Qwen3.8 Max
Alibaba's flagship Qwen model, served through Alibaba Cloud Model Studio with a 1M-token context window and a September 2 (0902) snapshot. It has the highest LMArena Vision score in this list and a top-ten score on the full LMArena Code Arena.
Alibaba (Qwen)Language modelReleased September 2026Data checked October 18, 2026
Where people rank it
#1
Web search and research (tied), 1,870 votes
#1
Reading images and documents (tied), 10,017 votes
#7
Coding and building apps, 13,812 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: Value #13, Reasoning #14 (tied), Speed #22.
Strengths and weaknesses
- Highest LMArena Vision score among the models here (statistically tied for first)
- Ranked 10th of 144 entries on the full LMArena Code Arena
- Moderate token price ($2 / $6 per million)
- Slow output on Artificial Analysis (38 tokens per second)
- Uses many tokens, so its Artificial Analysis cost per task ($5.41) is close to Claude Opus 5.5's
Best for
- Reading charts, documents and images
- Building web app prototypes
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)1,482.4Arena score 1482.4 (95% CI 1477.4 to 1487.3); 27,766 votes; entry "qwen3.8-max"; position 24 of 416 on the full board
- LMArena creative writing1,469Arena score 1469.0 (95% CI 1460.0 to 1478.0); 5,471 votes; entry "qwen3.8-max"; position 19 of 414 on the full board
- LMArena WebDev (coding)1,673.6Arena score 1673.6 (95% CI 1666.5 to 1680.7); 13,812 votes; entry "Qwen3.8-Max (0902)"; position 10 of 144 on the full board
- LMArena vision1,300.7Arena score 1300.7 (95% CI 1293.2 to 1308.1); 10,017 votes; entry "qwen3.8-max"; position 2 of 165 on the full board
- LMArena Agent Arena0.018IPS score 0.0176 (95% CI 0.0115 to 0.0237); 68,968 sessions; entry "Qwen3.8-Max"; position 25 of 52 on the full board
- Artificial Analysis Intelligence Index45.445.4 on the Artificial Analysis Intelligence Index v4.3.2 (AA shows 45), setting "Qwen3.8 Max (0902)"
- Humanity's Last Exam (Artificial Analysis)43.1%43.1% on Humanity's Last Exam (AA, 2,158 text-only questions), setting "Qwen3.8 Max (0902)"
- GPQA Diamond, graduate-level science92.8%92.8% on GPQA Diamond (AA; no longer part of the AA index since v4.2), setting "Qwen3.8 Max (0902)"
- Terminal-Bench 4.0 (Artificial Analysis run)38.9%38.9% on Terminal-Bench 4.0 as run by AA (mini-swe-agent harness), setting "Qwen3.8 Max (0902)"
- Long-context reasoning (AA-LCR)80.3%80.3% on AA-LCR (long-context reasoning), setting "Qwen3.8 Max (0902)"
- MMMU-Pro, image understanding82.8%82.8% on MMMU-Pro (image understanding, AA), setting "Qwen3.8 Max (0902)"
- AA-Omniscience, accuracy minus hallucination1212.0 on the AA-Omniscience Index (knowledge accuracy minus hallucination, range -100 to 100), setting "Qwen3.8 Max (0902)"
- Output speed, tokens per second38 t/s38 output tokens per second (AA median), row "Qwen3.8 Max (0902)"
- Cost per Artificial Analysis test task$5.41$5.41 to run one AA Intelligence Index task, row "Qwen3.8 Max (0902)"
- Artificial Analysis Coding Agent Index43.343.3 on the AA Coding Agent Index v1.5 (DeepSWE v1.1, Terminal-Bench 4.0, SWE-Atlas-QnA), best listed setup "Claude Code - Qwen3.8 Max"
- Terminal-Bench 4.0 leaderboard27.0%27.0% resolved on Terminal-Bench 4.0 (plus or minus 3.8, 95% CI) with Claude Code at max effort
Compare with
- GPT-6.1 SolOpenAI's mid-priced flagship, described by OpenAI as near-Astra performance at a lower cost.
- Claude Opus 5.5Anthropic's flagship for long-running agentic coding and knowledge work, and the model Anthropic suggests starting with for most jobs.
- 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 AstraOpenAI's most capable model, aimed at complex reasoning, coding, research and document creation.