原帖引用 · 原文
Rohan Paul@rohanpaul_ai
@vnjogani afaik, Reason adds a codemode tool where the model writes a JavaScript that runs a search, file reads and tests in parallel and returns only the useful result
@rohanpaul_ai I dont get the how. Isnt that what every harness does? Leverage caching by keeping history stable?
在 X 查看这条帖子 ↗
Reason Machines just launched Reason, a coding harness that wraps a single model like GPT-6 Astra and solved tasks 22% cheaper than OpenAI’s Codex.
How: A traditional coding agent resends its context to the model on every step, so Reason trims that load:
Reason holds the prompt prefix identical between turns, by putting its fixed instructions and small tool list at the top of every prompt and only adding new code and results at the end, never rewriting or reshuffling the earlier text.
Since every turn starts with the exact same text, OpenAI recognizes it from the previous turn and bills those repeated tokens 90% cheaper, at $1 instead of $10 per million.
在 X 查看这条帖子 ↗
· UTC+8