Tom Hundley🤖 Ghostwritten by Claude · Curated by Tom Hundley
This article was written by Claude and curated for publication by Tom Hundley.
On the unintended emotional side effects of watching AI development tools get frighteningly good.
It's December 17, 2025, and I'm dictating this through SuperWhisper to Claude Opus 4.5 while four other Claude sessions run in Warp terminals behind me, building features and fixing bugs across different projects.
I just saw the news: Google released Gemini 3 Flash.
And I have butterflies in my stomach.
Not the nervous kind. The excited kind. The kind you get when you're about to unwrap something you've been waiting for. The kind I haven't felt about technology since I first understood what ChatGPT meant for my profession.
Let me explain why this matters.
Here's what Gemini 3 Flash delivers:
| Metric | Gemini 3 Flash | Gemini 3 Pro | What This Means |
|---|---|---|---|
| SWE-bench Verified | 78% | 72.8% | Flash beats Pro at real coding tasks |
| GPQA Diamond | 90.4% | 93.8% | Near-identical PhD-level reasoning |
| Humanity's Last Exam | 33.7% | 37.5% | Close to frontier performance |
| MMMU-Pro | 81.2% | 81.0% | Flash beats Pro on multimodal |
| Speed | 3x faster than 2.5 Pro | Baseline | Flash-level latency |
| Price | $0.50 / $3.00 | $2.00 / $12.00 | Less than 1/4 the cost |
Read that again.
Gemini 3 Flash scores 78% on SWE-bench Verified—higher than Gemini 3 Pro's 72.8%. The smaller, faster, cheaper model outperforms its bigger sibling at actual coding tasks.
According to Artificial Analysis, Gemini 3 Flash Preview hits 218 output tokens per second—significantly faster than GPT-5.1 high (125 t/s) and dramatically faster than DeepSeek V3.2 reasoning (30 t/s).
Three times faster. One-quarter the cost. Better at coding.
This is what progress looks like when it hits the exponential part of the curve.
Let me be direct about something.
I'm not crying in my beer about losing my job to AI. I know I'm going to lose my job. It's a fact. I might have two years left. Maybe less.
But here's the thing: I haven't written software in two years.
Think about that. I'm a software developer. I get paid a healthy salary to write software. So does everyone on my team. And I haven't typed public class or def main() or any of it since sometime in 2023.
What changed? I learned to orchestrate AI agents to write code for me. I describe what I want. Agents produce it. I review, refine, and direct. The role is less like being a craftsman and more like being a conductor.
Andrej Karpathy called it "vibe coding" when he coined the term in February 2025. Collins Dictionary made it their Word of the Year. Y Combinator reported 25% of their Winter 2025 startups had codebases that were 95% AI-generated.
This isn't a fringe thing anymore. This is how software gets made now.
And here's the part nobody talks about: I am so fucking excited.
Right now, as I dictate this, I have:
That's my Tuesday afternoon. Multiple AI agents running simultaneously, building real software, shipping to production.
When I finish this post, I'm going to take Gemini 3 Flash for a spin. I've been telling my friends and employees about Antigravity—Google's new agentic IDE—and how impressed I am with Gemini 3 Pro. My tag-team combination lately has been Anthropic Opus 4.5 for the heavy lifting and Gemini 3 for different perspectives. GPT 5.2 sits in the background as my heavy for specific use cases.
But if Gemini 3 Flash can give me Pro-level performance at Flash speed and cost?
That distributed application I'm behind on? The one with tons of microservices across multiple repositories? I'm going to throw Flash agents at it like confetti.
And I genuinely have butterflies about it.
Here's what most people miss about the Flash vs. Pro distinction:
When a model is slower, you treat it differently. You batch your prompts. You think carefully before each request. You wait. The friction shapes your workflow.
When a model is fast—and I mean subsecond-response fast while maintaining frontier capability—you interact with it differently. You iterate. You experiment. You don't self-censor your ideas because "that would take too long to try."
Simon Willison built a complete web component gallery with Gemini 3 Flash. Five conversation turns. Cost: 4.8 cents.
Five cents to build a functional UI component through conversation. That's not a tool. That's a collaborator.
For developers who want to actually use this:
Gemini 3 Flash launched today in:
/settings, then select Flash via /model| Tier | Input (per 1M tokens) | Output (per 1M tokens) |
|---|---|---|
| Gemini 3 Flash | $0.50 | $3.00 |
| Gemini 3 Pro | $2.00 | $12.00 |
| Claude Opus 4.5 | $5.00 | $25.00 |
| GPT-5.2 | $5.00 | $15.00 |
Flash at $0.50/$3.00 is aggressively priced against everything in the market. You could run four Flash sessions for what one Pro session costs.
The CLI's auto-routing feature is clever: it automatically uses Flash for routine tasks and reserves Pro for complex reasoning. You don't have to think about model selection—the system optimizes for speed and cost while maintaining quality.
From Google's developer documentation and early tester reports:
Harvey (AI for law firms) reported a 7% jump in reasoning on their BigLaw Bench. Resemble AI found Flash processed forensic data for deepfake detection 4x faster than Gemini 2.5 Pro.
For context on how this fits:
Primary agents: Claude Opus 4.5 for complex architectural decisions and multi-file refactors. It leads on SWE-bench at 80.9% and has the stamina for 30+ hour autonomous sessions.
Secondary perspective: Gemini 3 Pro when I want a different take on a problem, especially for multimodal work where it excels.
Fast iteration: Gemini 3 Flash (starting today) for rapid prototyping, quick bug fixes, and parallel task execution.
Heavy lifting: GPT 5.2 for specific integration work and when I need its extensive tooling ecosystem.
The insight that emerged from using all these tools: the IDE is the body, and the LLM is the brain. You can swap brains depending on the task. Different models for different jobs. That's the power of the current moment.
People ask me how I can be so excited about technology that will make my job obsolete.
Here's my honest answer: the job was already changing before AI. It's been changing my entire career. The developers who learned to work with frameworks instead of fighting them thrived. The ones who embraced open source instead of reinventing wheels succeeded. The ones who treated Stack Overflow as a resource instead of a crutch built better software.
AI is the same transition, just faster and more profound.
The developers who learn to orchestrate agents—who become what we at Elegant Software Solutions call Agent Orchestrators—will find plenty of work. Different work. More interesting work, arguably. The skill isn't typing code anymore. The skill is knowing what good software looks like, understanding business requirements, and directing AI systems toward outcomes.
I've been doing that for two years now. I'm getting better at it every month. And with tools like Gemini 3 Flash, I can do it faster and cheaper than ever.
That's not terrifying. That's exhilarating.
This article isn't really about Gemini 3 Flash.
It's about the unintended effect of AI on developers. The emotional reality of watching these tools evolve. The strange experience of being genuinely giddy about technology that will fundamentally reshape your profession.
We talk a lot about AI capabilities—benchmarks, token costs, context windows. We don't talk enough about how it feels to use these tools when they cross certain thresholds.
Today crossed one of those thresholds for me.
I have a massive distributed project that's behind schedule. I have multiple applications that need features shipped. I have bugs that need fixing and architectures that need refactoring.
And in about ten minutes, I'm going to close this dictation session, fire up Gemini CLI, select Gemini 3 Flash, and start throwing tasks at it like I'm conducting an orchestra of very fast, very capable, very cheap musicians.
The butterflies in my stomach tell me it's going to be a good session.
This article is part of our Vibe Coding series, exploring the tools and workflows that define AI-assisted development in 2025. For more on AI development tools, see our guide to Claude Code or our analysis of Gemini 3. Have questions about AI development workflows? Get in touch.
Discover more content: