Skip to content
Developer312

Gemini 4 Neon: A #3-Overall, #1-Coding Model at $3/$15

Google's Gemini 4 Neon lands at $3/$15 per million tokens with a 1M-token context, ranking #3 overall and #1 on the Artificial Analysis Coding Agent Index at 58.6. That resets the price floor for agent loops.

Published Report an error

Google released Gemini 4 Neon, a mid-tier model priced at $3 per million input tokens and $15 per million output, with a 1M-token context window and four configurable thinking levels. Per Artificial Analysis, it scores 61, #3 overall on the Intelligence Index, and ranks #1 on the Coding Agent Index at 58.6. Google's own announcement adds #2 placements on the Vals Index and Agent Arena steerability, and claims roughly 50-80% lower cost per task than top frontier models.

The pricing lands in a week when frontier models cost $8/$40, both GPT-6.2 Astra and Claude Fable 5.2 sit on that exact number. A model ranked #3 overall and #1 at coding, selling at less than half the frontier input rate, changes the default answer to "which model should power my agent loop?"

Key Takeaways: What Matters Most?

  • Gemini 4 Neon is priced at $3 input / $15 output per million tokens with a 1M-token context window.
  • Per Artificial Analysis, Neon scores 61 for #3 overall on the Intelligence Index and #1 on the Coding Agent Index at 58.6.
  • Neon also ranks #2 on the Vals Index and #2 on Agent Arena steerability, per Google's announcement.
  • Google claims roughly 50-80% lower cost per task than top frontier models across four configurable thinking levels.
  • At Astra's $8/$40 and Fable 5.2's $8/$40, Neon undercuts the frontier tier by more than 60% on input price while holding the #1 coding spot.

Takeaways like these, every weekday morning. Get the briefing

The score-per-dollar argument

Index placement is only meaningful against price, so line them up. Neon is #3 overall at 61 on the Artificial Analysis Intelligence Index, one slot from the top, at $3/$15. The frontier tier, Astra and Fable 5.2 at $8/$40, is what Neon's 50-80% per-task cost claim is measured against. The Coding Agent Index is where the story sharpens: Neon holds #1 at 58.6, meaning the cheapest model in the top tier of overall intelligence is also the best-scoring coding agent on that particular index.

The honest caveat is that index rankings flatten a lot of detail. #2 on steerability means frontier models still beat Neon on that dimension, and #2 on the Vals Index leaves a slot above it. For builders, the operative question is which 10% of your workload actually lives on the frontier tier and which 90% does not. Coding agents are largely repetitive loops: read context, edit code, run tests, read output, repeat. That loop is exactly where a cheaper, still-#1-ranked model converts directly into margin.

The four configurable thinking levels are the cost-control mechanism that makes the price real in production. Rather than paying frontier rates for every step of a pipeline, you can spend deep thinking only where the task needs it, a knob frontier models have been slower to expose. That single control turns the pricing table into a dial you can tune per stage of your application instead of a flat rate you are stuck with across every request type.

What it means for the frontier tier

Neon's existence squeezes the frontier market from below while OpenAI and Anthropic compete at the top. Astra's 20% price cut to $8/$40 this week now looks less like aggression and more like defense: the frontier tier must justify a 2.7x input premium over a model that outscores it on at least one major coding index.

There is a plausible reading where this is the shape of the whole market from here: a frontier tier priced for genuinely hard, novel or high-stakes reasoning, and a mid-tier like Neon that wins the volume. Google's claim of 50-80% lower cost per task should be read as a vendor figure, the honest number for you depends on your context lengths, cache behavior and retry rates, but the direction is not in doubt.

The 1M-token context matters for the agent-loop use case specifically. Long-horizon agents accumulate context with every step, and a model that accepts 1M tokens at $3 per million input lets you keep the full session history resident instead of engineering aggressive truncation. That removes a whole class of context-management bugs, provided latency and output quality hold at scale, which is exactly what your own evals should check before a migration. It also means the per-task cost comparison against frontier models changes with session length: the longer the agent runs, the more the input-rate gap compounds in Neon's favor.

What To Do About It

  • Run your coding agent suite against Neon this week. If your current model is Astra or Fable 5.2 at $8/$40, a 20-task A/B on real tickets tells you whether Neon's 58.6 Coding Agent Index rank holds on your codebase.
  • Split your pipeline by difficulty, not by habit. Keep the frontier tier for architecture decisions and gnarly debugging; route the read-edit-test loop to Neon at $3/$15 and measure the bill delta after one week.
  • Tune the thinking-level knob per stage. Map each pipeline stage to one of Neon's four levels and record latency plus quality; the 50-80% claim is only real for you at the settings you measured.
  • Set your model-switch review at 30 days. Neon is #1 on one coding index today; re-check rankings and pricing monthly, because the $3/$15 floor will not stay uncontested.

The Bottom Line

Gemini 4 Neon is the strongest argument yet that the mid-tier is where AI value lives: #3 overall, #1 on the Coding Agent Index at 58.6, a 1M-token context, four thinking levels, and $3/$15 pricing that undercuts the $8/$40 frontier tier by more than 60%. For most agent loops, the default just changed, a top-ranked coding agent at mid-tier rates beats a frontier model on score per dollar. Keep a frontier model on retainer for the hard 10% of tasks, verify Google's 50-80% cost claim against your own traces, and expect this price floor to be the new baseline that every vendor launch is measured against.

Sources

  1. [1]Google Blog — Introducing Gemini 4 Neon (opens in a new tab)
  2. [2]AI Weekly — Gemini 4 Neon coverage (aggregator, verified live) (opens in a new tab)

Frequently Asked Questions

Get the next briefing

The lead story each weekday morning, with the key takeaways and a link to the full analysis.

One lead story each weekday morning: what happened, the key takeaways, and a link to the full analysis.

New posts daily, briefing weekday mornings. We use your email only to send the briefing. Unsubscribe anytime. Privacy policy.

Related Articles