Muse Spark 1.3
Meta's mid-priced Muse model, trained for agentic workflows and tuned for competitive coding, available through the Meta Model API. It is close to the leaders on LMArena text and ties for first on LMArena Vision at a fraction of flagship prices.
MetaLanguage modelReleased September 2026Data checked October 18, 2026
Where people rank it
#2
Chat and everyday questions (tied), 17,561 votes
#2
Reading images and documents (tied), 4,550 votes
#9
Coding and building apps, 9,068 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 #6, Speed #6, Reasoning #9.
Strengths and weaknesses
- Close behind the leaders on the LMArena text leaderboard, ahead of every OpenAI model
- Tied for first on LMArena Vision and 93.5% on GPQA Diamond
- Low price ($1.25 / $4.25 per million tokens) with a 1M-token context window
- Weak at agentic terminal tasks (14.6% on Terminal-Bench 4.0 in Muse Code)
- Trails Meta's own Muse Wonder on reasoning (48.7% vs 50.1% on Humanity's Last Exam)
Best for
- An everyday assistant at a low price
- Understanding images and documents
- Quick 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)1,494.4Arena score 1494.4 (95% CI 1488.8 to 1500.0); 17,561 votes; entry "Muse Spark 1.3 (Max)"; position 10 of 416 on the full board
- LMArena creative writing1,461.2Arena score 1461.2 (95% CI 1450.6 to 1471.7); 3,844 votes; entry "Muse Spark 1.3 (Max)"; position 29 of 414 on the full board
- LMArena WebDev (coding)1,658.4Arena score 1658.4 (95% CI 1650.5 to 1666.4); 9,068 votes; entry "Muse Spark 1.3 (Max)"; position 12 of 144 on the full board
- LMArena vision1,289.7Arena score 1289.7 (95% CI 1279.9 to 1299.5); 4,550 votes; entry "muse-spark-1.3-max"; position 12 of 165 on the full board
- LMArena Agent Arena0.038IPS score 0.0382 (95% CI 0.0299 to 0.0465); 73,912 sessions; entry "Muse Spark 1.3 (Max)"; position 14 of 52 on the full board
- Artificial Analysis Intelligence Index48.148.1 on the Artificial Analysis Intelligence Index v4.3.2 (AA shows 48), setting "Muse Spark 1.3 (max)"
- Humanity's Last Exam (Artificial Analysis)48.7%48.7% on Humanity's Last Exam (AA, 2,158 text-only questions), setting "Muse Spark 1.3 (max)"
- GPQA Diamond, graduate-level science93.5%93.5% on GPQA Diamond (AA; no longer part of the AA index since v4.2), setting "Muse Spark 1.3 (max)"
- Terminal-Bench 4.0 (Artificial Analysis run)33.3%33.3% on Terminal-Bench 4.0 as run by AA (mini-swe-agent harness), setting "Muse Spark 1.3 (max)"
- Long-context reasoning (AA-LCR)83.0%83.0% on AA-LCR (long-context reasoning), setting "Muse Spark 1.3 (max)"
- AA-Omniscience, accuracy minus hallucination2525.0 on the AA-Omniscience Index (knowledge accuracy minus hallucination, range -100 to 100), setting "Muse Spark 1.3 (max)"
- Output speed, tokens per second134 t/s134 output tokens per second (AA median), row "Muse Spark 1.3 (max)"
- Cost per Artificial Analysis test task$1.60$1.60 to run one AA Intelligence Index task, row "Muse Spark 1.3 (max)"
- Artificial Analysis Coding Agent Index54.354.3 on the AA Coding Agent Index v1.5 (DeepSWE v1.1, Terminal-Bench 4.0, SWE-Atlas-QnA), best listed setup "Muse Code - Muse Spark 1.3 (max)"
- Terminal-Bench 4.0 leaderboard14.6%14.6% resolved on Terminal-Bench 4.0 (plus or minus 2.9, 95% CI) with Muse Code at xhigh effort
Compare with
- 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-6.1 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.