TechLook
Google I/O didn't just ship models — it dropped a pricing grenade into the AI market. Gemini 3.5 Flash is now sitting at near-frontier performance, available across every Google product, and priced at a level that makes the previous cost baseline look absurd. Meanwhile, the talent market is moving too: Andrej Karpathy — one of the original architects of modern AI — just picked Anthropic over the company he helped found. And GitHub Copilot just became something meaningfully different.
The common thread today: the agentic era isn't coming. It's here. And it's getting cheap fast.
Gemini 3.5 Flash Goes Agent-First — At Half The Cost
Google dropped its most consequential model release in years at I/O, and the headline isn't a benchmark beat. It's the price-performance ratio.
Gemini 3.5 Flash delivers near frontier-level performance — comparable to Opus 4.7 and GPT-5.5 — while running 4x faster than comparable frontier models and at roughly half the cost. The model is already live inside the Gemini app, AI Mode in Search, Google Antigravity, AI Studio, Android Studio, and Google's enterprise tools.
Here's everything you need to know:
- Near frontier performance on standard benchmarks, matching Opus 4.7 and GPT-5.5 levels
- 4x faster inference than other frontier models at comparable accuracy
- Available now across Gemini app, AI Mode in Search, Google Antigravity, AI Studio, Android Studio, and enterprise products
- Beats Gemini 3.1 Pro on agent and coding benchmarks, scoring 76.2% on Terminal-Bench 2.1
- Gemini Spark rolling out to trusted testers — a 24/7 personal agent running on Google Cloud VMs, working across Workspace, Chrome, email, and chat
- Search getting its biggest redesign in a generation: cross-modal inputs, 24/7 info agents, generative UI for custom layouts
- Gemini Omni converts text, images, audio, or video inputs into video outputs
The 3.5 Flash benchmarks don't crush the competition on paper. But combining fast, cheap, near-frontier capability with a distribution footprint that spans billions of devices is a different kind of threat. The agentic stack — where speed and cost determine whether a tool actually works in production — is where this model matters most.
If your product's cost per task just got cut in half against the new baseline, that's a pricing problem that compounds fast.
Andrej Karpathy Walks Into Anthropic — And It Says Everything
One of the founding researchers of modern AI just chose his side. Andrej Karpathy — OpenAI's co-founder, Tesla Autopilot's former leader, and the researcher most associated with teaching the world about neural networks — has joined Anthropic's pre-training team under research lead Nick Joseph.
Karpathy will build an internal group focused on using Claude to automate Anthropic's own AI training pipelines. That's a self-referential loop that only a company at the frontier can attempt: using its best model to make its next model better.
Here's everything you need to know:
- Karpathy co-founded OpenAI in 2015, led Tesla Autopilot until 2022, briefly returned to OpenAI, then left in 2024 to build an AI education startup
- Joining Anthropic's pre-training team under Nick Joseph
- Will lead a new internal effort to apply Claude to Anthropic's own training pipeline automation
- Anthropic currently leads enterprise adoption at 34.4% vs OpenAI's 32.3%
- Karpathy's move mirrors a pattern: researchers who built OpenAI's early culture have been leaving for Anthropic for years, including the group that walked out in 2021 over a strategic rift
Karpathy didn't just join a competitor. He joined a company founded by researchers who left OpenAI because they disagreed with where it was going. His choice is a signal, not just a hire. The talent is going where the research culture feels right — and right now, that's Anthropic.
For founders watching the AI platform race: enterprise adoption is neck-and-neck, but the direction of talent movement matters more than a point-in-time survey number.
GitHub Copilot Is Now A Full Coding Agent
Microsoft used Build 2026 to quietly redefine what GitHub Copilot is. It expanded from autocomplete-with-guardrails into a coding agent that handles larger tasks end-to-end, works across multiple repositories simultaneously, and returns PR-ready changes without human intervention.
Here's everything you need to know:
- GitHub Copilot now handles full coding tasks — not just completions
- Works across multiple repositories in a single session
- Returns PR-ready changes, not just suggestions
- Context: Microsoft Build also positioned "open agentic web" as the center of its developer platform strategy — interoperable AI agents as a first-class platform primitive
The terminal is getting crowded. Grok Build entered the coding agent race last week (covered May 19). Cursor dropped its cost-per-task below $1. And now GitHub Copilot is going agent-native. The question for builders isn't whether to use a coding agent — it's which one gives you the best margin, not just the best accuracy.
Dell Sounds The Alarm On Agentic Cloud Costs
While the model announcements were flying, Dell's COO Jeff Clarke delivered a quieter warning that should concern every company building AI products on cloud infrastructure: agentic AI workloads are making cloud-only strategies economically unsustainable.
Here's everything you need to know:
- Dell COO Jeff Clarke says reasoning token usage is up 320x — a proxy for the compute demands agentic workloads generate
- Dell is urging enterprises toward on-prem "deskside to data center" AI infrastructure
- The argument: agents should run where company data already lives, not where cloud bills explode
- Y Combinator GP Tom Blom independently published a framework urging founders to stop treating AI as copilots and instead rebuild companies as recursive, self-improving loops — "record knowledge → AI-legible context → burn tokens instead of headcount"
The two messages align: the economics of cloud-native AI are under pressure from the bottom up. Agents that run thousands of tokens per task, across long sessions, against proprietary data, don't behave like API calls. They behave like infrastructure.
For builders: the cloud-first era was built on cheap, scalable compute. Agentic workloads are stress-testing that assumption at the layer where your cloud bill hits the P&L.
⚡ Quick Hits
- Google: Gemini 3.5 Flash and Spark are live — fast, cheap, near-frontier AI available across the entire Google product ecosystem.
- Google: Smart glasses are back. Gemini-powered frames from Warby Parker and Gentle Monster shipping this fall; Project Aura with displays to follow.
- OpenAI: Adding Google DeepMind's SynthID watermarking to images from ChatGPT, Codex, and the OpenAI API — becoming a C2PA Conforming Generator Product. A sign that AI media provenance is becoming a platform requirement.
- Microsoft: Build 2026 positioned "open agentic web" — interoperable AI agents — as the center of its developer platform. Not a product launch. A platform bet.
- Mistral AI: Acquired Emmi AI, pushing into industrial AI with simulation and engineering workflows. The small-model specialist is getting industrial.
- Anthropic: Launched MCP tunnels — credential-protected API connections purpose-built for enterprise agents. Safer agent-to-toolchain connectivity is now a product.
- Google: Partnering with Blackstone on a $5B joint AI cloud venture to rent out Google's TPU chips at scale.
- Neuralink: The Telepathy chip is in 21 patients. Blindsight — which feeds camera signals directly into the visual cortex — has secured FDA breakthrough designation with a first human implant targeted for year-end 2026.
Techlook — AI & tech signal for founders and builders.