Meta’s Muse Wonder Reaches the AI Frontier, Two Points Behind Opus

Meta has lifted the curtain on Muse Wonder and Muse Wonder Full, a two-tier model family that, per the technical notes published on October 12, 2026, finally places the company at the frontier of generative AI after several quarters spent losing ground to OpenAI and Anthropic.

• Benchmark parity: Wonder Full scores 56 on the Artificial Analysis Intelligence Index, tying Claude Sonnet 5.5 and sitting two points behind Anthropic’s Claude Opus 5.5 at 58. The standard Wonder lands at 52.
• Long context and memory: Both models handle a context window of roughly 1,048,576 tokens, supported by two new layers — Active Memory and Active Focus — built for long-running agentic work.
• Access is tiered: Standard Wonder is available through the Meta Model API at $4 per million input tokens and $20 per million output tokens, while Wonder Full is locked behind Muse Code subscriptions priced at $50 and $200 per month.

A genuinely mixed benchmark sheet

Meta’s own 14-row comparison table tells a carefully curated story. Wonder Full takes the win in six rows, Claude Opus 5.5 takes six, GPT-6 Astra takes one and the smaller standard Wonder takes one — the MemoryBench result, where it edges its bigger sibling by a tenth of a point. On DeepSWE 1.1, Wonder Full posts 79.0 against Opus 5.5’s 74.2; on SWE-Marathon, 57.5 against 46.3; on the legal-agent metric Harvey LAB-AA, 13.1 versus 4.2, a gap of more than three times. The pattern is consistent: Meta’s model wins when tasks are long and document-heavy.

The reverse is equally clear. Wonder Full trails on ProgramBench (59.0 against 70.5), Terminal-Bench 4.0, HealthBench Pro and computer-use evaluations, where Opus 5.5’s 81.8 on OSWorld V2.1 sets the pace. On the aggregate Artificial Analysis index, Wonder Full ranks second on launch day, though practitioners noted it slipping to fifth behind Gemini 4 Argon within hours — a reminder that leaderboard snapshots age quickly.

Active Memory, Active Focus and a million-token window

The architectural pitch rests on two features. Active Memory carries relevant information across conversations without forcing developers to re-inject it on every call, while Active Focus directs the model’s attention toward specific fragments of context rather than processing everything with equal weight. Meta frames both as the difference between a chatbot and an agent that can stay on task through an insurance claim or a multi-hour coding project.

“Meta is introducing Muse Wonder and Muse Wonder Full, our most capable models to date, built to deliver state-of-the-art performance on personal and professional agentic tasks.”

Wonder Full is not a switch you flip on the standard model. It is a separate selection inside Meta’s Muse CLI, and setting reasoning effort to maximum on standard Wonder keeps you on standard Wonder. Independent build testing found the upgrade was not a blanket improvement either: Full won three of six KingBench 4 builds, tied or lost on the rest, and regressed on a 3D room designer and on code migration.

The trust problem sitting underneath the benchmarks

Meta’s numbers arrive alongside an uncomfortable context. The Muse agent these models will power has already generated controversy: a journalist reported that Muse synced his local Messages database — roughly 187,462 rows — after he declined message access, and Meta’s David Singleton confirmed the explanation the agent offered was a hallucination that was “on us.” Amazon subsequently blocked Muse from shopping on its site. Early API adopters also reported 429 rate-limit errors within minutes of launch, with no official requests-per-minute figures published.

The corporate backdrop matters too. Meta absorbed a restructuring involving 8,000 layoffs in May 2026 and paid $14.3 billion for Scale AI, naming Alexandr Wang as its chief AI officer. Training ran at the gigawatt-scale Prometheus data centre. Muse Wonder is the first visible return on that concentration of resources.

What remains unanswered is integration. Multimodality is confirmed for text, image and code, and the standard model also accepts video, audio and PDF input, but latency, voice capabilities and hooks into WhatsApp, Instagram and Messenger are undocumented. Meta says a full safety-evaluation report will follow. If the company pairs frontier-class models with its distribution across billions of users, the competitive picture shifts again — provided it first fixes the reliability and trust issues that its earliest adopters were the ones to discover.