Epoch measures why long context feels slower on GPT than Claude

Epoch AI measured something developers feel but rarely quantify: how long a model takes to return its first token as you put more into the context window. They tested four frontier models up to about a million tokens and found the two families behave differently. GPT-5.6 Terra and Sol scale quadratically, so each additional block of tokens costs more latency than the last. Claude Sonnet 5 and Opus 5 scale close to linearly, adding a roughly fixed penalty per token no matter how full the context already is.

The gap is small at today's sizes and large at tomorrow's. At a one-million-token context, adding another 10,000 tokens costs Sonnet 0.129 seconds and Sol 0.324 seconds, a real but survivable difference. Extrapolated to ten million tokens, the GPT models would add 1.66 to 2.20 seconds per 10,000 tokens while Claude's overhead stays flat. Epoch ran 262 requests and checked the result with three separate statistical methods, then reads the pattern as an architecture choice: Anthropic appears to spend effort on long-context efficiency, while OpenAI accepts the quadratic cost and leans on context compression, which is why Codex compacts before 272,000 tokens even though it supports far more. Their analysis is here.

Why it matters

If you build agents that carry long histories, this tells you where to keep memory: linear scaling makes it cheap to leave history in the context, while quadratic scaling pushes you toward external memory and aggressive compaction. Test first-token latency at your own context sizes before you commit to one model for a long-running loop.