Looking Back at May — From "It Works" to "I Want It Still Working Three Months From Now"
This is Masaki Hirokawa from Gemini Lab.
May 2026 was less a month of flashy announcements and more a month I spent writing about the quiet, unglamorous design work that keeps things running long after launch day. The articles I published here in May, and how I spent my days outside the editor, both leaned in that direction with unusual consistency.
A lot of that came down to timing. The tense air before Google I/O 2026, the official release of Gemini 3.2 Pro and Flash, and the v2.1.0 rollout of my six Android wallpaper apps plus four iOS app updates all landed in the same four weeks. Instead of chasing new features, I kept asking myself the same question on most days: "Will the future me, three months from now, be glad I designed it this way?"
I started building apps independently in 2014, after first discovering the internet at age sixteen in 1997, and twelve years on, I can say May was a month where my relationship between indie development and AI quietly grew up a little more.
The Main Threads of May
Gemini 3.2 Pro / Flash Goes Official, and a 7-Day Migration
The headline of the month was the official launch of Gemini 3.2 Pro and Flash at the start of May. The gemini-3.2-pro and gemini-3.2-flash model IDs stabilized for production use, and I noticed a real, hand-feel improvement in long-context reasoning and first-token streaming latency.
Right after the release I published Gemini 3.2 Complete Guide — What Changed, How It Actually Feels, and How It Differs from the Previous Model, followed by the production-oriented Migrating from Gemini 2.5 Pro to Gemini 3.2 Pro in 7 Days — A Production Playbook.
What I personally appreciated most was that the thinking_budget parameter became reliable enough to depend on. As I wrote in Three Months with Gemini 3.1 Pro — An Indie Developer's Honest Review, being able to dial reasoning depth against cost makes Gemini much easier to embed inside production apps whose AdMob revenue swings month to month.
A Calm Comparison Against Claude Sonnet 4.6 and GPT-4o
Once 3.2 settled, I quietly re-ran my entire operations flow for my four iOS apps — Beautiful HD Wallpapers, Ukiyo-e Wallpapers, Relaxing Healing, and Law of Attraction Everyday — across Claude Sonnet 4.6, GPT-4o, and Gemini 3.2. That comparison became Gemini 3.2 vs Claude Sonnet 4.6 vs GPT-4o — An Indie Developer's Honest Comparison, May 2026.
The short answer is that there is no single winner. Short-form work like multilingual review replies fits Flash-tier models comfortably; long-form design reviews lean toward Pro-tier models; iterative coding loops still feel best with Claude Sonnet 4.6. Rather than picking "the one best model," I find running several in parallel — almost like a multi-track studio — fits indie work much better.
Cost Optimization Got Serious
May was also the month I genuinely went after API cost.
In How I Cut My Gemini API Bill from ¥30,000 to ¥6,000 a Month — A Production Guide to Context and Implicit Caching, I wrote up the caching patterns that worked on my own usage logs. The headline number sounds aggressive, but for use cases like "same system prompt + different user context" (which is roughly 90% of my wallpaper recommendation traffic), the savings show up exactly as expected.
On the vector search side, Cutting Gemini Embedding output_dimensionality from 768 to 256 Reduced Our Vector DB Storage by Two-Thirds does what the title says. The accuracy degradation was within practical tolerance for my recommendation flows.
And at the very end of the month, Classifying 8,000 App Reviews Overnight with the Gemini Batch API — Implementation Notes captured a real overnight run against the 8,000+ reviews accumulated across the 50-million-download wallpaper portfolio. Batch turned out to be dramatically cheaper than realtime, and "it finishes while I sleep" fits the rhythm of indie development surprisingly well.
Production Stability — The Quiet Equipment That Keeps the Lights On
Bookending the month were two clusters of production-stability articles.
Early May: Designing a Multi-LLM Failover Architecture Around Gemini API, Rotating Gemini API Keys with Zero Downtime in Production, and Before You Send PII to Gemini API — A Production Guide to Redaction, Audit Logging, and Operations.
Late May: Adding Circuit Breakers and Graceful Degradation to Gemini API — Design Notes from Indie Apps, Evolving Gemini Structured Output Schemas Safely in Production — A Design Record, and Continuous Quality Monitoring for Gemini API — A Golden Dataset and LLM-Judge Evaluation Architecture.
What I realized while writing these was that they weren't really articles about "how to use Gemini" — they were about helping the version of me three months from now. Both of my grandfathers were temple carpenters, and the feeling they passed down — that working with your hands is its own form of devotion — kept surfacing as I wrote about backoff timers and dual-emit schemas.
Art × Gemini — Turning Handmade Photo Collages into 120 Wallpapers
In week three, I wrote about my own decade-plus practice as a photo collage artist and how I now use Gemini 3.2 Pro and Imagen 4 to convert existing artwork into wallpaper-ready material at scale.
The two articles were 30 Days Expanding My Own Artwork into 120 Wallpapers with Gemini 3.2 Pro and Imagen 4 — An Artist-Indie-Developer's Asset Pipeline and Showing My Own Artwork to Gemini Vision — An Honest Review from a 17-Time International Art Award Recipient.
