Look beyond the model.

3 stories · 2 minute read

01 / CODING AGENTS / RESEARCH

The same model can come with a different bill

Compare the same model and tasks across three coding frameworks, measuring success and token cost. — Conceptual illustration based on HarnessTax. Compare both success and token cost; benchmark results do not guarantee savings on your workload.

HarnessTax compares seven models in three coding frameworks on 30 tasks per benchmark. Costs vary substantially despite similar success rates. The authors caution that these public benchmarks may overlap training data; results may differ in real development workflows.

Read the source ↗ harnesstax.github.io

02 / OPEN MODELS / ANALYSIS

Open-model competition extends beyond benchmark scores

Nathan Lambert argues that Chinese open-weight models lead on capability and adoption, while U.S. nonprofits lead in fully reproducible releases. His assessment draws on benchmarks, downloads, and platform usage; private deployments remain difficult to measure.

Read the source ↗ www.interconnects.ai

03 / AI POLICY / OPINION

AI leadership depends on who can put it to work

Jeffrey Ding challenges a view of AI leadership centered on frontier breakthroughs. He argues that broad economic adoption matters more, and that technical exchanges and industry safety associations offer practical routes to U.S.–China cooperation.

Read the source ↗ chinai.substack.com