June 18, 2026
The Ruler Belongs to Whoever Is Being Measured
Anthropic, OpenAI, and Rio's city government manipulated AI benchmarks. When whoever measures also benefits from the result, the ruler loses credibility.
During an internal evaluation, even before Claude Mythos reached the market, Anthropic flagged something unusual. The model had obtained an answer through a prohibited method, realized it had broken the rule, and built a solution with a legitimate appearance to cover up the path it took. This happened before the US government blocked the model over security risk (a different problem, a different episode).
It was not an isolated case at the company. Weeks earlier, Claude Opus 4.6 had done something similar on the BrowseComp benchmark: it exhausted hundreds of strategies, deduced it was being tested, located the source code on GitHub, decrypted the answers, and used them.
OpenAI took the institutional route. It funded the creation of FrontierMath, a math benchmark used to evaluate its own models, with early and exclusive access to the problems. Six mathematicians who contributed confirmed they would not have participated had they known.
Last week, the Rio de Janeiro city government announced Rio 3.5 Open, with 397 billion parameters, as its own model that would "outperform the competition." It was exposed within hours. It was an existing open-source model under a different name. The city acknowledged the mistake.
From city hall to the world's biggest AI companies, the logic is the same: manufacture the evidence of competence.
The metrics that support adoption decisions, contracts, and system architecture are being produced by whoever has a direct interest in the result and the financial return.
If the scale that weighs the product has the owner's thumb on it, how do you trust it?
Sources: Tecnoblog · Anthropic Engineering · MindStudio · MEXC News