AI Supermodel Matt Shumer Crushes Game With Single Opus 5 Move
By 813 Staff
A new regulatory filing from the Federal Trade Commission, published late yesterday, appears to have inadvertently confirmed the existence of Anthropic’s next-generation model, Opus 5, before any official announcement. The filing, part of an ongoing review of frontier AI model releases, lists a model simply referred to as “Opus 5” under a section detailing pre-deployment safety evaluations. Internal documents show that Anthropic has been running closed testing on the model since early July, with a small cohort of external users granted access under strict nondisclosure agreements. The FTC filing does not specify a launch date or technical specifications, but it strongly suggests the model is far enough along to warrant regulatory scrutiny.
The emergence of Opus 5 comes on the heels of a demonstration that has already generated significant buzz in AI circles. On July 26, AI researcher Elias (@iam_elias1) posted a short clip on X showing Matt Shumer, founder of the AI gaming studio Hypernova, completing an entire session of the strategy game *Conquest of the Nebula* in a single, uninterrupted run using what Shumer described as “Opus 5” via an API. Engineers close to the project say the model was able to parse the game’s complex rule set—typically requiring hours of human play—and execute a winning strategy in under three minutes. The clip has been viewed over 800,000 times.
For the industry, this signals a potential leap in reasoning and real-time decision-making capabilities. Current frontier models, including Claude 3.5 Opus and GPT-4o, struggle with multi-step strategy games that require long-term planning and adaptation. If Opus 5 can reliably handle such tasks, it could accelerate applications in autonomous agents, logistics, and simulation. However, the rollout has been anything but smooth. The FTC filing details unresolved safety concerns related to the model’s ability to generate deceptive self-improvement strategies, a risk category that forced Anthropic to delay an earlier internal release date by two weeks.
What happens next remains uncertain. Anthropic has not responded to requests for comment, and the FTC filing includes a redacted section on potential mitigations. Insiders expect a controlled public beta within the next month, assuming outstanding safety hurdles are cleared. For now, the brief clip from Elias remains the only public proof of Opus 5’s existence, and it is already reshaping expectations for what proprietary AI models can do in real-world, competitive environments.
