Descriptions:
Matthew Berman livestreams his real-time reaction to two concurrent AI industry developments: the release of Claude Opus 5 from Anthropic, and an open letter signed by Nvidia CEO Jensen Huang and other major AI figures calling for the United States to embrace open-weight AI models. The open source discussion frames the model release in broader strategic context, with Berman arguing that open-source AI drives down inference costs, prevents dangerous power concentration among a handful of closed-source labs, and ultimately benefits the entire ecosystem — including Nvidia, which sells chips regardless of who builds the models.
On the Claude Opus 5 benchmarks, Berman highlights a particularly striking result on ARC-AGI 3 — an evaluation where models must play completely unknown games with no rules or context provided — where Opus 5 jumped to 30% from a previous frontier of roughly 8%. He also covers computer use improvements, automation workflow benchmarks, and the cost-per-task charts showing Opus 5 outperforming both Fable 5 and GPT-5.6-Soul at lower price points across multiple evaluations.
Berman notes that Opus 5 deliberately trails the unguardrailed Mythos 5 model on cybersecurity exploit tasks, which he interprets as an intentional Anthropic safety decision rather than a general capability gap — and finds it notable that this targeted capability removal apparently did not hurt the model’s overall performance. The stream includes live audience Q&A and covers cost-per-task as the defining competitive metric of the current model generation.
📺 Source: Matthew Berman · Published July 24, 2026
🏷️ Format: Livestream