I want to be careful about one point. I do not use AI to create the photo collages themselves. The pieces that earned the 17 international awards were all assembled by hand. Where Gemini comes in is the operations layer — preparing existing works as multi-resolution assets that can be delivered through the wallpaper apps. Twelve years of running indie apps alongside an art practice have taught me that creating a work and releasing a work are two distinct jobs, and keeping that line clean has become a load-bearing principle for me.
The implementation side, against the 50-million-download wallpaper portfolio (Beautiful HD Wallpapers, Ukiyo-e Wallpapers, Relaxing Healing, Law of Attraction Everyday, plus two Android counterparts), shows up in Using Gemini Function Calling as a Recommendation Engine in a 50-Million-Download Wallpaper App and Parallelizing Gemini API with asyncio Made Multilingual Wallpaper Captions 12x Faster.
AdMob × Gemini — The Article That Paid for Itself by Month-End
Building a Gemini API Pipeline That Tells Me at 8 AM If AdMob Revenue Just Dropped from late week three, and Designing a Weekly AdMob Floor-Price Candidate Pipeline with Gemini 2.5 Pro from week four, paid me back in peace of mind more than anything else.
I have been operating AdMob since 2014, and there was a stretch where peak monthly revenue passed ¥1.5M. The flip side is that any sudden drop translates directly into living expenses. Automating "anomaly detection → an 8 AM ping → a rough first guess at the cause" via Gemini turned out to matter less as an engineering achievement and more as a way to sleep slightly better.
The Air Right Before Google I/O 2026
Toward the end of week three I published Getting Ready for Google I/O 2026 — A Gemini API Developer's Notes, capturing my expectations and quiet worries on the eve of the keynote.
I covered the actual announcements elsewhere, but leaving a clean before-and-after record turned out to be unexpectedly useful for calibrating my own forecasting habits. It's a small ritual I want to keep up next month.
A One-Week Trial of Gemini Computer Use
In the last stretch of the month I ran Gemini Computer Use and Claude in Chrome together for about a week, automating parts of the AdMob and App Store Connect consoles. I touched on it briefly in the week-four highlights. My honest read is "very useful" and "still a little scary" in roughly equal parts.
For billing and payment screens specifically, I would not run any of this without explicit guardrails. The pleasant surprise was that for tasks built around a human watching nearby — designer work, release-day screenshots, AdMob mediation tuning — it suddenly becomes a serious partner. I want to keep refining where exactly that line sits in next month's writing.
The Articles That Got the Most Reach — with a Few Editorial Notes
- Gemini 3.2 Complete Guide — What Changed, How It Actually Feels, and How It Differs from the Previous Model — Read steadily from launch week through month-end. As advertised, it is built on hands-on testing against 3.1.
- Gemini 3.2 vs Claude Sonnet 4.6 vs GPT-4o — An Indie Developer's Honest Comparison, May 2026 — I suspect the "don't crown a winner, split by use case" angle is what kept readers around.
- How I Cut My Gemini API Bill from ¥30,000 to ¥6,000 a Month — A Production Guide to Context and Implicit Caching — A lot of readers were stopped by the headline number. The internals are mundane; the effect is not.
- 30 Days Expanding My Own Artwork into 120 Wallpapers with Gemini 3.2 Pro and Imagen 4 — More readers than I expected were willing to follow me into the seam between art and indie development.
- AdMob + Gemini API Monetization Patterns for Indie iOS Wallpaper Apps — A design article rebuilding twelve years of AdMob operations through a Gemini lens.
What I'm Watching in June
A short look ahead.
First, the maturing Gemini 3.2 ecosystem. The "fast and stable" story is the headline today, but I expect third-party SDK wrappers and Function Calling template collections to multiply in June.
Second, safe-operation knowledge for Computer Use. Vendors will start publishing guidance on what kinds of automated screen actions are reasonable. As an indie developer, I want my own guardrails around billing flows in place before that happens.
Third, Imagen and Veo as a pair. Week three's wallpaper pipeline made the Imagen 4 still + Veo short-clip combination feel within reach. Loading screens and other quiet in-app surfaces look like good places to try this.
Fourth, syncing with the iOS release cycle. My four iOS apps are still working through new iPhone resolution support, the StoreKit 2 migration, and the CocoaPods → SPM move. The threads connecting Gemini to AdMob tuning and Crashlytics analysis will keep getting thicker in June.
Fifth, the ongoing shape of Gemini Lab itself. I want to keep "I want this still working three months from now" articles and "this saved me today" articles distinct, without letting either crowd out the other.
Closing — See You Here in June
May at Gemini Lab was less about chasing new features and more about writing down long-running production designs, mirrored against real apps I own. The artist's theme I have been carrying for years — exploring the architecture of collective psychology and shared consciousness against a backdrop of Japanese forms of prayer — sat on the same desk as twelve years of grounded indie development, both mediated by Gemini as a single tool.
I will keep writing here in June, from my own working examples and from what I notice when I do the work by hand. Thank you to those of you who have followed along from earlier months, and to those who happened to find your way here this month. I would be glad to have your company a little longer.