Theo - t3․gg

I can't believe they released this

Jul 14, 2026 26 min
ai modelsopenaicodexgpt-5.6-solsubagents
Watch on YouTube Follow Theo - t3․gg on Rundown — free

Summary

AI summaries can be incomplete or wrong. Verify anything important against the original video.

The creator reviews the "Codex" AI coding assistant and GPT-5.6-Sol models from OpenAI, highlighting issues with the "Ultra" setting and comparing V1 and V2 subagent implementations. They also briefly touch on the Firecrawl web scraping tool.

The video critically examines OpenAI's GPT-5.6-Sol models, focusing on the "Ultra" setting and its perceived flaws. The creator argues that "Ultra" is misleading and causes excessive token burn, leading to poor performance and cost inefficiency. They demonstrate how to use Codex CLI commands to select different effort levels, noting that "Ultra" isn't a standard reasoning level but rather a trigger for subagents to use more complex workflows, which can be detrimental. The creator also briefly compares the V1 and V2 subagent implementations in Codex, highlighting the shift from a worker pool model to a task tree model and the corresponding changes in communication and context handling. They touch on the Firecrawl web scraping tool as a potentially useful alternative for data extraction, praising its API and CLI capabilities, particularly its monitoring feature. The video concludes with a critique of OpenAI's subagent design, suggesting that a simpler approach might be more effective and less prone to issues like excessive resource consumption.

Verdict

GPT-5.6-Sol & Codex
AI models & coding assistant

The GPT-5.6-Sol 'Ultra' setting and Codex's subagent implementation have significant flaws that need rethinking, leading to inefficient and costly performance.

Wait

Pros

  • ✓ New multi-agent versions (V1 and V2) offer different approaches to agent management.
  • ✓ Codex CLI allows for selection of different effort levels (low, medium, high, xhigh, max, ultracode).
  • ✓ Firecrawl is a promising alternative tool for web scraping.

Cons

  • ✗ 'Ultra' setting is inefficient and costly. 4:09
  • ✗ Codex V2 subagent implementation is problematic.
  • ✗ Effort level slider and 'Ultra' setting in Codex are misleading.
  • ✗ The version of subagents Sol uses does not allow for selection of reasoning levels.
  • ✗ Running subagents on 'Max' effort leads to wasted resources.

Best for

  • Developers exploring advanced AI models
  • Users concerned about AI costs and efficiency

Claims & arguments

  • GPT-5.6-Sol "Ultra" Setting

    The 'Ultra' setting in GPT-5.6-Sol is inefficient and costly due to excessive token burn.

    • 6:16 Causes massive token burn for no good reason.
    • 14:16 The creator burnt through their 5-hour limit in 20 minutes using Ultra.
    • It leads to slow responses or overthinking tasks.
  • Codex Subagent Implementations

    The V1 and V2 subagent implementations in Codex have significant differences and current issues.

    • V1 is like a dispatcher hiring temporary helpers; V2 is like a named project team.
    • V2's implementation can be inefficient and lead to errors.
    • The version of subagents that Sol uses does not (currently) allow for selection of reasoning levels, leading to all subagents being on Max.
  • Codex CLI Effort Levels

    The Codex CLI's effort levels are not intuitive, particularly the 'Ultra' setting.

    • Ultra is not a reasoning level, but a workflow trigger that can lead to undesirable behavior.
    • The gradient on the slider is misleading; Ultra is not 'max' effort.

Key Points

  • 1:45 OpenAI's GPT-5.6-Sol models introduce new reasoning levels, including 'Ultra'.
  • 4:09 The 'Ultra' setting is criticized for being inefficient and costly, leading to excessive token burn and slower responses.
  • 18:10 The creator hopes for future improvements and suggests waiting for future model updates.
  • The creator demonstrates using Codex CLI commands to adjust effort levels, noting that 'Ultra' is not a reasoning level but a workflow trigger.
  • The video contrasts V1 and V2 subagent implementations in Codex, highlighting V1's worker pool approach vs. V2's task tree model.
  • Firecrawl is introduced as a potential alternative for web scraping, praised for its features like API, CLI, and monitoring.
  • The creator believes OpenAI needs to rethink their subagent design, as the current implementation can be inefficient and costly.
  • The current implementation forces all parent and subagents to run on 'Max' effort, potentially leading to wasted resources and errors.

Worth watching if: If you're interested in the practical performance and cost implications of using advanced AI models like OpenAI's GPT-5.6-Sol and Codex, this video offers a critical analysis of their 'Ultra' setting and subagent implementations. It's also relevant if you're looking for alternatives in AI coding assistants or web scraping tools.

Get every Theo - t3․gg video extracted like this

One daily email with structured extracts of every channel you follow. Free tier covers 15 videos a month.

Sign in with Google

No credit card. Free tier forever.

Watch on YouTube