The company launched its Muse Wonder AI model, which it said outperformed some rivals on technical benchmarks.

Meta introduced Muse Wonder and its larger sibling, Muse Wonder Full, on October 12, its first frontier-model release since its AI reorganization in August 2025. Meta says Wonder Full led rival models on two long-horizon coding tests: it scored 79.0 on DeepSWE 1.1, versus 74.2 for Claude Opus 5.5 and 74.1 for GPT-6 Astra, and 57.5 on SWE-Marathon, versus 46.3 and 42.5. Results are not uniformly dominant: on Terminal-Bench 4.0, Wonder Full scored 57.1, below Opus 5.5’s 59.6 and Astra’s 59.1. The models accept text, images, video and audio, support a one-million-token context window, and return text; Meta highlights Active Memory and Active Focus for recalling context and persisting through difficult tasks. Wonder is available via Meta’s API at $4 per million input tokens and $20 per million output tokens; Wonder Full is limited to Muse Code subscribers. The launch builds on Muse Spark, which Meta positioned as a faster, lighter model, and advances its effort to compete with OpenAI and Anthropic.