GPT-6 Astra on Medium: Why Codex Feels Easier to Work With

The most noticeable part of GPT-6 Astra in Codex, for me, is the interaction. On Medium, it feels easier to work with than the other models I have been using. The responses feel useful, the output quality feels strong, and the agent appears willing to make a decision and carry the work forward.

That combination matters more to my experience than a single impressive answer. I want the conversation and the execution to fit together. Astra has felt particularly comfortable in that respect during my September 1–5 observation window. This is my assessment of the experience, not a benchmark establishing that it outperforms every other agent.

What the official guidance supports

OpenAI lists GPT-6 Astra for work across coding, computer use, research and documents. Its model documentation explicitly includes Medium reasoning effort. Current client guidance describes Medium as balancing speed and depth, and warns that model options depend on the account and rollout.

OpenAI also describes Astra as better at incorporating guidance and answering side questions while retaining the original goal. That product description is consistent with the quality I appreciate, but it is not independent evidence of my own results.

A setting that works for my experience

Medium is the setting behind this note. I have not supplied matched runs on Low, High or Extra High, so there is no basis here for calling Medium universally optimal. My narrower point is that I already find the collaboration and resulting work satisfying at this setting.

Willingness to act is valuable when the agent understands the intended outcome. It still leaves me responsible for judging whether the decisions and final result are appropriate. A confident completion message is not, by itself, evidence that a task was completed correctly.

Other experiences add a cost question

Matt Shumer's first-person Astra review describes sustained work across applications and a positive experience with its computer use. He also notes that ambitious long runs still require careful coordination and can lose momentum. That is another user's account, not a result reproduced for this article or a controlled comparison.

A separate Reddit user reported meaningful progress and commits on a project, but also rapid consumption of their allowance with Astra Medium. They had not yet assessed the finished changes. This is a useful counterpoint: an agent can feel productive while its usage cost still makes long sessions difficult. One report cannot establish a typical consumption rate.

My current conclusion

Astra on Medium feels like a strong fit for the way I want to collaborate with an agent: useful interaction, strong output and willingness to follow through. To turn that impression into a comparison, I would need a defined task, saved outputs, correction counts and usage records. For now, this is a personal reason to keep using the setting, not a general performance or value ranking.