CharlieKey takeaways from your favorite podcasts
← 11 November 2026AI & I

LIVE Vibe Check: Gemini 4 Carbon

The Every team's live vibe check finds Gemini 4 Carbon a strong, diligent, natively multimodal second-opinion model that is not yet a daily driver, with dense design output and weaker writing.

Watch on YouTube ↗

Charlie's summary

  1. Carbon is the model you reach for as a second opinion, not your daily driver

    The host says Gemini has firmly become one of the three models the team tests, because it seems uncorrelated with the other two and comes up with different ideas, sometimes worse but often fresh. Randy found that having Fable and Carbon review each other's work unlocked complicated tasks where one model alone got stuck on one path. The team's shared verdict: an “amazing second opinion”, not a daily driver.

    ▶ 0:00
  2. Carbon writes maintainable code, and Kieran ranks it second for coding

    Kieran now reaches for Carbon more than OpenAI's model and places it behind Anthropic's. He says it follows best practices on large code bases and favors maintainable code over lean code, though it is still more verbose than Fable and has a new personality to learn. The jump from Argon is large, and it verifies its own work reasonably well.

    ▶ 2:02
  3. Google has made the model diligent and long-running, to the point of timing out

    On the host's “junior engineer benchmark” (a frozen Notion backlog of about 100 tasks), other models finished in an hour with 8 or 9 tickets. Carbon did 15 or 16 tickets over four and a half hours before he stopped it. It also timed out on tasks in his dashboard benchmark because it was unwilling to stop. The host says it is best run in the background, and the team now has to tell agents to stop rather than keep going.

    ▶ 5:39
  4. Native multimodality lets Carbon catch audio and video flaws other models miss

    Randy found that Fable, Astra and Opus repeatedly missed audio joins where a line was clipped off in Every Cut's timeline, while Carbon spotted it at once and opened a pull request to fix it. The host explains that Claude slices video into scenes and transcribes audio, whereas Gemini is natively multimodal. His workaround is to give Claude a Gemini API key for anything multimodal, and Randy now runs a Gemini pass on every video output.

    ▶ 8:13

Give your email to unlock the rest of the key points — for every episode.

Do more with Charlie

Free — sign in with your email, no password.