Today for AI
HOT RADAR
llmHEAT 162.4°

AI CLUSTERED EVENT · 10/13/2026

Meta Launches Muse Wonder: Trained on Prometheus Data Center, Surpasses Claude Opus 5.5 in SWE-Marathon Benchmark

9 reports archived8 independent sourcesupdated 10/13/2026, 02:48:29
Synthesis & Latest Updates
8 Sources Cross-Validated

Meta's Superintelligence Lab has officially launched Muse Wonder, its first frontier-class closed-source model family, marking a significant return to the top tier of AI development. Trained on the gigawatt-scale Prometheus data center in Ohio, the series comprises Standard and Full versions, with core breakthroughs in long-horizon agentic tasks. In key benchmarks, Muse Wonder Full scored 57.5 in SWE-Marathon, substantially outperforming Claude Opus 5.5's 46.3, and achieved 79.0 in DeepSWE coding tests, demonstrating superior capabilities in complex engineering scenarios.

Cross-verified reports indicate support for 1 million token context windows and multimodal inputs including text, image, video, and audio. The Full version reached an intelligence index score of 56 in Artificial Analysis v4.3, an 8-point jump from its predecessor and only 2 points behind top-tier models. Notably, it tripled competitors' scores in the legal automation benchmark Harvey, proving robustness in hallucination-free task execution. Access is stratified: developers can use the Standard version via API or Muse Code, while the powerful Full version is restricted to specific subscribers, leaving general user access timelines unclear.

This release highlights Meta's advancements in compute infrastructure and architecture, aiming to build developer ecosystem barriers through commercial strategies like Muse Code subscriptions. Despite community calls for open-sourcing, Muse Wonder's hard metric breakthroughs in long-cycle coding and agent tasks pose a substantive challenge to the market dominated by Anthropic and OpenAI, accelerating competition in the personal superintelligence race.

LATEST/Meta confirmed that Muse Wonder is now available on OpenRouter and has entered the Agent Arena evaluation system for real-world task tracking. While no timeline for consumer access has been announced, community pressure mounts for open-source alternatives or lightweight versions to address concerns over restricted access tiers.

HEAT TRENDHourly heat curve

3 fully observed hours
Now
162.4
Peak
166.603:00
24h change
–

The line compares only sources observed throughout; gap hours are interpolated to keep the trend continuous.

TIMELINECoverage timeline

Total 8 reports · Latest first
  1. The DecoderT2·86 pts
    • Muse Wonder Full scored 56 on the Artificial Analysis Intelligence Index v4.3, with the standard version at 52, significantly surpassing the predecessor Muse Spark 1.3 (48).
    • Performance on Terminal-Bench 4.0, which measures autonomous complex task handling, jumped dramatically from 33% to 57%, indicating top-tier agentic capabilities.
    • The new models introduce an 'Active Memory' feature for retaining key information and reduce output token usage by approximately 10% compared to predecessors, enhancing inference efficiency.
  2. Rohan Paul (@rohanpaul_ai)T2·78 pts
    • Muse Wonder scores 57.5 on SWE-Marathon, creating a significant gap compared to Claude Opus 5.5's 46.3 and GPT-6 Astra's performance.
    • SWE-Marathon requires AI to independently build full software projects like a C compiler or Slack clone within 2-10 hours, counting success only if all hidden checks pass.
    • Meta positions Muse Wonder as its largest and most capable model, emphasizing competitiveness in both personal and professional agent scenarios.
  3. Arena.ai (@arena)T1·68 pts
    • The Muse Wonder series represents Meta's most capable models yet, specifically engineered for long-horizon agentic tasks.
    • Models are accessible via Meta Model API and Muse Code, with Muse Wonder Full restricted to subscribers.
    • The series has joined Agent Arena, where performance is measured against millions of real-world tasks using causal tracing, alongside support for various code generation modalities.
  4. TechCrunch AIT2·86 pts
    • Muse Wonder represents Meta's first dedicated attempt at a closed-source frontier model, benchmarking against GPT-6 and Claude.
    • Tiered access strategy: Wonder is available via API, while Wonder Full is exclusive to Muse Code subscribers, signaling a commercial moat.
    • Infrastructure milestone: Training leverages the newly operational Prometheus data center, highlighting Meta's compute capacity expansion.
  5. 宝玉 (@dotey)T2·86 pts
    • Muse Wonder comes in Standard (speed-focused) and Full versions, trained on Meta's gigawatt-scale Prometheus data center.
    • Key strengths lie in agentic and long-horizon tasks, outperforming rivals in Personal Superintelligence Index (59.1 vs 55.2) and DeepSWE (79.0 vs 74.2).
    • Access is tiered: developers can use APIs/Muse Code today, Full version is subscription-only, while general user rollout timelines remain ambiguous and inconsistent across official channels.
  6. Artificial Analysis (@ArtificialAnlys)T1·78 pts
    • Muse Wonder Full achieves an Intelligence Index score of 56, an 8-point improvement over its predecessor, outperforming Muse Spark 1.3 across all ten evaluations.
    • Significant gains in agentic performance: Terminal-Bench 4.0 scores jumped from 33% to 57%, and GDPval-AA v2.1 Elo increased by 176 points.
    • Supports text, image, video, and audio inputs; Muse Wonder is available via API, while Muse Wonder Full is currently exclusive to Muse Code users.
  7. Rohan Paul (@rohanpaul_ai)T2·86 pts
    • Muse Wonder flagship beats Claude Opus 5.5 on 8 of 14 benchmarks and leads GPT-6 Astra in long-running tasks.
    • Scores 57.5 on SWE-Marathon (vs. 46.3 for competitor) and 79.0 on DeepSWE 1.1, demonstrating superior long-horizon engineering capabilities.
    • Achieves 13.1 on Harvey’s legal agent benchmark, more than triple Claude Opus 5.5’s score, highlighting robustness in hallucination-free task completion.
  8. Yuchen Jin (@Yuchenj_UW)T1·68 pts
    • Muse Wonder surpasses competitors Opus 5.5 and Astra on SWE-Marathon and DeepSWE v1.1 coding benchmarks.
    • Artificial Analysis Intelligence Index shows a minimal gap (only 2 points) behind top-tier model Opus 5.5.
    • Strong community demand for Meta to open-source Muse Wonder or provide lightweight alternatives like Spark 1.3.