GPT-6 Astra
OpenAI's most capable model, aimed at complex reasoning, coding, research and document creation. It leads ARC-AGI-2, GPQA Diamond and Scale's Humanity's Last Exam board, yet people rate its chat answers only mid-table on LMArena.
OpenAILanguage modelReleased September 2026Data checked October 18, 2026
Where people rank it
#2
Coding and building apps (tied), 6,014 votes
#5
Web search and research (tied), 18,483 votes
From blind votes on LMArena, where people pick the better of two anonymous answers. See the full tables on the AI leaderboard.
On paper
Fixed-test rankings from the benchmarks tab: Reasoning #5, Value #10, Speed #18.
Strengths and weaknesses
- Best ARC-AGI-2 score in this list (95.0%) and the top GPQA Diamond result measured by Artificial Analysis (96.1%)
- Second-highest score on both the LMArena Code Arena and the LMArena Agent Arena
- First on Scale AI's Humanity's Last Exam leaderboard (54.8%, plus or minus 1.9)
- Expensive ($10 / $50 per million tokens) and slow (57 tokens per second on Artificial Analysis)
- Only mid-table on the LMArena text and creative writing leaderboards
Best for
- Hard math and science problems
- Complex coding projects
- Deep research reports
Every score we have
Each bar compares this model with the best result any model on this page has on the same test.
- LMArena text arena (human votes)1,477.3Arena score 1477.3 (95% CI 1471.0 to 1483.6); 12,163 votes; entry "gpt-6-astra-max"; position 30 of 416 on the full board
- LMArena creative writing1,449.9Arena score 1449.9 (95% CI 1437.7 to 1462.2); 2,718 votes; entry "gpt-6-astra-max"; position 38 of 414 on the full board
- LMArena WebDev (coding)1,777.2Arena score 1777.2 (95% CI 1766.9 to 1787.5); 6,014 votes; entry "GPT-6 Astra (Max)"; position 2 of 144 on the full board
- LMArena vision1,286Arena score 1286.0 (95% CI 1275.2 to 1296.8); 3,590 votes; entry "gpt-6-astra-max"; position 19 of 165 on the full board
- LMArena Agent Arena0.141IPS score 0.1413 (95% CI 0.1183 to 0.1643); 13,168 sessions; entry "GPT-6 Astra (Max)"; position 2 of 52 on the full board
- Artificial Analysis Intelligence Index52.752.7 on the Artificial Analysis Intelligence Index v4.3.2 (AA shows 53), setting "GPT-6 Astra (max)"
- Humanity's Last Exam (Artificial Analysis)54.7%54.7% on Humanity's Last Exam (AA, 2,158 text-only questions), setting "GPT-6 Astra (max)"
- GPQA Diamond, graduate-level science96.1%96.1% on GPQA Diamond (AA; no longer part of the AA index since v4.2), setting "GPT-6 Astra (max)"
- Terminal-Bench 4.0 (Artificial Analysis run)59.1%59.1% on Terminal-Bench 4.0 as run by AA (mini-swe-agent harness), setting "GPT-6 Astra (max)"
- Long-context reasoning (AA-LCR)80.7%80.7% on AA-LCR (long-context reasoning), setting "GPT-6 Astra (max)"
- MMMU-Pro, image understanding86.9%86.9% on MMMU-Pro (image understanding, AA), setting "GPT-6 Astra (max)"
- AA-Omniscience, accuracy minus hallucination43.443.4 on the AA-Omniscience Index (knowledge accuracy minus hallucination, range -100 to 100), setting "GPT-6 Astra (max)"
- Output speed, tokens per second57 t/s57 output tokens per second (AA median), row "GPT-6 Astra (max)"
- Cost per Artificial Analysis test task$3.26$3.26 to run one AA Intelligence Index task, row "GPT-6 Astra (max)"
- Artificial Analysis Coding Agent Index61.661.6 on the AA Coding Agent Index v1.5 (DeepSWE v1.1, Terminal-Bench 4.0, SWE-Atlas-QnA), best listed setup "Codex - GPT-6 Astra (max)"
- Terminal-Bench 4.0 leaderboard58.2%58.2% resolved on Terminal-Bench 4.0 (plus or minus 2.8, 95% CI) with Codex at max effort
- ARC-AGI-2 puzzles95.0%95.00% on ARC-AGI-2 (semi-private set), best listed setting "GPT-6 Astra (Max)", $1.12 per task
- ARC-AGI-3 puzzles62.7%62.71% on ARC-AGI-3 (semi-private, standard harness; "provider adapter" runs excluded), best listed setting "GPT-6 Astra (Max)"
- Humanity's Last Exam (Scale AI)54.8%54.80% on Scale AI's Humanity's Last Exam leaderboard (plus or minus 1.94), entry "GPT 6 Astra"
Compare with
- Claude Fable 5.1Anthropic's top tier for the hardest reasoning and long-horizon agent work, which Anthropic recommends when Opus 5.5 still falls short.
- GPT-6.1 SolOpenAI's mid-priced flagship, described by OpenAI as near-Astra performance at a lower cost.
- Qwen3.8 MaxAlibaba's flagship Qwen model, served through Alibaba Cloud Model Studio with a 1M-token context window and a September 2 (0902) snapshot.
- Claude Opus 5.5Anthropic's flagship for long-running agentic coding and knowledge work, and the model Anthropic suggests starting with for most jobs.