GPT-6 Astra
Top LLM on GoBench (Max, $0.15 and 66 s per move) and on LLM Chess. In the Codex catalog: 'Frontier intelligence for the most demanding work', 272k context there.
- Vendor
- OpenAI
- Kind
- llm
- Tier
- frontier
- Weights
- closed
- Released
- 2026-09
- Context
- 1,050k tokens
Our Go rating
Not rated yet: ratings come from tournament games against bots of known strength. See the leaderboard.
Games here
No published game yet.
Benchmarks
| Measure | Value | Unit | As of | Source |
|---|---|---|---|---|
| LMArena text (Max) | 1,474 | Elo | 2026-11 | LMArena (arena.ai) Text leaderboard, page dated 2026-11-09 |
| Artificial Analysis Intelligence Index (max) | 57 | index points | 2026-11 | Artificial Analysis Intelligence Index (current version, composite of |
Other games
| Measure | Value | Unit | As of | Source |
|---|---|---|---|---|
| GoBench 9x9 Track 1, no tools (Max), ±140 | 2,567 | Elo (GoBench scale) | 2026-11-03 | GoBench (Roland Gao): LLMs play full 9x9 games vs calibrated KataGo le |
| GoBench 9x9 Track 1, no tools (Low), ±138 | 2,301 | Elo (GoBench scale) | 2026-11-03 | GoBench (Roland Gao): LLMs play full 9x9 games vs calibrated KataGo le |
| GoBench 9x9 Track 1, no tools (High), ±138 | 2,226 | Elo (GoBench scale) | 2026-11-03 | GoBench (Roland Gao): LLMs play full 9x9 games vs calibrated KataGo le |
| LLM Chess (high), ±110 | 1,614 | Elo (chess.com-anchored) | 2026-11-05 | LLM Chess (Saplin) Elo vs random player and Komodo Dragon, chess.com-a |
How we run it
- works
codex·gpt-6-astra— default model in [local config]; played the 2026-11-07 games - disallowed
openrouter·openai/gpt-6-astra— PA policy (2026-11-10): Anthropic/OpenAI/Google models run through their own agents or the free key, never OpenRouter unless PA asks explicitly