Google I/O 2026 — what stood out to our team

Google I/O is the year's clearest signal of where the platform is heading. We watched the keynotes, dug into the developer sessions and pulled out the four threads we think actually change how we build for clients in the next 12 months.

This isn't a transcript — Google's announcements page covers that. It's our so-what filter: what's going to land in real production code, what's a year out, and what's noise.

1. Gemini keeps getting more capable, and cheaper

The headline every I/O for the last three years has been the next Gemini family. This year was no different, with another step-change in reasoning, longer context and stronger tool use. The pattern that matters to us is the cost curve: features that were premium two model generations ago now sit in the mid-tier — which is the tier most production AI features should actually run on.

What that means for us:

• We're more aggressive about defaulting to the mid-tier for everything except the genuinely hard reasoning steps.
• Long-context use cases that were uneconomic in 2024 (think: feed-the-whole-repo-and-ask-it-questions) are increasingly viable. We're now writing them into RAG-replacement evaluations rather than RAG-mandatory ones.
• Tool calling reliability keeps climbing — the gap between "demo works" and "shipped agent works on 95% of inputs" is finally narrowing without bespoke retries and re-prompts.

The honest catch: cheaper doesn't mean free, and the new models aren't universally better at everything. We still run our own eval suite on every model swap. The press release isn't the eval.

2. On-device AI gets serious on Android

Android's bet on running models on the device — not the cloud — is finally hitting a useful tier. The combination of stronger NPU hardware and smaller, more capable models means the latency and privacy story for on-device features has improved meaningfully.

For us this matters in three concrete ways:

• Privacy-sensitive workflows (legal, medical, financial assistive features) become viable to ship without a server hop.
• Latency-critical UX (live transcription, suggestion-as-you-type, camera-frame analysis) feels native rather than tethered.
• Cost — for high-volume, low-margin features, on-device inference flips the unit economics.

The constraint hasn't changed: you have to think about model size, memory pressure and battery from day one. But the envelope of "what can ship on a mid-range Android phone in 2026" is a lot bigger than it was in 2024.

3. The web platform is quietly catching up

The web platform announcements at I/O rarely make headlines, but they cumulatively matter more than any single Gemini delta. This year we paid attention to:

• Continued Core Web Vitals tightening — the INP (Interaction to Next Paint) metric is now firmly the bar to clear, and the budgets are tighter. Our build defaults already pass it; teams running on older stacks should plan a migration.
• Better native streaming UI primitives — closer alignment with what we already do with React Server Components, Vercel AI SDK and friends.
• WebGPU is finally something to plan around — not yet a default, but for the right workload (in-browser inference, complex visualisations) it's no longer a science project.

The pattern: the web is becoming a more capable surface for the kind of AI-augmented products we build, while staying the most accessible distribution channel on the internet. We're not abandoning native — but the "ship a web app first, native if you must" calculus keeps getting stronger.

4. Developer tooling: AI in the inner loop

The most under-appreciated thread at I/O this year was developer tooling. The IDE story, the test-generation story, the codebase-aware assistants — all of it took a step forward. We're already deep users of AI-augmented dev workflows, and the new generation closes some of the rough edges:

• Refactors that span multiple files that actually understand the architecture, not just the syntax.
• Tests that interrogate intent, not just lines of code.
• Documentation that stays in sync with the codebase, generated on commit.

For us, this changes the math on what we can deliver in a 4-week sprint. A senior engineer with the right AI dev loop now does what a 2-person team did 18 months ago — without losing the architectural rigour that the second person was supposed to bring.

What we're not doing yet

A few things we saw and are not (yet) putting on client roadmaps:

• Anything that requires very-bleeding-edge model features in production paths. We give them 3–6 months to stabilise before relying on them.
• Heavy bets on early-stage Android features that don't yet ship on a majority of devices our clients' users own.
• Replacing well-tuned retrieval with massive context windows by default — long-context is a tool, not a strategy. We still pick per-problem.

TL;DR

The platform keeps moving in the direction we bet on when we started TechKis: AI-augmented engineering becomes the default, the web stays the most cost-effective distribution surface, and senior craft compounds with every model generation. Google I/O 2026 didn't shift our roadmap — it confirmed it.

If you're a founder or product team thinking about where to spend AI engineering budget in the next 12 months, we'd love to talk.