AI
“Meta released its new AI model family called Muse Wonder (codenamed Watermelon), with the top version Wonder Full scoring 56 on Artificial Analysis' Intelligence Index, behind only Claude Opus 5.5 at 58.”
Plain restatementMeta released a model family named Muse Wonder, internally codenamed Watermelon. Its highest-capability variant, Muse Wonder Full, scored 56 on the Artificial Analysis Intelligence Index, and the only model scoring higher was Claude Opus 5.5 at 58.
Meta really did release a model family called Muse Wonder on 12 October 2026, its internal codename really was Watermelon, and the top version, Wonder Full, really did score 56 on the Artificial Analysis Intelligence Index. That score came from Artificial Analysis itself, an independent tester, not from Meta, which makes it stronger evidence than a company self-report. The part that does not hold up is the ranking. The post says Wonder Full was behind only Claude Opus 5.5 at 58, but Artificial Analysis' own leaderboard also lists Claude Sonnet 5.5 at 56, and on the unrounded numbers Sonnet sits at 56.0 against Wonder Full's 55.8. Artificial Analysis describes Opus 5.5 and Sonnet 5.5 as the highest-scoring models, with Wonder Full in the group that follows, so the honest description is a near-tie for third rather than an exclusive second place. Two further details are missing: the 56 was achieved at maximum reasoning effort, which is an expensive setting, and the other four benchmark wins listed in the post come from Meta's own scorecard and leave out tests where rival models score higher. Rankings on this index change often, so this assessment is as of 19 October 2026.
Meta released [drifted from the evidence:] its new AI model family [drifted from the evidence:] called Muse Wonder (codenamed Watermelon), [drifted from the evidence:] with the top version Wonder Full [drifted from the evidence:] scoring 56 on Artificial Analysis' Intelligence Index, [drifted from the evidence:] behind only Claude Opus 5.5 at 58.
Meta released [added by the neutral restatement:] a model family [added by the neutral restatement:] named Muse Wonder, [added by the neutral restatement:] internally codenamed Watermelon. [added by the neutral restatement:] Its highest-capability variant, Muse Wonder Full, [added by the neutral restatement:] scored 56 on [added by the neutral restatement:] the Artificial Analysis Intelligence Index, [added by the neutral restatement:] and the only [added by the neutral restatement:] model scoring higher was Claude Opus 5.5 at 58.
Red-tinted words in the claim drifted from the evidence. Green-tinted words are what a neutral restatement needs.
The trace / claim to source
- Meta Superintelligence Labs released a model family called Muse Wonder, in two variants, Muse Wonder and Muse Wonder Full, on 12 October 2026.
- The internal codename was Watermelon, stated on the record by Meta's AI chief in an interview with CNBC.
- Muse Wonder Full scored 56 on the Artificial Analysis Intelligence Index at max reasoning effort. This figure comes from Artificial Analysis, an independent evaluator that ran the tests itself, not from Meta.
- Claude Opus 5.5 at max effort scores 58 on the same index and is the single highest-scoring entry.
- Muse Wonder Full does score higher than GPT-6 Astra, Gemini 4 Argon and Claude Fable 5.1, which Artificial Analysis places at 53 each.
- The caption's availability description is correct: the standard Muse Wonder is on the Meta Model API and Muse Code, and Wonder Full is restricted to paying Muse Code subscribers. Meta's own developer blog states exactly this split.
- The claim says Wonder Full is "behind only Claude Opus 5.5." Artificial Analysis' own board places Claude Sonnet 5.5 (max) at 56 above it and Claude Opus 5.5 (xhigh) at 56 alongside it, and the operator's own prose names Opus 5.5 max and Sonnet 5.5 max as the highest intelligence models with Wonder Full in the group that follows. Dropping Sonnet 5.5 from the comparison converts a tie for third place among distinct models into an exclusive second place.
- The 56 is a score at maximum reasoning effort on index version 4.3, and it is a rounded display of 55.8. Opus 5.5's 58 is likewise 57.6 at max effort. On the unrounded numbers Wonder Full sits 0.2 points below Claude Sonnet 5.5 rather than above it, which is the specific fact that makes the "behind only" phrasing fail. Max-effort scores are also not what a typical user gets: the same Opus 5.5 model scores 42 at low effort.
- Benchmark cherry picking (secondary, applies to the caption's four bullets): the post lists four evaluations where Wonder Full leads and omits rows in the same comparison where it trails, including ProgramBench, where press transcriptions of Meta's card show 59.0 for Wonder Full against 70.5 for Opus 5.5, and Terminal-Bench 4.0, where Opus 5.5 is ahead.
