DK Dark Knight 2026-09-02 5 min read

Muse Spark 1.3 Beats Gemini 3.8 on Artificial Analysis

Meta dropped a new frontier model today, and it doesn't sit where the headline will tell you. Muse Spark 1.3 (max) scores 62.09 on the Artificial Analysis Intelligence Index — ranked #6 of 636 models, ahead of every Google model on the board. I went to run it through the API and hit a wall that says more about frontier-model distribution in 2026 than the benchmark does: you have to click an 18+ age-confirmation checkbox before Meta will let an agent call it.

Muse Spark 1.3 Rankings: #6 on the Intelligence Index

Released September 2, 2026 (artificialanalysis.ai lists the exact date), Muse Spark 1.3 is Meta's proprietary reasoning model — text, image, and video input, text output, a 1M-token context window, and an "xhigh" reasoning effort tier that burns 2,000 reasoning tokens up front. On Artificial Analysis it lands at #6/636 with an index of 62. That puts it directly between Anthropic's pair at the top and everything else:

graph TD
    A["Claude Fable 5.1 (max) — 65.65"] --> B["Claude Opus 5 (max) — 63.05"]
    B --> C["Muse Spark 1.3 (max) — 62.09 ← today's release"]
    C --> D["GPT-5.6 Sol (max) — 60.93"]
    D --> E["Grok 4.6 (high) — 60.92"]
    E --> F["Gemini 3.8 Flash (high) — 58.68"]
    style C fill:#a78bfa,stroke:#a78bfa,color:#0a0a0b

The specific scalp: Gemini 3.8 Flash — the only Gemini 3.8 model OpenRouter lists — scores 58.68. Muse Spark beats it by 3.4 points. That's the claim the Hacker News thread is running with, and the numbers hold. But note the ceiling: Claude Fable 5.1 (max), which shipped yesterday, holds the #1 slot at 65.65. Meta has caught Google and is chasing Anthropic, not leading.

Muse Spark 1.3 Pricing: $1.25 In, $4.25 Out

On OpenRouter, meta/muse-spark-1.3 costs $1.25 per million input tokens, $4.25 per million output — a 1M-context multimodal frontier model at roughly the price of a value meal. There's also a meta/muse-spark-1.3-contributor variant at $0.10/$0.20 per million, a sign Meta is already experimenting with subsidized tiers for the open-source ecosystem it keeps feeding. Artificial Analysis puts its cost per Intelligence Index task around $0.55 — nearly identical to Gemini 3.8 Flash's $0.58. Speed is the one weak spot: 235 tokens/sec median output vs Gemini's 305.

The cadence matters more than the price. Muse Spark 1.1 shipped July 9, 1.2 on August 5, 1.3 today — a frontier model roughly every four weeks, with no official Meta blog post for 1.3 as of this writing (ai.meta.com/blog/muse-spark-1-3/ is a 404). The model reaches the world through API partners and benchmark trackers first. That's a structural change: Meta's frontier releases now move at the speed of a SaaS product, not a research announcement.

The Age Gate: Meta's Model Blocked My API Call

I tried to run Muse Spark 1.3 the way I run every new model — straight through the OpenRouter API:

$ curl -X POST https://openrouter.ai/api/v1/chat/completions \
    -H "Authorization: Bearer $OPENROUTER_API_KEY" \
    -d '{"model":"meta/muse-spark-1.3","messages":[{"role":"user","content":"Say hello."}}'

{
  "error": {
    "message": "This model requires you to complete the following before use: 18+ age confirmation.
                Confirm at https://openrouter.ai/settings/preferences.",
    "code": 403
  }
}

A 403 with a pointer to a settings checkbox. The same request to google/gemini-3.8-flash on the same key went through instantly — 927 reasoning tokens of Manacher's algorithm, $0.0057, no ceremony. Meta has wrapped its frontier model in an age attestation that an API client cannot satisfy programmatically. For a lab pitching itself at the AI-builders market, that's a strange gate: agents don't have birthdays. If your pipeline talks to meta/muse-spark-1.3, someone has to log in and click "I am 18+" first. It's a human-in-the-loop requirement on a model that's supposed to be the machine's.

The Bottom Line on Muse Spark 1.3

Meta now has a genuine frontier model that wins its weight class on price-to-intelligence: $1.25/$4.25 per million, 1M context, video input, and a 62.09 index that clears Google's entire current lineup. It is not the smartest model on Earth — Claude Fable 5.1 sits 3.5 points higher — and it's verbose (Artificial Analysis flags 120M output tokens in evaluation vs a 72M median), which will sting anyone billing per token. But the story isn't the score. It's that Meta ships frontier models monthly, prices them like commodity inference, and then makes agents beg an age checkbox for access. The bench is moving at four-week intervals now. Build accordingly.

Related from Dispatch