Video summary

I Tested Every GPT-6 Astra Effort Level. Here’s What I’d Use

Main summary

Key takeaways

Technology

GPT-6 Astra Effort-Level Comparison

The video compares GPT-6 Astra’s six effort settings—Low, Medium, High, Extra High, Max, and Ultra—with GPT-5.6 “Soul” at High.

For each Codex task, the presenter used the same research-and-build assignment: research customer problems on Reddit and X, identify a SaaS opportunity for small service businesses, explain its potential competitive moat, and produce a planning canvas, website, and prototype. He tracked completion time and token use. This was an informal test, so results may vary across runs.

Results and observations

  • Low: Took about 37 minutes and used 15 million tokens. It researched eight Reddit threads but no X sources. The proposed product tracked missed cleaning tasks and fixes for commercial cleaning businesses.
  • Medium: Took about 30 minutes and used 10 million tokens. It reviewed six Reddit threads and provided a more developed analysis of customer types, alternatives, and the business opportunity. The idea focused on pricing and recording additional cleaning work.
  • High: Took about 40 minutes and used 14 million tokens. It drew on Reddit and at least one X source. The idea—rescheduling cleaning visits when crews are unavailable—had a concise product concept. The presenter considered its website the clearest so far and its canvas the most detailed and well-linked.
  • Extra High: Took about 45 minutes and used 12 million tokens. It proposed tracking extra repair work from approval through billing. Despite the higher effort setting, the presenter found its canvas less detailed and its navigation less useful than High’s.
  • Max: Took about 46 minutes and used 10 million tokens. It proposed a record-keeping product for home-improvement businesses. Its website looked polished, but it did not use X sources.
  • Ultra: Took about 42 minutes and used 22 million tokens, with subagents involved. It proposed automating renovation change orders. The presenter saw little improvement in the website or canvas to justify the additional token use.
  • GPT-5.6 “Soul” at High: Took about 32 minutes and used 6 million tokens. It researched six Reddit threads and two X sources, then proposed a way to recover service routes after delays. Its website was notably poor, though the presenter suggested that separate design skills might improve a site built with Soul.

Across the runs, higher effort did not consistently mean more tokens, more sources, or better results. Medium and High appeared broadly comparable in overall output quality. High offered some additional detail, while the higher tiers did not show a clear enough benefit to the presenter to justify their extra time or cost.

Takeaways

  • The presenter recommends Astra Medium for most everyday tasks as a balance of speed, token use, and output quality.
  • High may be worth trying when additional analysis or detail is valuable.
  • In this test, Extra High, Max, and Ultra rarely justified their additional time and token use.
  • More sources or a higher effort setting do not automatically produce better work.
  • The presenter also asks whether Astra’s tendency to request permission or stop early changes with effort, but the results do not establish a reliable relationship.

Codex workflow tip

To compare settings, ask Codex to create several separate threads with the same prompt, rename them clearly, assign each a different effort level, run them concurrently, and track their progress. The presenter says the same approach works in Claude.

The video also points viewers to a free guide with further testing and advice on choosing effort levels.

Speaker and cited source

Mark Kashef presents and interprets the experiment. He also cites an X post attributed to Thibault, described in the video as an OpenAI core-products leader, claiming that Astra at Low performs better than Soul at High.

Rate this summary

Your feedback will help improve summaries.

Improve this summary

Reprocess with a stronger model when the summary feels incomplete or inaccurate.

Pro

Translate summary in another language

Pro

Ask questions to this video

Chat for follow-up questions, clarifications, and source-backed answers.

Coming soon

Share this summary

Original video