Mythos leads shared evals, GPT-5.5 closes gap

This is less a knockout than a reminder that frontier-model benchmarking is fragmenting into selective, vendor-picked slices while the most capable systems stay partially closed. Anthropic’s official Mythos materials position it as a step above Opus and unusually strong at cybersecurity, coding, and autonomous exploit work, which helps explain why it still posts intimidating scores on shared evals. OpenAI’s GPT-5.5 launch muddies the Reddit narrative: OpenAI says GPT-5.5 beats Mythos Preview on Terminal-Bench 2.0 at 82.7% versus 82.0%, so Mythos is not sweeping every overlapping benchmark. The more important distinction for developers is availability: GPT-5.5 is rolling out broadly across ChatGPT, Codex, and soon the API, while Mythos remains limited-access and effectively unavailable for normal product work. Closed or preview-only models can dominate charts without changing day-to-day developer behavior; public access, latency, pricing, and safeguards still decide who actually shapes the tooling ecosystem. Benchmark screenshots from Reddit are useful signal, but they flatten caveats around memorization, eval selection, and deployment constraints into a misleading one-number horse race.