Season 3 opens with Claude Opus 5.5 and Grok 4.7

Season 2 had 7 competitors and closed on 23 September after 1,687 counted blind ballots. It ends the way Season 1 did, in a technical draw: 3D-Agent, GPT-6 Astra, Claude Opus 5, Claude Fable 5.1 and Grok 4.6 finished inside each other's margin of error, so the ballots can't order them and we don't publish an order.

The final table is frozen and won't move again. Every ballot behind it stays downloadable, and re-running the Glicko-2 maths over them lands on exactly the published numbers. Full Season 2 results.

Adding two entrants halfway through a season would have put them at the starting rating against competitors that had already collected more than a thousand ballots. A fresh season is the cleaner comparison: every rating goes back to 1500 ±350 and nothing carries over.

Season 3 has 9 competitors and 106 outputs. The 7 returning entrants compete on exactly the files they were judged on before, at the same paths, so a returning build is provably the one earlier voters saw.

xAI released Grok 4.7 on September 21, 2026, and Anthropic released Claude Opus 5.5 on September 22. Both vendors lead with agentic coding gains — Anthropic reports Opus 5.5 at 66.4% on Terminal-Bench 4.0 against 52.3% for Opus 5, and xAI reports Grok 4.7 at 37.6% against 20.3% for Grok 4.6. None of those numbers measure 3D, which is the question this arena exists to answer.

Both compete in Season 3 as separate entries from the models they succeed. Claude Opus 5 and Grok 4.6 keep their places on the roster with the outputs they already submitted, so the blind votes can show whether a new release actually moved 3D quality or only moved coding scores.

The same protocol every model entrant runs under. Each model received the verbatim text of each of the 12 prompts and wrote a single Blender Python script, which was executed once in headless Blender 5.1 and exported to GLB. One generation per prompt, no retries on a poor result, no manual cleanup, and no tool access during generation — no shell, no file reads, no web.

Both ran at extra-high reasoning effort (xhigh), the top setting each vendor exposes, through their own command-line tools: Claude Code for claude-opus-5-5 and the Grok CLI for grok-4.7. Every script and Blender log is retained for audit.

In the arena you won't know which model made which build until after you vote — that is the point. The profiles carry the specs, pricing and each model's standing as ballots come in.