Muse Spark 1.3 2026: what the AA Intelligence Index 61 means
Muse Spark 1.3 enters AA’s tie band with Grok and GPT. Split xhigh from max before you rank it.
Meta shipped Muse Spark 1.3 on 2 September 2026. The next day OfficeChai matched it to Artificial Analysis: the xhigh setting scored 61, tying Grok 4.6 and GPT-5.6 Sol — a third-party index, not another lab-owned table.
The query is Muse Spark 1.3 and the Artificial Analysis Intelligence Index. What follows checks the AA release page, OfficeChai’s 3 Sep 2026 chart, and CryptoBriefing / OpenRouter on the 2 Sep launch. An index point is not a general crown.
What the index page prints
AA marks 1.3 as a proprietary September 2026 release with weights “Not publicly available.” Intelligence Index v4.1.1 blends nine evals: GDPval-AA v2, τ³-Banking, Terminal-Bench v2.1, SciCode, Humanity's Last Exam, GPQA Diamond, CritPt, AA-Omniscience, AA-LCR. The card lists text / image / video in, text out, and a ~1M-token context.
How to read the chart
- 1
A third-party composite, not a vendor harness
OfficeChai describes coverage of knowledge work, agents, coding, and scientific reasoning. That is the evaluator’s blend, not an unpublished Meta suite. Same rule as our Gemini 3.7 Flash AA note: name the index before the product claim.
- 2
Out of the “next tier” cluster
OfficeChai: about 57 for 1.2, 51 for 1.1, 43 for the April original; the 4-point 1.2→1.3 step is the family’s largest. On that chart Fable 5.1 (max+fallback) is 66 and Opus 5 (max) is 63; xhigh ties Grok 4.6 (high) and GPT-5.6 Sol (max) at 61. Kimi K3 and GLM-5.3 sit at 60; Gemini 3.8 Flash (high) at 59. Safety evals: Fable 5.
- 3
Speed and task-price are separate columns
AA gives xhigh ~182 tokens/s and about $0.55 per Intelligence Index task. OfficeChai cites a faster speed chart (235 tokens/s); this page uses the AA release page. CryptoBriefing relays ~20% fewer tool calls and ~25% fewer tokens versus 1.2 — vendor efficiency language, not an AA reproduction.
Where it is live
CryptoBriefing: Muse Code (beta coding agent) and the Meta Model API. OpenRouter lists meta/muse-spark-1.3 dated 2 Sep 2026 at about $1.25 / $4.25 per million input/output tokens, 1M context.
Max reasoning is not a full public SKU yet
CryptoBriefing says a “max reasoning” mode still needs more safety testing. AA already prints a max row at 62, with price and speed blank on the release page. A chart row is not an API product everyone can buy.
Control is a different axis
A higher index does not answer controllable deployment. See Guidelight. Weights are closed, so this is not an open-checkpoint audit.
| Claim | Checkable source | How to read it |
|---|---|---|
| xhigh 61 / max 62 | AA Muse Spark 1.3 release page | Split the effort rows first |
| Tie at 61 with Grok 4.6 and GPT-5.6 Sol | OfficeChai 2026-09-03 | That article’s chart, not AA’s full rank dump |
| Shipped 2 Sep; fewer tools/tokens | CryptoBriefing / OpenRouter | Date checks; 20%/25% is vendor language |
OfficeChai: 1.3 is the first Muse Spark release that ties OpenAI and xAI’s top rows instead of sitting in the next-tier cluster.
# Muse Spark 1.3 · Artificial Analysis
# release: 2026-09-02 (CryptoBriefing / OpenRouter)
family: April → 1.1 Jul → 1.2 Aug → 1.3 Sep
index: AA Intelligence Index v4.1.1 (9 evals)
xhigh: 61 · max: 62
xhigh_speed_aa: 182 tok/s
xhigh_cost_per_index_task: $0.55
weights: not publicly availableBoundaries
- A composite ≠ a single-task crown
- Nine weighted evals. Coding or science slices can still trail higher-scoring Opus / Fable rows.
- A tie ≠ interchangeable APIs
- Latency, price, tool traces, and content policy differ. OpenRouter list prices are not AA’s dollars-per-task column.
- This page is not a meeting brief
- Release and eval only. For a short shared board, open wbmeet and create or join a room.
Questions worth checking
Is 61 or 62 the “official” score?
Both are AA rows for 1.3. OfficeChai’s “caught the frontier” line uses xhigh=61. Write the effort setting on any buy or reproduce note.
Why do speed numbers disagree?
The AA release page lists ~182 tokens/s for xhigh. OfficeChai cites 235. This page quotes AA and labels 235 as a different chart.
Can we self-host the weights?
AA says not publicly available. Today that means hosted API / Muse Code, not a Hugging Face checkpoint.
Does this prove Meta passed Anthropic?
No. On the same OfficeChai chart Fable 5.1 is still 66 and Opus 5 is 63. 1.3 entered the tie band; it did not take the top row.