Dev.to AI 🤖 Ai 👁 0 📖 3 min read

Gemini 4 Neon brings complex agent workflows to Copilot and Google's AI tools

Gemini 4 Neon Is Google's "Cheaper Than Carbon, Smarter Than Hydrogen" Bet — and the Benchmarks Deserve a Second Look Google has a new middle child. Gemini 4 Neon sits between Carbon — the flagship for the hardest long

Gemini 4 Neon Is Google's "Cheaper Than Carbon, Smarter Than Hydrogen" Bet — and the Benchmarks Deserve a Second Look

Google has a new middle child. Gemini 4 Neon sits between Carbon — the flagship for the hardest long-horizon work — and Hydrogen, the fast everyday model. The pitch: intelligence approaching Carbon, at a fraction of the time and token cost, built specifically for long-running agents and multistep tool use. It's now generally available in GitHub Copilot, which is where most developers will actually meet it.

That sounds great. But the launch numbers need reading with the fine print intact — because Google's own announcement quietly shows how much the harness matters, and how little a headline score transfers to your setup.

The benchmarks are real — and slippery

Neon scores 61 on the Artificial Analysis Intelligence Index (third overall) — but that's at the xhigh thinking level. Its FrontierCode 2.0 result of 46.2% is at the default high setting. Those are not measurements of one fixed configuration, and Google doesn't pretend otherwise. Anyone quoting a single Neon score without naming the thinking level is selling you a number, not a model.

The most honest figure in the whole announcement might be the SWE-sweep spread: 12.8% with the mini-SWE harness versus 39.0% inside Google Antigravity. Same model, nearly 3x difference — because the agent system around it does the heavy lifting. That 39% belongs to Antigravity's harness, not to Neon in your Copilot session.

Steerability is the stat that actually matters for agents

Neon ranks second on the new Agent Arena Index for steerability — adapting to user corrections mid-task — across more than 4,200 real-world agent sessions. For an agent handling a changing assignment, completion alone is worthless if the model keeps pursuing instructions you've already revised. This is the capability the agent era lives or dies on, and it's the one benchmark worth watching. The caveat: Google hasn't published the detailed methodology, so treat the ranking as promising, not proven.

The cost claims are vendor math, not your bill

Google claims 32% lower cost per task than Carbon at high, and 50–80% lower than top competing frontier models on some indexes. Token prices: $3/million input, $0.15/million cached input, $15/million output. But cost per task depends on how much work a task consumes and whether the output is usable — and the announcement doesn't give enough underlying detail to turn those percentages into a forecast for your team's bill. Don't budget on a vendor's percentage.

What you'll actually get

  • A 1M-token context window that synthesizes spreadsheets, PDFs, charts, audio, and video into structured outputs — with chart-reading scores (73.5% on Chartography) that cover specific evaluations, not every document you'll upload.
  • In Antigravity, Neon can orchestrate parallel Hydrogen subagents — a documented workflow, not a promise that every product exposing Neon offers the same controls.
  • In Copilot, it's rolling out gradually to Pro+, Max, Business, and Enterprise — and under default model enablement, admins who haven't disabled it will see it appear automatically.

The take

Neon is a genuinely interesting model — the steerability focus and the agent-first positioning are real. But Google's own data makes the case against taking any single benchmark at face value: the harness, the thinking level, and the cost basis all move the numbers. Judge it on how it behaves in your tooling, with your corrections, on your tasks. Everything else is launch marketing.

Go deeper

The full analysis — with the benchmark tables, pricing breakdown, and rollout details — is on Agentic Affairs:

👉 Gemini 4 Neon brings complex agent workflows to Copilot and Google's AI tools

Originally published on Agentic Affairs (https://agentic-affairs.com)

📰 Read the original article on Dev.to AI

Originally published by Dev.to AI. Aggregated on AIWithGhost for educational purposes — full credit and traffic to the original publisher.