Descriptions:
Matt Wolfe’s weekly AI news digest for the week of April 10, 2026 leads with an in-depth breakdown of Anthropic’s Claude Mythos and the accompanying Project Glass Wing. Wolfe walks through Anthropic’s own published benchmarks: Mythos Preview scored 83.1% on cybersecurity vulnerability reproduction versus 66.6% for Claude Opus 4.6, beat Opus by 24 percentage points on SWE-bench Pro, and nearly doubled Opus on SWE-bench Multimodal. Mythos independently discovered a 27-year-old OpenBSD flaw, a 16-year-old FFmpeg bug, and chained multiple Linux kernel vulnerabilities — capabilities Anthropic cites as the reason for withholding general availability. Instead, access is being granted selectively to cybersecurity teams at named partner companies via Project Glass Wing.
The episode also highlights an open-weight model release that Wolfe argues has been underreported. The model achieves near-GPT-5.4 and Opus 4.6 coding performance on standard benchmarks — positioning it as the strongest freely downloadable, fine-tunable coding model to date — while scoring second on math benchmarks behind Gemini 3.1 and GPT-5.4.
Additional segments cover new Gemini features (interactive simulation generation, paralleling recent OpenAI and Anthropic releases), and Wolfe closes with a callout for collaborators to help design “average Joe” benchmarks for everyday model comparisons rather than focusing exclusively on advanced math and science tasks.
📺 Source: Matt Wolfe · Published April 10, 2026
🏷️ Format: Roundup
