OpenRating is an independent, open-source AI rating platform. Join the community
OpenRating Research
Technical deep dives into benchmark integrity, LLM serving infrastructure forensics, and empirical pricing analysis.
OpenRating is an independent model evaluation platform. Every article here starts from primary measurements — benchmark runs, inference traces, and pricing data we collect ourselves — so the numbers behind the leaderboard are explainable.
A comprehensive technical teardown of AI silicon in 2026: NVIDIA Blackwell & Rubin, Google TPU v8, AMD Instinct MI355X, Cerebras WSE-3, Groq LPUs, and OpenAI's custom Jalapeño ASIC.
Why behavioral evaluations fail against proxy wrappers, and how tokenizer merges, template offsets, and error boundaries expose the true underlying model.
What we cover
OpenRating's leaderboard is built on measurements we run and verify ourselves. The blog publishes the methods, teardowns, and raw analysis behind those numbers — including the cases where popular benchmarks fail.
How leaderboards can be gamed — and the forensic methods we use to keep OpenRating measurements honest.
Silicon architectures, inference economics, and what watts-per-token means for the cost of every model you run.
Empirical breakdowns of token pricing, cache economics, and the true per-task cost of frontier models.
Fingerprinting, proxy detection, and verification techniques for knowing which model is really answering.
Monthly empirical updates when labs drop new models. No marketing spam.
Join 4,200+ ML engineers and researchers. Unsubscribe at any time.