Three things happened this week that developers are processing differently depending on which tools they already use. Here's what you need to know about each one — with the actual numbers, not the marketing version.
1. Gemini 3.5 Flash is out. It is not cheap. It is very good.
Gemini 3.5 Flash launched May 19 at Google I/O 2026 and is now the default model in the Gemini app, AI Mode in Google Search, and the Gemini API. That is the headline Google wanted.
Here is the headline Google buried: 3.5 Flash costs $1.50 per million input tokens and $9.00 per million output tokens — 3x the previous Gemini 3 Flash at $0.50/$3, and 6x the Gemini 3.1 Flash-Lite tier.
So what you are getting for that price increase: it beats Gemini 3.1 Pro on coding, agentic, and multimodal benchmarks while running 4x faster in output tokens per second. That is a real and meaningful result. A Flash-tier model that outperforms last year's Pro on the tasks that matter for agents — tool use, multi-step reasoning, code generation — genuinely changes the cost-quality calculus.
The correct framing is: Google moved the frontier line down to the Flash tier. The old two-tier mental model — "Pro for hard problems, Flash for throughput" — has collapsed on the workloads that actually matter for agents.
What developers should actually do:
If you were on Gemini 3.1 Pro ($2.50/$15), this is a price cut. Switch now. If you were on Gemini 3 Flash for low-cost throughput, this triples your bill and you need to evaluate whether the capability jump justifies it. For low-complexity tasks — basic summarisation, translation, classification — staying on Gemini Flash-Lite or Haiku makes more financial sense. The unglamorous move is: run your actual workload on both, measure cost-per-task, decide from your own numbers.
One more thing: Gemini 3.5 Flash fully supports MCP — its MCP Atlas score is 83.6%. If you have built MCP tools for Claude, they should largely drop in without modification. That is a practical piece of good news for developers who have already invested in MCP infrastructure.
2. Google's Colab CLI is the most quietly important developer tool shipped this week
This one barely made the tech press. It deserves more attention.
On June 5, Google shipped the Colab CLI — an open-source, Apache 2.0 command-line interface that connects your local terminal to remote Colab runtimes. One command provisions a GPU: colab new --gpu A100. colab exec ships your local .py file to the runtime and runs it. No browser, no upload step, no notebook UI required.
That is useful on its own for developers who use Colab for ML experiments but hate the browser interface. But the part that matters more is this: any terminal-based AI agent — Claude Code, Codex, Antigravity — can drive it via a bundled COLAB_SKILL.md file. The agent can provision a runtime, run Python, retrieve results, and tear down the session, all without a human in the loop.
Think about what that means in practice. An AI agent working on a machine learning task can now provision its own GPU compute, run experiments, retrieve the outputs, and iterate — without you paying for a dedicated cloud GPU instance or managing infrastructure. The compute is pay-as-you-go Colab credits. The agent drives the whole thing.
This is not a cosmetic change. It turns Colab sessions into a remote compute service that AI agents can operate directly from the shell.
Install it:
uv tool install git+https://github.com/googlecolab/google-colab-cli
colab new --gpu T4 # provision a GPU runtime
colab exec script.py # run your code on itOfficial announcement on Google Developers Blog has the full command reference.
3. Gemini Spark is Google's "always-on agent" bet
Tucked inside the I/O announcements and now rolling out to early testers: Gemini Spark is a persistent personal agent built on top of Gemini 3.5 Flash, initially rolling out to AI Ultra subscribers ($100/month) in the US. It handles recurring tasks in the background — scheduling, monitoring, summarising, acting on your behalf without you explicitly triggering each task.
This is the "always-on agent" pattern that every major AI lab has been talking about for two years and nobody has shipped in a way that actually stuck. Spark is Google's serious attempt.
The developer angle: Spark runs on the same Antigravity agent stack and MCP infrastructure that 3.5 Flash uses. If you are building developer tooling that plugs into the Google ecosystem, Spark is the surface area where your MCP server could end up running as a background task on someone's machine 24/7.
The skeptic's angle: "always-on" agents have a trust problem that has nothing to do with technical capability. Developers have been burned by background processes that took unexpected actions. The adoption curve for Spark will depend entirely on whether Google can make it feel safe and predictable rather than capable and unpredictable.
Worth watching. Not yet worth building around.
One thing to do before next Friday
If you have any production workloads running on Gemini 3 Flash or Gemini 3.1 Flash-Lite, check your billing settings this week. The new 3.5 Flash is the default model in the API going forward. If you are not pinning model versions explicitly in your API calls, you may already be on the more expensive tier.
Pin your model version. Always pin your model version. gemini-3.5-flash will eventually get updated again, and gemini-latest is a bill waiting to surprise you.
Missed something big this week? Drop it in the comments — next Friday's edition is better when readers are part of the research.
Comments (0)
Login to post a comment.