I'm scared to make this video

Summary of Technological Concepts, Product Features, and Analysis

The speaker frames the video as a warning about Google’s product direction and ecosystem changes, with emphasis on model performance vs. cost/accountability, CLI tooling reliability, and cloud service uptime.

---

1) Gemini 3.5 Flash — performance claims vs. cost/efficiency critique

Performance benchmarking (speed + quality)

The speaker argues that Gemini 3.5 Flash outperforms Gemini “31 Pro” in almost all tasks, with an exception on SkateBench where 31 Pro allegedly performs better.

They also highlight Google’s apparent push toward stronger agentic/workflow capability, citing results from:

• Coding-style / terminal-style benchmarks (Terminal bench scoring described as top-tier, near frontier models like GPT 5.5)
• SWAG Bench Pro (still beating 31 Pro, but not matching all higher-end models)
• Tool/agent evaluations such as Toolathon and finance-agent-style tests
• A multi-benchmark run including MMMU Pro and other tasks

“Intelligence index” / speed-to-performance emphasis

They note a high “speed to performance” ratio, presenting Gemini 3.5 Flash as extremely fast (near ~300 tokens/sec) and positioning this speed as a key advantage.

Main critique: pricing & token economics are misleading

The speaker claims the pricing presentation is intentionally opaque (e.g., no dollar sign shown on the page) and asserts Google tripled Flash-tier prices.

Reported pricing:

• $1.50 per million input tokens
• $9 per million output tokens

Comparison to older models (as cited by the speaker):

• Gemini 3 Flash: much cheaper (approx. ~$3 out / $0.50 in)
• Gemini 2.0 Flash: even cheaper (approx. ~$0.10 in / $0.40 out), described as >20× increases

Token efficiency criticism

They argue Gemini models consume far more tokens during benchmarks than competitors (example figures mentioned for models like DeepSeek v4 flash, GPT 5.4 mini, Sonnet, etc.).

Even though raw “efficiency” may appear slightly worse-but-close between older Gemini Flash variants (example: 72M vs 73M tokens mentioned), the speaker claims overall cost worsens because the model burns additional tokens and produces more output tokens.

Cost conclusion

They conclude Gemini 3.5 Flash ends up among the most expensive models in the comparisons being discussed, and that being “fast” does not justify the higher token costs.

---

2) “Anti-Gravity CLI” vs. “Gemini CLI” — workflow tooling changes

Sponsor-independent tooling preview: Trigger.dev

A sponsor segment covers Trigger.dev, described as providing:

• Real-time visibility into job/agent workflow execution (clickable process steps)
• Multi-step agents broken into smaller JavaScript functions for image/text pipelines (with TypeScript emphasized)
• Support for local development servers as workflows become more complex

Gemini CLI replacement / closed-source transition (speaker’s claim)

The speaker alleges Google is transitioning “Gemini CLI” into “Anti-Gravity CLI.”

Anti-Gravity CLI is described as:

• Rewritten from scratch in Go (Golang)
• Focused on invoking, monitoring, and interacting with Anti-Gravity agents
• Not positioned as a full GUI—more as a fast, lightweight CLI

The speaker further claims:

• Gemini CLI and Gemini Code Assist IDE extensions will stop serving requests for Google AI Pro/Ultra
• Users in those subscription plans can supposedly only use models through Anti-Gravity going forward

Open-source complaint

They argue the previously open-source Gemini CLI is effectively removed/closed in favor of a closed-source Anti-Gravity CLI, harming community iteration and trust.

Usability & bug reports from Anti-Gravity CLI

The speaker reports multiple issues, including:

• Broken terminal UI scrolling
• Control-C doesn’t exit; requires a special “slash exit” command
• Potential email exposure (their email allegedly appears in the UI/logs)
• UI instability (input box “moving all around”)
• Freezing/locking during generation
• Random/unreliable UI states and feedback prompts

Overall claim: it is the “buggiest CLI” they’ve used recently and doesn’t function as expected.

---

3) Anti-Gravity model “coding” test results (agentic coding quality)

Game rewrite/refactor with source access

The speaker runs a task: rewrite/refactor a game (“Fish Slap”) using source access and rebuild it more stably.

Claimed result:

• Gemini 3.5 Flash failed:
  • produced broken code
  • did not verify/run checks
  • improvements are described as malfunctioning (game mechanics not working)
• Some output included images, but the speaker describes them as low quality and incorrectly rendered (including transparency issues)

Comparison

A different model (GPT “5.5” mentioned) reportedly handled the task much better, including producing a working 3D version.

---

4) Google Cloud / Railway outage — account ban and reliability concerns

Core incident claim: Railway outage

The speaker claims Railway’s service went down because Google Cloud banned/disabled Railway’s account.

Additional details claimed:

• Railway is described as paying $2 million+ per month
• They argue the web-facing layer (CDN/web layer) went offline, even if some underlying infrastructure may remain

Pattern-based argument: cloud mishaps

They cite historical examples, including:

• An Australian customer allegedly having their cloud subscription deleted due to misconfiguration, causing severe disruption

They compare Google Cloud unfavorably to AWS/Azure in terms of responsiveness and reliability (as characterized by the speaker).

---

5) Meta/organizational critique (developer ecosystem impact)

The speaker attributes issues to internal politics and organizational prioritization, claiming:

• Teams/groups that built open tooling/community have been replaced or deprioritized

They also mention “biting Codex” style behavior shown in demo content (a referenced Codex folder), framing it as embarrassing for Google’s internal strategy.

Trust summary

They argue their lack of trust is driven by repeated problems:

• Misleading benchmark presentation
• Token economics
• CLI quality/bugs
• Cloud disruptions

---

Key Reviews / Guides / Tutorials Mentioned

Trigger.dev workflow guide (sponsored segment)

Described content includes:

• Orchestrating agentic workflows with real-time step visibility
• Building pipelines from simple JavaScript functions
• Example workflow ideas such as:
  • image generation → filter → upscale → send result
  • text prompt refinement steps

Link mentioned: soydev.link/trigger

No other formal “how-to tutorial” sections appear beyond the sponsor workflow explanation and the speaker’s “I tested X” analysis.

---

Main Speakers / Sources

• Main speaker: The YouTube channel creator (unnamed in the subtitles). The summary notes references via Google’s earlier actions on “their channel,” and mentions “Demetri, Jack, and Gal” as internal contacts.
• Companies/products referenced: Google (Gemini 3.5 Flash, Anti-Gravity CLI, Gemini CLI, Gemini Code Assist extensions, and the Anti-Gravity/open-source transition), Railway, and Google Cloud.