Video summary

I Tested GPT-6.1 Sol vs Astra. Here’s What I’d Use

Main summary

Key takeaways

Technology

Summary

The video compares GPT-6 Astra with GPT-6.1 Sol by asking each model to build Relay Voice, a Mac dictation app modeled on Wispr Flow. The app combines voice transcription with tool use: it can look up Gmail information, conduct web research through Firecrawl, and display tool calls and results in its interface.

App and build setup

  • The app supports a keyboard shortcut for dictation, a history and settings view, and a minimize mode.
  • Both models received the same main prompt and technology stack: Gemini for transcription, Firecrawl for web search, Google CLI for Gmail and other Google services, and Electron for the Mac desktop app.
  • The prompt also asked the models to use computer control to open the app and configure the required macOS permissions. Both could organize their work into separate project phases or branches.
  • A third version used a hybrid workflow, with Astra leading the planning and Sol handling implementation.

Demonstration and results

  • In the hybrid build, the app successfully retrieved Gmail messages and conducted web research, returning sources and a summary.
  • Astra’s standalone build handled voice input on the first try, but its research workflow returned several broken links. The interface also had limitations, including difficulty dragging the window.
  • Sol’s build looked simpler and had some latency and interface issues. Voice input required several rounds of prompting and troubleshooting, though its research links worked.
  • The standalone builds took nearly the same amount of time: about 42 minutes for Astra and 40 minutes for Sol. Their token usage was also similar, at approximately 69 million and 66 million, respectively.
  • The estimated costs were substantially different: about $203 for Astra versus $43 for the Sol-started build. The video also cites lower per-token prices for Sol, including cheaper input, output, and cached input.

Takeaway

The presenter favors a hybrid approach: use Astra for planning or for repairing and polishing problem areas, while using Sol for much of the implementation to reduce costs. Sol was considerably cheaper but needed more help with voice input; Astra was more reliable on that feature but had problems with research citations. The presenter describes GPT-6.1 Sol as a significant improvement over the earlier Sol model.

Guide and resources

The video walks through the app demos, prompt, technology stack, project phases, and comparison of time, token usage, and cost. The presenter says the Relay Voice code, prompt, and setup resources are available through links in the video description.

Main speaker/source: Mark Kashef, presenting his own tests of GPT-6 Astra and GPT-6.1 Sol. Other named technologies and services discussed include Gemini, Firecrawl, Google CLI, Electron, and Codex.

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