I landed my first Point of Sale (POS) SaaS client project and built it inside Google Antigravity, trusting Gemini 3.1 Pro for all agentic coding. Bad move. Midway through, severe bugs hit—fixing sales inputs completely broke my dashboard analytics.
Frustrated, I switched to the GPT stack (Codex, GPT-5.5, GPT-5.6 Sol). But that brought a new headache: token quotas drained in minutes and API bills went through the roof.
The breakthrough? Community stats on Agent Arena showed Flash models beating 3.1 Pro in agentic tasks. Desperate, I tested Gemini 3.6 Flash and it was surprisingly solid.
1. The Beginning: Falling for 3.1 Pro Hype
When Google dropped Gemini 3.1 Pro in February 2026, tech Twitter hyped its ARC-AGI-2 benchmark scores as the ultimate problem solver.
Early on, 3.1 Pro handled high-level whiteboard concepts fine:
• System Architecture: Mapping schemas and ERDs.
• Business Logic: Outlining business logic flows.
• Edge Cases: Spotting potential edge cases.
2. Client Pressure & Regression Hell
I secured a major deal with a client who paid 50% upfront. Expectations were high, and that's where things went downhill.
A retail POS requires tight database relations: transactions, inventory sync, multi-outlet stock, and live dashboard revenue summaries. When I instructed Gemini 3.1 Pro in Antigravity to build the sales input module:
3. Fleeing to GPT & The Token Trap
Panicking about deadlines, I switched to the OpenAI ecosystem:
• **Codex / GPT-5.5 (Medium)**: Daily coding assistant.
• **GPT-5.5 (High Reasoning)**: Complex SQL logic.
• GPT-5.6 Sol: Heavy artillery for deep debugging.
While reasoning was solid, a new problem popped up: tokens burned out insanely fast. Running heavy models like GPT-5.6 Sol and 5.5 High across autonomous agentic loops burned quotas in no time. I needed a workhorse that didn't drain my wallet.
4. Community Outrage: It Wasn't Just Me
My frustration wasn't unique. On the Google AI Developers Forum, a thread titled "Unacceptable Antigravity Quotas for Gemini 3.1 Pro – Workflow Completely Blocked!" hit 62 upvotes and 19 replies in two days.
"Quota usage is totally intransparent, and the random 'refresh' policy is worse. It's like a car that sometimes works, sometimes it doesn't... ok for a hobbyist, but not for a cab driver." — Hauke_Walden (20 upvotes)
"My quota dropped from 60% to 0% without any prompt, with a reset time set to 50+ hours." — madnz-08 (10 upvotes)
"It is unacceptable to be stuck at 10 prompts on Gemini 3.1." — MrTos (25 upvotes)
The pattern was identical: vanishing quotas, lockout timers hitting 133–167 hours (over a week!), and random lockouts halting paying devs mid-project. Gemini 3.1 Pro's platform reliability was a total joke for professional work.
5. The Pivot: Testing Gemini 3.6 Flash
Hunting for a leaner alternative, I saw rankings on Agent Arena praising Flash models for agentic tool use.
Curious, I tested Gemini 3.6 Flash in Antigravity:
• Blazing Speed: Instant feedback loops without lag.
• Quota Sanity: No panic about hitting rate limits mid-session.
• Sane Diffs: It stayed in its lane without aggressively trashing outside files.
It proved lightweight models were getting unwarranted hate.
6. August: Brother's WhatsApp Tip & The Jump to 3.7 Flash
In August, my younger brother sent me a WhatsApp message: "Bro, Google just dropped Gemini 3.7 Flash."
As reported by VentureBeat, Google built 3.7 Flash specifically for coding and agentic loops, with a 50% introductory discount ($0.75 / 1M input, $3.75 / 1M output).
Having seen good results with 3.6 Flash, I wanted to see if Google really pulled off deep reasoning on a Flash model without killing speed.
7. The Proof: Gemini 3.7 Flash (High Mode) in Antigravity
I opened Antigravity, picked Gemini 3.7 Flash, and enabled Thinking Mode: High. The leap in quality was immediate:
Conclusion
This client project taught me a valuable lesson: in daily agentic coding, diff discipline, fast inference, and token efficiency matter way more than benchmark scores on paper.
From Agent Arena tests with 3.6 Flash to the leap in Gemini 3.7 Flash, it proved Flash models with thinking mode aren't just "cheap alternatives"—they are solid workhorses that handle heavy code cleanly and safely.
If you're building a SaaS and stuck between regression bugs or sky-high token bills, give Gemini 3.7 Flash a run in Google Antigravity.
Written by Achmad Junaedi — Founder of setvy.id & Vocational Educator at SMKN 10 Surabaya, Indonesia.