Astra's 99.9% Score Came From the Harness, Not the Model
OpenAI's ARC-AGI-3 results sheet lists 62.7% and 99.9% side by side for the same model. Tracing why one model ends up with two scores reveals who actually deserves credit for this result.
Amazon's cloud platform. It rents out servers that AI companies use to run their models, and sells models from multiple companies through Bedrock. Its flagship services include EC2, the computing infrastructure, and Bedrock, the AI model service. In the case of the Astra model, which drew attention in a September 2026 benchmark, some assessments suggested that the deciding factor in performance was not the model itself but the harness surrounding it—that is, its execution environment.
Current rank (1M)
66
Rank over the last 6 days
2026.09.08 – 2026.09.13 · High 65 · Low 74 · Now 66
OpenAI's ARC-AGI-3 results sheet lists 62.7% and 99.9% side by side for the same model. Tracing why one model ends up with two scores reveals who actually deserves credit for this result.
DuckLabs, which spent five years without venture funding, is joining AWS in early September, and DuckDB stays under its MIT license
Princeton and UC San Diego researchers ran 8,135 experiments to uncover what actually makes skills work — and where they fall short
MCP, the standard for connecting AI tools, has revised its roadmap just five months after the last one. Servers no longer need to hold onto state, and agent identity authentication is up next.
That's the last story.