Back to Blog

Gemini 4 Neon Coding Benchmark: 99.4% on the First Try, Our Best Score Yet, at 3.8× Claude Opus 5.5's Cost

2026-11-18·7 min read·CodeRouter Team

Disclosure: we build CodeRouter, a router for coding agents. We called both Gemini models through OpenRouter, which sent every Neon request and most Carbon requests to Google's Vertex endpoint. The other models' numbers are the ones already on our benchmark page, measured on November 3, and in our Qwen4-Max and GPT-6.2 benchmarks from November 16 and 17.

Google released Gemini 4 Neon on November 18. It costs $3 per million input tokens and $15 per million output, the same list price as Claude Sonnet 5.5, with cached input at $0.15 and a 1,048,576-token context. Google places it between Gemini 4 Hydrogen and Gemini 4 Carbon and says it matches or beats Carbon's quality at about three times the speed. We ran both Neon and Carbon on the same tasks as every model on our benchmark page. Neon is available by name in CodeRouter's Direct mode since November 18, Carbon since November 10; setup is below the results.

Gemini 4 Neon solved 99.4% of the tasks on the first try and 100% with one retry, the best first-try score we've measured. The run cost $9.87, 3.8× what Claude Opus 5.5 cost for 98.3% and 100%, and took a median of 19.4 seconds per task against Opus's 4.8.

The results

Same 175 single-file Python tasks (Exercism plus Aider's hard set). The model sees the instructions and the stub file and returns the whole file; we run the tests, and on a failure it gets one retry with the failing output.

Model First try With one retry Cost of all 175 tasks Median time per task
Gemini 4 Neon 99.4% 100% $9.87 19.4 s
Claude Opus 5.5 98.3% 100% $2.61 4.8 s
Claude Sonnet 5.5 93.7% 100% $1.48 3.9 s
Gemini 4 Carbon 97.1% 99.4% $3.57 ($7.14 from Dec 1) 20.2 s
GPT-6.2 Astra 92.6% 97.1% $3.17 3.3 s
GPT-6.2 Sol 92.0% 97.1% $0.88 2.7 s
Qwen4-Max 83.4% 96.0% $1.47 11.3 s
Kimi K2.6 86.9% 94.9% $0.57 9.4 s
Claude Haiku 5.5 72.0% 93.7% $0.15 5.5 s

Costs are list prices applied to the tokens each request reported, retries included; for both Gemini models that equals what OpenRouter charged, request by request. Carbon's $3.57 is at Google's introductory price of $2 and $10, which ends on November 30. From December 1 Carbon costs $4 and $20, and the same run would cost $7.14.

Neon's only first-try miss was tree-building from Exercism, which Opus 5.5 and Sonnet 5.5 also missed on the first try; all three solved it on the retry. Carbon failed that task on both attempts with a different error message from the one the test expects. Both Gemini models solved all five tasks that both GPT-6.2 models failed. No request returned an error or hit our 32,768-token output limit; the longest reply was 25,683 tokens, Carbon's.

Why Neon costs 3.8× Opus: 91.5% of its output is thinking

Neon has Sonnet 5.5's list price, and its run still cost more than any model in the table. It wrote 631,481 output tokens across the 175 tasks, 6.7× Opus 5.5's 94,663, and 91.5% of them were thinking. The median first answer was 2,112 output tokens, 1,854 of them thinking. Google bills thinking tokens as output.

Carbon thought less: 329,164 output tokens, 83.5% of them thinking. At its introductory price that cost $3.57; at December's price it would cost $7.14.

Both models ran at Google's default thinking level, which is high for Neon. We didn't set a lower one, and a lower level may change the cost, the time and the score. We haven't measured it.

"About three times the speed": per token yes, per task no

Google says Neon matches or beats Carbon's quality at about three times the speed. On these tasks the per-token half of that holds:

Gemini 4 Neon Gemini 4 Carbon
Time to first token (median) 1.9 s 5.8 s
Output speed (median) 113 tokens/s 54 tokens/s
Output tokens per first answer (median) 2,112 1,076
Time per task (median) 19.4 s 20.2 s
Time for all 175 tasks 5,505 s 5,535 s

Neon starts about three times sooner and writes about twice as fast, but it writes about twice as much. So a finished task took the same time: about 20 seconds at the median, with Neon faster on 98 of the 175 first attempts. The quality half holds too, by a few tasks: 174 against 170 on the first try, 175 against 174 with the retry.

Try Gemini 4 Neon in your coding agent

CodeRouter serves it by name: gemini-4-neon. Gemini 4 Carbon is gemini-4-carbon. Both are billed at Google's list price with no markup, plus a 5.5% fee when you buy credits; exact prices are on the models page. Smart routing doesn't use either: you pick the model.

  1. Sign in, add credit, and create an API key. The free trial covers the models smart routing uses, so it doesn't cover Gemini 4 Neon or Carbon.
  2. In the dashboard, switch the account to Direct mode (Team keys always use smart routing). In smart routing, the default, a model name in the request is treated as auto. In Direct mode the model you name answers, with no routing and no fallback. More: Direct mode.
  3. Point your tool at CodeRouter.

Claude Code. Put the key in ANTHROPIC_AUTH_TOKEN, not ANTHROPIC_API_KEY, and set every model variable, or some calls will still ask for a Claude model:

export ANTHROPIC_BASE_URL="https://www.coderouter.io"
export ANTHROPIC_AUTH_TOKEN="cr_YOUR_KEY_HERE"
export ANTHROPIC_MODEL="gemini-4-neon"
export ANTHROPIC_DEFAULT_OPUS_MODEL="gemini-4-neon"
export ANTHROPIC_DEFAULT_SONNET_MODEL="gemini-4-neon"
export ANTHROPIC_DEFAULT_HAIKU_MODEL="gemini-4-neon"
export CLAUDE_CODE_SUBAGENT_MODEL="gemini-4-neon"
claude

Cursor, Cline or any OpenAI-compatible tool. Base URL https://www.coderouter.io/api/v1, your CodeRouter key, model gemini-4-neon (or gemini-4-carbon). With the OpenAI SDK:

from openai import OpenAI

client = OpenAI(
    base_url="https://www.coderouter.io/api/v1",
    api_key="cr_YOUR_KEY_HERE",
)
resp = client.chat.completions.create(
    model="gemini-4-neon",
    messages=[{"role": "user", "content": "Refactor this function"}],
)
print(resp.choices[0].message.content)

CodeRouter doesn't pass a thinking level to Gemini 4 models yet, so Neon thinks at Google's default, high. Hard prompts can think for minutes before the first answer token: CodeRouter keeps such a stream open for up to 300 seconds of silence. Details are in the integration notes. We haven't run a Claude Code session on Neon ourselves yet.

What it means for CodeRouter's routing

Neither Gemini model is in smart routing. On this suite Neon is the most accurate model we've measured and the most expensive per task, and Opus 5.5 reached the same 100% with one retry for about a quarter of the cost and time. Single-file tasks aren't long agent sessions, though, and whether Neon earns a place, for example on the hardest steps, needs agent-session tests.

Limits

Per-task data: coderouter-gemini4-175-2026-11.json. The other models' data is on the benchmark page and in the GPT-6.2 benchmark.

To try Gemini 4 Neon on your own code, sign in and name gemini-4-neon in Direct mode.

Try phase-aware routing on your own workload

One base URL for Claude Code, Cursor or Aider. Free trial, no credit card.

Start free trial

Get weekly AI cost optimization tips

Join developers cutting their agent bills