Reason's chart shows Codex running GPT-6 Astra xhigh, gains attributed to harness
The author notes that Reason's own chart labels Codex as running GPT-6 Astra xhigh, and argues the extra 3 points and ~$1 saved per task come from the harness rather than the model, while questioning the inconsistent benchmark name and lack of independent reproduction.
Read the label under Reason's star on their own chart: GPT-6 Astra xhigh. That's the same model Codex runs at the point it beats. So the 3 extra points and the ~$1 saved per task come from the harness: how it plans, what it reads, when it stops. That's where I'd look for cheap gains this year too, not in the next model. Two things I want before I fully buy it. The tweet says Terminal-Bench 4.0 but the chart's axis says Coding Agent Index. And it's their own run. If someone independent reproduces it, every agent team should be studying this harness.
This post quotes another post: “Reason launches as a free terminal agent harness, 22% cheaper than Codex+Astra”