441e85161ClaimNews eventMark Zuckerberg revealed a new AI model named "Muse Wonder" that Meta says is as powerful as the latest technology from OpenAI and AnthropicAgreeDisagree
441e85161ClaimMeta releases Muse Wonder, its first frontier AI model, to rival OpenAI and AnthropicAgreeDisagree
441e85161ClaimAt its October 2026 launch, Meta's Muse Wonder Full model was available only in Meta's Muse Code coding tool, not through the Meta Model APIAgreeDisagree
441e85161ClaimMeta priced its Muse Wonder model at $4 per million input tokens and $20 per million output tokens on the Meta Model APIAgreeDisagree
441e85161ClaimMeta reported that its Muse Wonder model outscores Anthropic's Claude Fable 5.1 and Claude Opus 5.5 and OpenAI's GPT-6 Astra on Meta's internal MemoryBench long-conversation memory benchmarkAgreeDisagree
441e85161ClaimMeta's Muse Wonder models accept text, image, video and audio input and produce text outputAgreeDisagree
441e85161ClaimMeta trained its Muse Wonder AI models at its gigawatt-scale Prometheus data center in central OhioAgreeDisagree
441e85161ClaimMeta released its Muse Wonder and Muse Wonder Full AI models on October 12, 2026AgreeDisagree
441e85161ClaimMeta's Muse Wonder Full hallucinated on 23% of questions on Artificial Analysis's AA-Omniscience knowledge benchmark, down from 33% for Meta's Muse Spark 1.3AgreeDisagree
441e85161ClaimMeta said in October 2026 that its Muse Wonder model would power Meta's Muse personal agent app, starting with paying subscribersAgreeDisagree
441e85161ClaimMeta reported state-of-the-art results for its Muse Wonder models on personal agent, professional agent, memory and coding benchmarks at their October 2026 launchAgreeDisagree
441e85161ClaimMeta's Muse Wonder Full model captured 92 flags on the ExploitGym hacking benchmark in Meta's testing, but only 21 by exploiting the intended vulnerabilityAgreeDisagree
441e85161ClaimMeta's Muse Wonder Full model refused 98.7% of high-risk English biology prompts on SecureBio's BioTIER safety benchmark in Meta's testingAgreeDisagree
441e85161ClaimMeta's Muse Wonder Full scored 57% on the Terminal-Bench 4.0 coding benchmark in Artificial Analysis's testing, up from 33% for Meta's Muse Spark 1.3AgreeDisagree
441e85161ClaimRunning in Meta's Muse Code, Meta's Muse Wonder Full scored 68 on the Artificial Analysis Coding Agent Index, level with Anthropic's Claude Sonnet 5.5 in Claude CodeAgreeDisagree
441e85161ClaimMeta's Muse Wonder is the company's first frontier AI model since Alexandr Wang restructured Meta's AI efforts in August 2025AgreeDisagree
441e85161ClaimMeta rated its Muse Wonder Full model at moderate or lower risk in cybersecurity, chemical and biological threats, and loss of control once safety mitigations were appliedAgreeDisagree
441e85161ClaimMeta lists visual and spatial reasoning, especially 3D tasks, as a known limitation of its Muse Wonder modelsAgreeDisagree
441e85161ClaimMeta could not rule out its Muse Wonder Full model reaching the high risk threshold for chemical and biological threats under Meta's Superintelligence Scaling Framework without safety mitigationsAgreeDisagree