Latest / UpNext AI / Meta’s Muse Wonder Models, Epic Games’ AI-First Shift, and the Limits of AI Benchmarks
Transcript
- Welcome to the UpNext AI podcast. It's Tuesday, October 13th, 2026, and here's what matters in AI today. Meta has released Muse Wonder and Muse Wonder Full, its first attempt at frontier-scale models. TechCrunch reports that the pair marks a step up from Muse Spark, which was designed for lightweight, low-latency use. Meta says the larger models perform better on complex tasks such as coding. In the company’s own tests, their benchmark scores were comparable to those of Fable 5.1 and GPT-6 Astra. Those are Meta’s comparisons, rather than an independent verdict on how the models will perform for every user. The access arrangements matter as much as the launch claim. Wonder is available through the Meta Model API, priced at $4 per million input tokens and $20 per million output tokens. Wonder Full is reserved for Muse Code subscribers and will not be served through that API. That gives developers a way to try one model in their own applications, while the larger variant has a different distribution path. TechCrunch says both models will also power Meta’s Muse agent. Meta says they were trained at its Prometheus data center in central Ohio. For teams weighing a new model provider, the immediate questions are whether its coding results hold up on their work, and whether the available model fits their access and cost requirements. That brings us to what Meta wants these models to do over longer stretches of time. WIRED highlights features intended to help keep agents on course. Meta also emphasizes what it calls active memory: retaining and combining information a user provides over time, so an assistant can draw on it during later tasks. In Meta’s testing, Muse Wonder scored higher on the MemoryBench memory evaluation than several competing models. That claim speaks to a real challenge for persistent assistants. A model handling a task today may need information the user supplied days or weeks earlier, and it needs to bring back the relevant detail. The company is positioning memory as part of the agent experience, alongside its claims about more demanding coding work. WIRED also reports that the release includes features intended to keep agents from going off course. There is no demonstrated effectiveness attached to that description here. Builders assessing an agent should test recall and behavior over the kinds of extended tasks they actually plan to run. A different kind of AI workflow change is being reported at Epic Games. Ars Technica, citing Bloomberg’s reporting, says Epic is shifting its Disney collaboration toward AI-first development for a Marvel- and Star Wars-themed shooter. According to that account, models would generate code and guide artwork, while human developers manage and direct the resulting assets. Internal messages described in the report call for no hand-written code and artwork made through full AI workflows. The project has reportedly been delayed indefinitely. The report also describes understaffing and slower-than-expected development, so the delay should not be attributed to the AI shift alone. For game studios, this is a substantial proposed change in how work moves through production: developers would spend more of their time directing and reviewing model output. It remains a reported internal approach, rather than a confirmed Epic announcement or evidence that the approach has delivered a finished game. And there is a useful research check on claims about model performance. A paper published earlier this week asks whether the benchmark scores used to discuss safety and deployment readiness reflect the product people actually use. The researchers compared responses through ChatGPT’s chat interface and OpenAI’s API, with web search turned on and off. They drew prompts from the BBQ and SafetyBench evaluations and repeated each prompt, checking consistency, citations and abstentions as well as accuracy. With search disabled, the chat interface was less accurate than the API on both benchmarks. Turning on web search reduced accuracy by as much as eight percentage points in the tests, while repeated runs produced inconsistent responses for up to twenty-one percent of prompts. Citation choices and abstentions also varied. This is a study of one widely used model under particular test conditions, not a universal ranking of interfaces or search tools. Its operational lesson is clear: evaluate the interface and settings people will use, and run prompts more than once before treating a score as a readiness signal. ...Are you building apps with voice? Elevate your app's voice capabilities with ElevenLabs. Their API is a game changer for embedding dynamic, responsive voice interactions in your applications, providing unprecedented realism, flexibility and latency. Visit up next dot fm slash eleven to check out their latest offerings. ... Two brief policy notes before we go. Bloomberg reports that former U.S. Commerce Secretary Gina Raimondo warned against countries going it alone on AI. She called for governments and companies to cooperate on the technology’s economic and security risks. And the Financial Times reports that UK minister Kanishka Narayan urged companies not to release frontier models if they cannot adequately control the risks. It is a stated position on release decisions, rather than a new rule announced in the report. Before we wrap up, a quick note: this podcast is generated with the assistance of AI and is intended for informational purposes only. All referenced articles, research, and commentary remain the property of their original authors and publishers. And that's your briefing for today. If you enjoyed this episode, share it with a friend or colleague—and follow UpNext AI on Apple Podcasts, Spotify, or wherever you get your podcasts! Full source links are in the episode notes, and we'll be back tomorrow with what's up next!