Sonnet 5.5 vs GPT 6.1 Sol chess match via MCP: 18x more reasoning tokens, $16 vs $2.37
adigrazia80 · reddit · 2026-10-04
A Reddit user wired Sonnet 5.5 (Claude Desktop) and GPT 6.1 Sol (Codex) into a local chess app via MCP, having each play a full game in a single conversation with no engine or legal-move hints. Score is 1-1 after two games.
Key numbers (both at Medium reasoning):
- Sonnet 5.5: 33m45s of claimed turns, 269,076 output tokens (224,770 reasoning), 57.44M cumulative input tokens, $16.21 API equivalent, 1 illegal move
- GPT 6.1 Sol: 21m12s, 34,956 output tokens (12,128 reasoning), 16.81M input tokens, $2.37, zero illegal moves
Subscription usage also diverged sharply: Claude's $20/month five-hour budget went from 23% to 56% during the match, while Codex's weekly display on the $200/month plan stayed at 5%. Sonnet called its pawn promotion "unstoppable," later admitted missing the Bf3 defense, and had one move rejected then self-corrected; Codex also misread a passed pawn once. The author cautions two games prove nothing about playing strength and plans a rematch with colors swapped.
More from Fun
- Having an AI assistant answer debt-collection calls? 'Billion dollar startup idea' — AIandDesign · 2026-10-04
- AI researcher Pedro Domingos' airport joke about an 'exterminate humanity' symposium — pmddomingos · 2026-10-04
- Yacine teases dragging a nonexistent 'Opus 5.5' out of distribution and forcing it to think — yacineMTB · 2026-10-04
- Laptop running Claude Code caught fire — literally 'burning tokens' — Miles_Brundage · 2026-10-04
- Meme: ChatGPT requests access to the neurons in your ventral tegmental area — ZeroStateReflex · 2026-10-04
- Redditor asks Grok, ChatGPT, Claude, Gemini and DeepSeek to draw themselves — Outrageous-Ad-9080 · 2026-10-04