GPT-6 Astra vs Opus 5.5: The End Of Chess

Summary
Chess commentator Levy Rozman (GothamChess) analyzes a two-game chess match between OpenAI’s GPT-6 Astra and Anthropic’s Claude Opus 5.5. Using an on-screen digital analysis board, he walks through both games move by move, evaluating their strategic plans, endgame conversions, tactical blunders, and illegal move attempts. He concludes by presenting the self-estimated Elo ratings generated by each AI and sharing his own rating assessment.

What is shown

• Match Introduction [00:00 - 01:23]: Levy introduces the contest between GPT-6 Astra and Claude Opus 5.5, discussing the recent evolution of LLM chess capabilities.
• Game 1 [01:24 - 14:34]: GPT-6 Astra plays White against Claude Opus 5.5 (Black) in a Nimzo-Indian Defense (1. d4 Nf6 2. c4 e6 3. Nc3 Bb4). White accepts an isolated queen’s pawn (IQP) setup. After White overlooks a tactical sequence and pushes pawns carelessly [11:15], Black seizes an endgame with outside passed pawns [13:50]. White resigns at [14:21].
• Game 2 [15:14 - 27:25]: Claude Opus 5.5 plays White against GPT-6 Astra (Black) in the Four Knights English Opening (1. c4 e5 2. Nc3 Nf6 3. Nf3 Nc6). Both sides contest the queenside [18:20]. Claude attempts illegal queen moves ([23:01], [26:42]) which Astra flags. Astra eventually blunders into hanging a rook [25:57], leading to White winning significant material and Black resigning at [26:52].
• Elo Ratings Discussion [28:03 - 30:41]: Review of post-match Elo self-estimations generated by Claude Opus 5.5 and GPT-6 Astra, followed by Levy’s evaluation of their performance.

Claims & numbers

• The presenter states this video was recorded in October 2026 as the third entry in a series testing frontier models [00:43].
• The presenter notes that GPT-6 Astra previously lost a match against Claude Fable [00:55].
• Claude Opus 5.5 estimated Astra played at 1200–1400 Elo and that it played at 1500–1700 Elo [28:14, 28:25].
• GPT-6 Astra estimated its own play at 1600 Elo in Game 1 and ~1100 Elo (900–1400 rapid range) in Game 2 [29:03, 29:08].
• GPT-6 Astra estimated Claude Opus 5.5 played at 2200 Elo in Game 1 and 2100 Elo in Game 2 [29:21, 29:28].
• The presenter estimates Astra played at around 1200–1300 Elo and Claude Opus 5.5 played around 1800 Elo [29:41, 29:44].

Notable quotes

• “Because for many years, ChatGPT, Claude, and others have been laughable at chess... They’re now good.” [00:18]
• “Astra resigned the game and said, ‘I’m down four pawns, I have no attack remaining, I resign.’” [14:26]
• “Claude made an illegal move. For the second time in a couple of videos, the same with Fable, at some point in the game it makes an illegal move. Astra never makes an illegal move.” [23:00]

Assessment
This is an independent gameplay review and demonstration by a chess creator testing frontier AI models prompted to play chess. The games were run off-camera through model prompts with moves manually entered onto an analysis board, and the presenter candidly highlights both impressive master-level positional concepts and persistent failure modes, such as Claude hallucinating illegal queen moves.

Described by gemini-3.8-flash on 2026-10-15 from the video’s audio and frames.