Compare Models

Wonder

Wonder is a model from Meta, released October 12, 2026. It ranks #10 of 47 models on the Vals Index with 61.06%. Its best result is #1 of 47 on CyberBench v1.1.

Release Date: Oct 12, 2026

Developer Meta  🇺🇸
Context Window 1M
Max Output Tokens 131k
Token Costs (in/out) $4.00/20.00
Weights Private
Input Modalities

Accuracy

61.06 % ± 1.07

Cost / Test (Vals Index)

$ 11.43

Latency

37 min 8 s

Fallback Rate 0.00%
Refusal Rate 0.22%
Vals Index
BenchmarksAccuracyRankings
Vals Index

61.06%

±1.07
10/47
Code Migration

62.55%

±4.35
6/77
CyberBench v1.1

80.36%

±5.13
1/47
EMB

67.08%

±2.89
18/74
Finance Agent (v2)

56.64%

±0.38
16/78
Legal Research Bench

50.48%

±3.48
6/78
MysteryMechanism

27.48%

±3.00
15/27
ProofBench v1.1

84.00%

±3.69
12/50
Public Benefits Bench v1.1

70.23%

±1.19
6/50
Tax Agent Bench

36.27%

±3.47
10/69
Vibe Code Bench v1.1

83.09%

±2.49
18/112
BioMysteryBench

67.04%

±0.37
19/27
Harvey's Legal Agent Benchmark

17.08%

±3.10
8/79
IOI

66.94%

±4.91
15/44
Terminal-Bench 4.0

49.49%

±2.20
9/47
Terminal-Bench Science

22.86%

±5.05
9/40
Proprietary BenchmarksAcademic BenchmarksIndustry Partners
Vals
Default Provider : Meta
Temperature: 1
Top P: Default
Top K: Default
Max Output Tokens: 131,072
Reasoning Effort: max