Gemini 4 Neon is Google's new middle model for coding agents. It costs more per token than Argon does this month, and less from December 1

Google released Gemini 4 Neon on November 18 for the Gemini API, AI Studio, Antigravity, and the Gemini app. It costs $3 input and $15 output per million tokens and leads Artificial Analysis's Coding Agent Index. Price it against Argon's December 1 rates, not today's, before you move agent traffic.

Google's launch post places Neon between Carbon, which it pitches for the hardest long-horizon projects, and Hydrogen, which handles fast everyday execution. Google says Neon "balances depth and efficiency" and costs 32% less per task than Carbon at its default thinking level.

We haven't run Neon on our own tasks. Every benchmark below is Google's figure as published in the launch post, and the cost examples are our arithmetic on list prices.

Where you can use it

SurfaceWho gets itNotes
Gemini API (AI Studio, Android Studio)Paid tierRolling out from November 18; no free tier mentioned
Google AntigravityPaid Google AI Pro and Ultra subscribersCan act as an orchestrator over parallel Hydrogen subagents
Gemini appGoogle AI Pro and Ultra subscribersModel drop-down
Gemini for businessEnterprise customersThrough the API and model drop-down; Vertex AI isn't named
Fairwind ProgramVetted security defendersA cyber-tuned Neon rolls out "in the coming days"

The launch post doesn't mention Gemini CLI, Jules, GitHub Copilot, or Cursor. If you reach Gemini through one of those, wait for that tool's own changelog before you plan around Neon.

The API model ID is gemini-4-neon, per Google's Neon developer guide. It works with generateContent and the Interactions API on the paid tier only; there is no free tier.

What it costs

Per million tokens. Neon is from the launch post and the Neon developer guide, Argon from the pricing page, Hydrogen from our Hydrogen pricing guide, and the rest from our GPT-6.2 guide:

ModelInputCached inputOutput
Gemini 4 Neon$3.00$0.15$15.00
Gemini 4 Argon, through November 30$2.00$0.10$10.00
Gemini 4 Argon, from December 1$4.00$0.20$20.00
Gemini 4 Hydrogen$1.50$0.075$7.50
GPT-6.2 Sol$2.00$0.10$10.00
Claude Opus 5.5$4.00$0.20$20.00
GPT-6.2 Astra$8.00$0.40$40.00

Our arithmetic for one uncached agent step with 200,000 input tokens and 20,000 output tokens:

  • Gemini 4 Neon: $0.60 + $0.30 = $0.90
  • Gemini 4 Argon: $0.60 today, $1.20 from December 1
  • Gemini 4 Hydrogen: $0.30 + $0.15 = $0.45
  • GPT-6.2 Sol: $0.60, and Claude Opus 5.5: $1.20

So for the next ten days Neon costs 50% more per token than Argon. After December 1 it costs 25% less. If you are on Argon today, the switch only pays off on price once the new rates land.

Cached input at $0.15 is a 95% discount. Agent loops resend the same repo context every turn, so a step whose 200,000 input tokens all hit the cache costs $0.03 for input plus $0.30 for output. Your real bill depends on your cache hit rate far more than on the headline rate.

The Neon developer guide settles two billing questions. The $15.00 output rate includes thinking tokens, so a higher thinking level shows up directly on the output line. And pricing is flat across the full 1M-token window, with no long-context surcharge like the 272K cliff on GPT-6.2.

Caching is implicit only. It applies to prompts of at least 4,096 tokens, and explicit caching calls such as caches.create and cached_content aren't supported. If your agent manages cache objects by hand today, that code has to go.

What the coding results say

All of these are from Google's launch post. Google doesn't say which harness it used for most of them, and its full comparison table is an image.

BenchmarkNeon resultWhat to keep in mind
Artificial Analysis Coding Agent Index58.6, ranked #1An independent index, as quoted by Google
Terminal-Bench 4.165.2%Harness not stated
FrontierCode 2.046.2%At the default high thinking level
SWE-sweep12.8% in mini-SWE, 39.0% in AntigravitySame model, three times the score with Google's own harness
Terminal-Bench Science70.5%Google calls it best in class
Coding cheating eval2.45% of runs took shortcutsLow, but not zero; review diffs that touch tests

The SWE-sweep gap is the most useful number on the page. Going from 12.8% to 39.0% by changing only the harness shows how much of a coding agent's result comes from the tool around the model. Don't expect Antigravity-level numbers if you drive Neon from your own loop or a third-party CLI.

The Artificial Analysis rank is the strongest outside signal. On its broader Intelligence Index, Google says Neon scores 61 at the top thinking level and ranks third, "just behind larger, slower models." For context, The Decoder reported GPT-6.2 Astra at 62 on that index earlier this week.

Catches before you switch

  • Thinking level drives cost. Set thinking_level in generation_config to low, medium, high (the default), or xhigh. Neon rejects minimal with an error, so agents that use it for cheap routing calls will break. Carbon and Argon default to medium, so moving to Neon raises the default spend per call.
  • Output caps. Input, thinking, and output share one 1,048,576-token window. A synchronous request stops at 65,536 output tokens; for more, set background to true and raise max_output_tokens, which counts thinking too. Google also says to set HTTP client timeouts to at least 15 minutes.
  • Sampling parameters are gone. The guide says to remove temperature, top_p, and top_k. In-range values are ignored, and out-of-range values return a 400 error.
  • Declare tools natively. Prompting for raw-text JSON tool calls with no declared tools can return MALFORMED_FUNCTION_CALL. Use function calling or structured outputs.
  • Still unpublished. Google hasn't stated a knowledge cutoff or a preview-or-stable label, and rate limits sit on a separate page we didn't check. Don't put Neon behind a production SLA until it does.
  • Price drift. Google is doubling Argon's prices on December 1, after a launch-price period. Neon has no announced change, but plan for the possibility on any newly launched Gemini model.
  • Security work is gated. The general Neon posts strong security-benchmark numbers, but the cyber-tuned version goes only to vetted defenders through the Fairwind Program.

What we'd do

Argon users: run Neon beside Argon this week on a batch of closed tickets at the high thinking level. If Neon matches Argon on your tasks, switch on December 1, when it becomes the cheaper of the two.

Antigravity users: try Neon as the orchestrator with Hydrogen subagents. That is the setup Google tuned for, and the SWE-sweep numbers suggest the harness matters as much as the model.

Everyone else: at $0.90 for our example step, Neon sits between GPT-6.2 Sol at $0.60 and Opus 5.5 at $1.20. It is worth a trial if you want a second vendor for fallback, but the numbers above are Google's, so measure on your own repo before you move a default.

Sources

Researched and drafted by an AI agent from the linked sources. How posts are made.