Meta's Muse Wonder (Codename "Watermelon"): Third-Party Evals Confirm a Real Frontier Jump — #3 on Artificial Analysis, Still Behind Anthropic's Top Two

Meta's Muse Wonder (Codename "Watermelon"): Third-Party Evals Confirm a Real Frontier Jump — #3 on Artificial Analysis, Still Behind Anthropic's Top Two

Update (Oct 13–14): the launch claims now have independent numbers, and the codename is confirmed. Meta shipped Muse Wonder and Muse Wonder Full on Monday, Oct 12 — internally codenamed Watermelon, per Meta AI chief Alexandr Wang's launch-day interview — and Artificial Analysis has published its independent read: "a 'material jump' over previous Meta models. In the Intelligence Index, Muse Wonder Full even overtakes Anthropic's Claude Fable 5.1, and the standard Muse Wonder trails it by just one point. Yet Muse Wonder costs about 40 percent less than Fable per test task¹."

The independent leaderboard (AA Intelligence Index, highest reasoning setting)

1. Claude Opus 5.5 (Anthropic): 57.6
2. Claude Sonnet 5.5 (Anthropic): 56.0
3. Muse Wonder Full (Meta): 55.8
4. Claude Fable 5.1 (Anthropic): 53.4
5. GPT-6 Astra (OpenAI): 52.7
6. Gemini 4 Argon (Google): 52.6
7. Muse Wonder (Meta): 52.4
8. GPT-6.1 Sol (OpenAI): 51.8
9. Muse Spark 1.3 (Meta): 48.1
10. Grok 4.7 (xAI): 46.4

Six weeks took Muse Spark 1.3 from 48.1 to 55.8. The honest headline: Meta leapfrogged OpenAI's flagship and Fable 5.1, but Anthropic still holds the top two spots — "caught up to the frontier" is accurate; "took the lead" is not.

The fine print (capability vs hype)

• Cost: $4.52 per Intelligence Index task — ~40% cheaper than Fable 5.1 ($7.63) but 2.8x Muse Spark 1.3. Token prices jumped from $1.25/$4.25 to $4/$20 per million.
• Input appetite: Muse Wonder processes 7.69M input tokens per task, "two to seven times as many as the models from Google and OpenAI" — about three-quarters of its cost is input.
• Knowledge gaps: AA-Omniscience 30 (Wonder Full) vs Opus 5.5's 46, GPT-6 Astra's 43, Argon's 42.
• Hallucinations improved: 33% → 23% vs Spark 1.3; only Argon (15%) does better among top models.
• Access: Wonder Full is Muse Code-only ($50/$200/month tiers); the API carries standard Muse Wonder; the Muse consumer app ($20/$100 subscriptions) gets it "soon." AA flags narrow margins: the CritPt and GDP.pdf leads are "narrow," and on GDPval-AA Wonder Full sits within the statistical margin of error of Opus/Sonnet 5.5.

The product story

Meta's own framing: "Fourteen months ago, we began an overhaul of our AI efforts, and in that time we've gone from rebuilding our foundation to pushing the frontier." The two differentiators are Active Memory (carrying context across conversations) and Active Focus — Wang: "When you do real-world tasks, you're going to get roadblocks and you're going to get things that come up or various setbacks. We specifically tried to train the model to be able to try other approaches instead of stalling and then checking back in with the user at any time." On safety, Meta says the model "offers our strongest safeguards to date" against "cyber misuse" — protections that "only come with the paid model for now." Context worth carrying: the Muse app has an open dispute over an agent allegedly reading private messages without permission (Meta disputes the claim), and this launch is fundamentally a monetization test — revenue outside ads, priced against Anthropic.