Every frontier lab now leads on a test of its own choosing, so the decisions that matter have moved to price per token and to agent liability, where the evidence is harder
Frontier leadership is now contested benchmark by benchmark, with vendors grading their own work, while cheaper open-weight and low-tier models erode price premiums and regulators move from guidance to enforcement over agent behavior. Three strands support that claim this week. On rankings, Claude Opus 5.5 leads one independent index while Google claims a separate coding benchmark for Gemini 4 Argon using a score it calculated itself, so no single model holds the top position across measures. On price, Zhipu and OpenAI both shipped cheaper options that undercut their own prior or flagship offerings. On governance, the agent breach at Hugging Face has moved from investigation into a lawsuit and an industry-wide federal inquiry, with Australia weighing direct reporting duties. Together these shift attention away from which model is best and toward what a task costs and what an agent is permitted to do.
60.3 Claude Opus 5.5 score leading the LLM Stats overall index
77.9% Gemini 4 Argon DeepSWE v1.1 score, calculated by Google itself
one-tenth the price GLM-5.3-Flash cost relative to GLM-5.2, per Zhipu
$0.10 input and $0.50 output per 1M tokens Pricing for GPT-6 Luna, OpenAI's lowest-cost GPT-6 tier
Signal Pro adds the evidence behind the argument, how this week sat against previous weeks, the policy read for every market and what it changed for planning.
Get Signal Pro →The week's read in your inbox every Monday. Unsubscribe anytime.
A research seat runs Mireto's research on any organization in your plan's market and helps you prepare the conversation. Partner firms deliver the result to their clients in their own name.