Epoch AI compared how OpenAI's and Anthropic's flagship models handle very long prompts, and found GPT-5.6 Terra and Sol slow down quadratically as context grows while Claude Sonnet 5 and Opus 5 stay almost linear -- the marginal latency to process the next 10,000 tokens for GPT-5.6 Terra rose from 0.071 seconds at a 100,000-token prompt to 1.66 seconds at a 10-million-token prompt

Epoch AI measured time to first token for OpenAI's GPT-5.6 Terra and Sol and Anthropic's Claude Sonnet 5 and Opus 5 as prompts grew longer. For GPT-5.6 Terra, the extra latency to process another 10,000 tokens rose from 0.071 seconds at a 100,000-token prompt to 1.66 seconds at a 10-million-token prompt. Epoch found the GPT models scaled quadratically while the Claude models stayed close to linear.