- Marketing as evidence (secondary, applies to the caption's four bullets): the Harvey LAB, SWE-Marathon, DeepSWE and MemoryBench figures come from Meta's own model card at settings Meta selected, and MemoryBench is described in press transcriptions as an internal Meta benchmark. The post labels these "Source: Meta," which is honest, but they sit in the same list as the independently run index score without distinguishing the two kinds of evidence. The SWE-Marathon 57.5 figure is also reported as an xhigh-setting result in at least one transcription of the card, a setting qualifier the post does not carry.
- Whether the Harvey LAB-AA and Terminal-Bench 4.0-AA rows quoted in press comparison tables were run by Artificial Analysis or by Meta against those evaluations. Different outlets present the same table under both framings, and the attribution could not be resolved.
- The cause of the gap between Meta's model card figure of 59.0 and Artificial Analysis' 56 for the same model on the same named index. Candidate explanations reported include index version, run date and harness differences, but no party has published a reconciliation.
- Whether the current board ordering has changed in the last few days. Dated readings from 13, 14 and 19 October agree, and no newer model release appears to have entered above 56, but the index is revised often enough that this is an as-of-2026-10-19 statement only.
- Whether Wonder Full's benchmark standing translates to deployed performance. Reporting notes that independent real-world performance remains to be established, and the variant that produced the 56 is not the variant available through the API.
The underlying release is real and the headline number is correct. Artificial Analysis reports that Meta released Muse Wonder and Muse Wonder Full, achieving Artificial Analysis Intelligence Index v4.3 scores of 52 and 56 at max reasoning effort, each a material gain over Muse Spark 1.3 at 48, with a release gap of only six weeks. Artificial Analysis separately records Claude Opus 5.5 at max effort scoring 58 on the Intelligence Index, which it described at the time as the highest score it had measured by several points. The ranking element does not hold. Artificial Analysis' own comparison page currently states the ordering as Claude Opus 5.5 (Max, Default Fallback) at 58, Claude Sonnet 5.5 (Max, Default Fallback) at 56, Claude Opus 5.5 (Xhigh, Default Fallback) at 56, and Muse Wonder Full (Max) at 56. Its leaderboard page phrases the same result in prose: Claude Opus 5.5 (max with fallback) and Claude Sonnet 5.5 (max with fallback) are the highest intelligence models, followed by Claude Opus 5.5 (xhigh with fallback) and Muse Wonder Full (max). On the unrounded figures the gap runs against the claim. A dated reading of the board on 14 October lists Claude Opus 5.5 at 57.6 (displayed as 58), Claude Sonnet 5.5 at 56.0, and Muse Wonder Full at 55.8. A daily-updated mirror of the same board shows the same order: Claude Opus 5.5 (max) 57.6, Claude Opus 5.5 (xhigh) 56.0, Claude Sonnet 5.5 (max) 56.0, Muse Wonder Full (max) 55.8. Contemporaneous reporting described the position as a tie rather than second place: the results put Wonder Full in a tie with Claude Sonnet 5.5 at 56, ahead of GPT-6 Astra, Gemini 4 Argon and Claude Fable 5.1 at 53 each, but behind Claude Opus 5.5 at 58. The codename checks out on the record. Meta AI chief Alexandr Wang said in an interview that the company developed Muse Wonder, codenamed Watermelon, to improve the capabilities of personal and professional agents. On availability, Meta's own developer channel states the split: Muse Wonder launched on Muse Code and the Meta Model API at $4 input and $20 output per million tokens, while Muse Wonder Full is available exclusively by subscription in Muse Code.
Complete reasoning
The reply is formatted for pasting into the thread where the claim is circulating.
Ask this case
Answers come only from the case file above; nothing is added.
Did Meta really release a model called Muse Wonder with the codename Watermelon?
Yes. Meta Superintelligence Labs released Muse Wonder and Muse Wonder Full on 12 October 2026, and Meta's AI chief confirmed on the record that the internal codename was Watermelon.
Is it true that Muse Wonder Full scored 56 on the Artificial Analysis Intelligence Index?
Yes. That score came from Artificial Analysis, an independent evaluator that ran the tests itself, not from Meta, and it was achieved at maximum reasoning effort.
Is Muse Wonder Full really behind only Claude Opus 5.5, as the claim says?
No. Artificial Analysis' own board also lists Claude Sonnet 5.5 at 56, and on unrounded figures Sonnet actually scores slightly higher than Wonder Full, 56.0 versus 55.8. Artificial Analysis itself describes Wonder Full as part of the group following Opus 5.5 and Sonnet 5.5, which is closer to a tie for third than an exclusive second place.
Does the 56 score reflect what typical users would experience with the model?
Not necessarily. The 56 was achieved at maximum reasoning effort, a costly setting, and the variant tested is not the same one generally available through the API, so the investigation did not establish how this score translates to typical deployed performance.
Are the other benchmark wins mentioned in the post reliable?
They come from Meta's own model card rather than independent testing, and they leave out tests in the same comparison where Wonder Full trails, such as ProgramBench and Terminal-Bench 4.0, where rival models score higher.