Gemini 3.5 Flash is just... fine

Summary of technological concepts, features, and analysis

Gemini 3.5 Flash release (main topic)

• Claimed performance vs. cost (marketing): Google positions Gemini 3.5 Flash as having “frontier” performance at ~4× the speed and under half the cost of competitors. The video argues real-world results don’t match these claims.

• Model specs / modalities:

  • Context window: 1 million tokens
  • Output limit: 64,000 output tokens
  • Inputs supported: text, images, video, audio, and PDFs
  • Emphasized as strong on multimodal workloads (consistent with Google’s historical strengths)

• Coding & benchmark performance (mixed, per the speaker):

  • On Google’s own benchmarks, Gemini 3.5 Flash is reported as:
    • roughly in line with GPT 5.5 for coding, with small differences on SW Bench Pro and Terminal Bench
    • outperforming Opus 4.7 on Terminal Bench by ~10%
    • while Claude Opus beats Gemini on SW Bench Pro by ~10%
  • Agentic benchmarks: described as winning on MCP and Toolathon benchmarks (per the speaker)

• Third-party benchmark discrepancy (where results “worsen”):

  • Using Artificial Analysis (third-party), coding performance is described as:
    • ~45 on the “coding index”
    • below models like Kimiko 2.6
    • not beating Gemini 3.1 Pro, despite Google’s internal benchmarks showing the opposite
    • only slightly above Gemini 3 Flash

• Speed is the standout:

  • Reported 278 tokens/second, described as significantly faster than Opus 4.7, GPT 5.5, Haiku, and some open models.
  • The speaker concludes it offers the best “intelligence vs. speed” balance if speed is the priority.

• Cost analysis (major critique):

  • Published prices:
    • $1.50 per 1M input tokens
    • $9 per 1M output tokens
  • Measured cost (via Artificial Analysis), per the speaker:
    • $1,552 to run the “intelligence index”
    • ~5.5× more expensive than Gemini 3.5 Flash
    • ~75% more expensive than Gemini 3.1 Pro
    • more expensive than GPT 5.5 on high reasoning (the speaker says GPT 5.5 beats Flash on coding)
  • Conclusion: the video argues it is not “half the cost” and can be worse than cheaper models that also code better.

• Token-hungry behavior (why cost rises):

  • In agentic evaluation, it averaged ~49 turns per task, suggesting it burns through input tokens.

• Overall verdict on Gemini 3.5 Flash:

  • Labeled as “meh” / mixed bag
  • Suggested mainly for agents (speed + agent benchmarks)
  • Not recommended as an overall best coding model due to:
    • weak coding quality vs alternatives
    • potentially unfavorable cost

Second major announcement: “Antigravity 2” + new CLI

• Antigravity 2 app (standalone agent IDE/app):

  • The speaker says it’s become hard to distinguish from other “agent coding” tools (compared visually to Codex/Cursor/etc.).
  • UI elements described:
    • Conversations (left)
    • Projects
    • Scheduled tasks
    • Ability to open files and view diffs

  • Key point: it’s not an “Anti-Gravity IDE” anymore, but a standalone app.

  • Test prompts demoed:

    • Cafe website (simple prompt):
      • Produces a functional single index.html
      • Speaker likes the UI/design output, suggesting Gemini 3.5 Flash performs well at UI design
      • Notes it can have an “AI look” (e.g., card/gradient style)
      • Claims it looks better than what Opus 4.7 produced in a one-off test
    • Full-stack personal finance dashboard (complex prompt):
      • The app “works,” but UI looks AI-generated
      • Negative reaction to the name “Aura Wealth”
      • Speaker says Opus 4.7 produced a much nicer UI
      • Also mentions time spent: ~20 minutes for Opus vs ~5 minutes for Gemini, implying Flash is faster but doesn’t invest enough time to polish UI

• Anti-Gravity CLI:

  • Gemini CLI shutdown: speaker says Gemini CLI will be unavailable after June 18
  • New CLI characteristics:
    • rewritten in Go
    • closed source
    • functionally similar to the old CLI at launch (no major new capability demonstrated)
  • Speaker frames it as “Killed by Google”-style churn and expresses disappointment about the closure of the codebase.

Product/market positioning conclusion

• The speaker suggests Google may be prioritizing everyday consumer integration (Gmail, Search, Workspace, Android, etc.) more than being the top developer/coding leader.
• Expectation:
  • Gemini 3.5 Flash is not the clear coding winner
  • Gemini 3.5 Pro is mentioned as coming next month as a potential improvement

Main speakers/sources

• Main source/speaker: the video narrator/reviewer (author of the analysis)
• Referenced/compared model providers:
  • Google (own benchmarks)
  • Artificial Analysis (third-party benchmarks)
  • Competitors: OpenAI (GPT 5.5), Anthropic (Claude Opus 4.7)
  • Others mentioned: Kimiko 2.6, Haiku, and open-source OpenAI models
• Product sources mentioned:
  • Gemini CLI, Antigravity 2, Anti-Gravity CLI
  • Other coding-agent apps referenced by name: Codex, Cursor