Hover any bar for the details. The twelve solid columns are completed quarters through 30 June 2026; the faded thirteenth is Q3 2026 so far, current to 25 July.
Flagship, not largest window. Each bar is the camp's most capable model that quarter, then its context window — not the biggest window the camp could field. That's why the line isn't always upward.
Maximum advertised. Where a window is tiered, the top tier is plotted and flagged in the readout. GPT‑5.4 is 272K standard and 1M through the API and Codex; the 1M figure is plotted.
"Leading Asian lab" is a judgment call. No single Chinese lab held the frontier throughout, so the pick moves between Alibaba, Moonshot and DeepSeek. Reasonable people would draw some quarters differently.
Advertised ≠ effective. Every model benchmarked so far degrades on recall well before its stated limit.
Two step changes, not a smooth curve. 32K to 128K over late 2023, then a long plateau at 128K–200K that lasts nearly two years, then everyone lands on 1M inside a single quarter.
Anthropic sat at 200K for nine straight quarters — Claude 2.1 through Opus 4.5 — before Opus 4.6 moved the Opus line to 1M in February 2026.
Q1 2026 is the convergence point. GPT‑5.4, Opus 4.7 and Qwen 3.6 Plus all advertise roughly 1M within weeks of each other, and the number stops being a differentiator.
Q2 2026 dips on the Asian bar because the quarter's strongest Asian model, Kimi K2.6, ships 256K. MiniMax M3 hit 1M the same quarter but wasn't the capability leader.
Three weeks into Q3 2026, the gap has closed entirely. Kimi K3 (16 July), Claude Opus 5 (24 July) and GPT‑5.6 Sol all sit at 1M or just over. Every camp is now flat at the same number, and the next axis of competition is effective recall and price per filled window, not the headline figure.
Compiled 25 July 2026. Pre‑2025 figures are from vendor model cards and release posts at the time.