Gemini 4 Neon
TL;DR Highlight
Gemini 4 Neon cuts cost per task 32% versus Carbon, at $3 input and $15 output per million tokens
Who Should Read
A backend developer choosing a Gemini model for coding agents or workflow automation based on cost and performance
Core Mechanics
- Pricing is $3 in, $15 out
- Input is $3 per million tokens and output is $15 per million tokens.
- Cached input costs $0.15 per million tokens, a 95% discount.
- The context window is 1 million tokens.
- Four thinking levels to tune cost
- The levels let teams dial intelligence and token spend per workload.
- At the default high level, cost per task is 32% lower than Carbon while keeping a high Artificial Analysis Intelligence Index score.
- At xhigh it scores 61, ranking #3 overall, at 36%+ lower cost per task than competitors at their highest settings.
- Strong coding agent results
- It ranks #1 on the Artificial Analysis Coding Agent Index at 58.6.
- It scores 65.2% on Terminal-Bench 4.1 and 46.2% on FrontierCode 2.0 at the default high setting.
- On SWE-sweep it gets 12.8% with the mini-SWE harness and 39.0% in Google Antigravity, more than triple.
- Top-tier knowledge work benchmarks
- It ranks #2 on the Vals Index at 70.9% and #1 on Legal Research Bench and Tax Agent Bench.
- It ranks #1 on Zapier AutomationBench at 56.3% and #2 on Arena's Agent Arena Index at 1336.
- It handles spreadsheets, PDFs, charts, audio, and video in a single context.
- Safety and alignment numbers published
- Refusal rate is 1.98% on Terminal-Bench 4.0 and 0% on Terminal-Bench-Science.
- It took shortcuts on 2.45% of tasks in the coding cheating evaluation.
- On Gray Swan's indirect prompt injection benchmark, it had a 0.0% attack success rate at its highest thinking level with real-time safeguards.
- Its place in the Gemini 4 family
- Carbon targets the hardest long-horizon projects.
- Hydrogen handles fast everyday execution.
- Neon balances depth and efficiency between them.
Community Reactions
- Big speed gap on OpenRouter
- One commenter saw Neon at 127 tokens per second versus 46 for Carbon on OpenRouter.
- They said Carbon is too slow and outputs too much per task.
- Cost comparison on a shared benchmark
- One commenter ran a 43-question data analytics exam and saw costs of Hydrogen $6, Carbon $11, Argon $13, Neon $17, with every Gemini 4 model scoring an A.
- They noted their harness uses heavy prompting, so default behavior is unclear.
- More output tokens, higher cost
- One commenter said Carbon, Hydrogen, and Neon used 3 to 4 times more tokens than Argon when generating demos.
- They argued bigger models can cost less overall and respond faster despite higher unit prices.
- Perceived quality is split
- One commenter felt it falls short of Opus 5.5 despite the benchmarks.
- Another found it efficient at driving tasks to completion and strong at writing.
- Naming scheme confusion
- Commenters said it's hard to map Neon, Carbon, Argon and similar names to use cases.
- One commenter corrected their own assumption that Neon was the larger model.
How to Apply
- Compare cost on 10 of your real prompts
- Pick 10 representative prompts from today's traffic and run them on Neon at high and on your current model.
- Record quality, output tokens, and latency in a table.
- Split workloads by thinking level
- Keep ordinary requests on the default high level.
- Use xhigh only for the hardest analysis and debugging tasks and measure cost per task separately.
- Apply cached input to repeated long prompts
- Find calls with large repeated input such as system prompts or documents.
- Recalculate monthly cost with the cached input price of $0.15 per million tokens.
- Test against Carbon in your agent pipeline
- Run the same tasks on Carbon for long-horizon work and Neon for complex high-frequency work.
- Set routing rules based on completion rate and output tokens.
Try It
# Paste into Claude or ChatGPT — model selection experiment design prompt (simple version based on the article's figures)
Our service is {{service description}}. We currently use {{current model}}, with {{monthly calls}} calls per month, averaging {{input tokens}} input tokens and {{output tokens}} output tokens per call.
Gemini 4 Neon pricing is $3 per million input tokens, $0.15 cached input, and $15 output.
Create the following:
1. A monthly cost comparison formula between the current model and Neon
2. Monthly cost when cached input is applied
3. A table template comparing quality, output tokens, and latency across 10 representative prompts
Next step: Paste in the results and help me decide whether high or xhigh is better value.Terminology
Related Papers
AI is killing All About Berlin
After Google AI Overviews, a Berlin guide site lost 70-75% of its traffic.
Not Everything Needs an LLM
PostHog used small decision models to skip 98.7% of expensive LLM calls
GPT‑6.2 Astra and Sol
GPT-6.2 Astra and Sol cut costs sharply, but over-engineering and early-stop reports are mixed
Show HN: TrnRT – An open-source LLM inference engine for AWS Trainium
One trn2.48xlarge serves GLM-5.3-Flash through an OpenAI-compatible API using NKI megakernels.
Know When to Zoom: How Small Is Too Small for Vision-Language Models?
Splitting large photos into tiles raised Claude and Gemini accuracy by 11.1–14.5 points
Understanding Positional Embeddings in Modern LLMs
A walkthrough of how LLMs encode token position, from GPT-2 to RoPE
Show HN: DeepSeek V3 in ~200 lines of pure Jax
DeepSeek V3's core design in 190 lines of JAX, trained on a MacBook Air in 2.6 hours
Nvidia allegedly engineered scarcity by funding AI startups to buy its GPUs
WSJ alleges NVIDIA funded startups that then bought its own GPUs to prop up demand.
Related Resources
- https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-neon/
- https://openrouter.ai/google/gemini-4-neon
- https://openrouter.ai/google/gemini-4-carbon
- https://micjm.com/blog/carbon-is-that-slow/
- https://agentduel.com/leaderboard?excludeEarlyVotes=true
- https://www.vals.ai/benchmarks/vals_index
- https://artificialanalysis.ai/#intelligence-efficiency-tabs