very AI company wants its model to be the best. But who comes out on top often depends on who is holding the yardstick.

Practically every AI model released over the last year has come with the label “state of the art,” inching out the competition on standard benchmarks and evaluations for metrics such as performance, alignment, and context window length. But now some firms are developing their own assessments.

Two major model firms have released new benchmarking and evaluation tools in the last week:

• On Friday, Anthropic introduced Bloom, an open-source framework for generating behavioral evaluations of frontier AI models. Bloom allows researchers to quickly develop tests for specific model traits they’re interested in tracking. 
• And on Wednesday, OpenAI released FrontierScience, a benchmark that evaluates AI capabilities for “expert-level scientific reasoning” in domains like physics, chemistry and biology.

Of course, in testing these measurements, Anthropic found that its Claude Opus 4.5 model outperformed competitors like OpenAI, xAI, and Google at reining in troublesome behaviors, including delusional sycophancy, self-preferential bias, and self-preservation. And OpenAI’s benchmark revealed that GPT-5.2 beats other frontier models in research and Olympiad-style scientific reasoning.

While these benchmarks might not be lying about these models’ capabilities, they likely tell you about these systems’ specific features, but “don’t necessarily really create a fair way of comparing different tooling,” said Bob Rogers, chief product and technology officer of Oii.ai and co-founder of BeeKeeper AI, told The Deep View. These tests emphasize the things that the model developer is proudest of, rather than serving as an objective barometer.

“This is a big part of the old school big tech playbook,” said Rogers. “What you do is you build a benchmark that really emphasizes the great aspects of your product. Then you publish that benchmark, and you keep moving your roadmap forward and keep being ahead of everybody else on that benchmark. It’s a natural thing.”