SDSignal Desk

Gemini 4 Neon is Google’s bid to make the middle tier the default

· Updated · Signal Desk newsroom

Analysis of reporting by Google DeepMind. How we work

Image: Google DeepMind

DeepMind’s new model sits between Carbon and Hydrogen and claims near-frontier scores at much lower cost per task. The pricing and safety disclosures are strong; nearly every benchmark claim is still Google’s own reporting.

Key takeaways

  • Neon costs $3 per million input tokens, $0.15 cached and $15 per million output, with a 1 million-token context window.
  • Google says it ranks #3 on the Artificial Analysis Intelligence Index at xhigh and #1 on its Coding Agent Index.
  • At default high thinking, Google claims 32% lower cost per task than Carbon and roughly 50% to 80% lower than top competitors.
  • A cyber-focused version goes only to vetted defenders in the Fairwind Program.
  • Google reports a 0.09% misaligned-action rate and its lowest fabrication rate yet in internal testing.

Where Neon fits

Gemini 4 now has a lineup, and Neon is the piece meant to do most of the work. Google positions Carbon for the hardest long-horizon projects, Hydrogen for fast everyday execution, and Neon as the balance: frontier-level reasoning for complex, high-frequency workflows at lower cost. It rolls out in the Gemini API, Gemini for business, Antigravity for paid Pro and Ultra subscribers, and the Gemini app.

This is the right product shape. Most enterprise AI spending is not on the single hardest problem but on thousands of medium-hard ones: spreadsheet modeling, document synthesis, multi-file code changes. A model that does those well, cheaply, matters more to budgets than another leaderboard-topping flagship. The lineage traces back to the Gemini 4 launch, when Google first restricted its most capable cyber features to trusted defenders.

The numbers, with caveats

Google’s scoreboard is long. Neon is #3 overall on the Artificial Analysis Intelligence Index with a 61 at xhigh, #2 on the Vals Index at 70.9%, #1 on Artificial Analysis’s Coding Agent Index at 58.6, and #1 on Zapier’s AutomationBench and Andon Labs’ Vending-Bench 2. On SWE-sweep it scores 12.8% in a minimal harness and 39.0% inside Antigravity, a gap that says as much about the harness as the model.

Some of these are third-party leaderboards, which is better than purely internal evals. But the selection is Google’s, and many named benchmarks are new enough that we cannot judge how well they predict real work. The cost claims are the more useful signal: 32% lower cost per task than Carbon at the default setting and about half to a fifth of competing frontier models. Those are testable, and developers will test them within days.

Safety reporting worth copying

The most encouraging part of the release is the safety section, because it reports specific failure rates rather than adjectives. Google says Neon refused just 1.98% of Terminal-Bench 4.0 tasks while keeping Carbon-level misuse safeguards, cheated on 0.0% of Terminal-Bench 4.0, took shortcuts on 2.45% of its coding-cheating evaluation, and had a 0.09% misaligned-action rate on its agentic safety suite. It also claims a 0.0% attack success rate on Gray Swan’s indirect prompt injection benchmark at its highest thinking level with real-time safeguards.

These are self-reported, and internal evaluations reflect what the developer chose to measure. Still, publishing low single-digit and fractional rates for shortcuts and misalignment sets a bar rivals should meet. The cyber variant follows the gated pattern Google set with Argon: offensive-adjacent capability goes to vetted defenders first through Fairwind. That is a defensible approach for a model Google says is state-of-the-art on exploit and vulnerability benchmarks.

Who it affects

Developers
Competitive pricing with cached-input discounts makes Neon a candidate default for agent and coding workloads.
Enterprises
Strong spreadsheet, legal and tax benchmark claims target the routine knowledge work that drives volume.
Security teams
A cheaper, capable cyber model reaches defenders first; attackers will look for equivalent open tools.
Rival labs
Mid-tier pricing pressure is now as important as flagship benchmarks.

What to watch

  1. Independent reproduction of the cost-per-task claims on real workloads.
  2. Whether Neon’s fabrication and shortcut rates hold up outside Google’s own tests.
  3. Fairwind access expansion and any misuse incidents tied to cyber capabilities.
  4. Price responses from competing mid-tier models.

FAQ

How much does Gemini 4 Neon cost?

$3 per million input tokens, $0.15 per million cached input tokens and $15 per million output tokens, with a 1 million-token context window.

How does Neon differ from Carbon and Hydrogen?

Google describes Carbon as the model for the hardest long-horizon work, Hydrogen for fast everyday tasks, and Neon as the balance of depth and efficiency.

Who can use the cyber version?

Eligible members of Google’s Fairwind Program for trusted defenders, starting over the next few days.

Are the benchmark results independent?

Some come from third-party leaderboards such as Artificial Analysis and Vals, but Google selected and reported them, and safety figures are internal.

Read the original

Continue at the source.

Google DeepMind

Companies: Google

More on Google

All Google coverage →