Oleksandr Kuzmenko, EditorPublished 7 min readAI-assisted · editor-reviewed
TL;DR
- Prompt caching cuts input token rates by 95% ($0.15/1M vs $3.00/1M) over a 1M context window.
- Scores 58.6 on the Artificial Analysis Coding Agent Index to take the top ranking.
- Features four thinking tiers and reduced context-compaction decay for long multi-agent sessions.
Key facts
- Input token price
- $3.00 / 1M tokens
- Cached input token price
- $0.15 / 1M tokens (95% discount)
- Output token price
- $15.00 / 1M tokens
- Context window
- 1,000,000 tokens
- Artificial Analysis Coding Agent Index
- 58.6 (#1 overall)
- Mini-SWE score in Antigravity
- 39.0% (self-reported)
Source: Key facts — Context window
Source: Key facts. Context window: 1,000,000 tokens. Each figure is copied from the story.
- Context window1,000,000 tokens
Source: Key facts — Input token price, Output token price
Source: Key facts. Input token price: $3.00 / 1M tokens. Output token price: $15.00 / 1M tokens. The chart compares only values that share a unit. Each figure is copied from the story.
Metric chart
- $3.00 / 1M tokensInput token price
- $15.00 / 1M tokensOutput token price
- Input token price$3.00 / 1M tokens
- Output token price$15.00 / 1M tokens
Pricing and Context Architecture
Google has officially launched Gemini 4 Neon, positioning it as an efficient frontier model optimized for long-horizon agentic workflows, multi-step instruction following, and production coding.
Key economics and operational limits include:
- Input pricing: $3.00 per million tokens
- Cached input pricing: $0.15 per million tokens (a 95% prompt-caching discount)
- Output pricing: $15.00 per million tokens
- Context capacity: 1 million tokens
- Thinking levels: Four granular tiers, scaling from default
hightoxhighfor deep multi-step reasoning
Coding and Agent Benchmarks
Neon is built to act standalone or as a high-judgment orchestrator delegating to parallel subagents in harnesses like Google Antigravity. It is engineered to minimize duplicate file reads and maintain coherence during context compaction.
On external evaluations, Gemini 4 Neon scores:
- Artificial Analysis Coding Agent Index: 58.6 (#1 overall)
- Terminal-Bench 4.1: 65.2%
- FrontierCode 2.0: 46.2% at
highthinking - SWE-sweep mini-SWE harness: 12.8% standalone, scaling to 39.0% inside Google Antigravity
- Arena Agent Arena Index: 1336 (#2 globally across 4,200+ sessions)
Cyber Defense and Availability
For security engineering, Neon achieved 90.5% on SEC-bench Pro, 51.26% on WebExploitBench v1.2, and 70.7% on CWE-bench v2 for patching flaws. Gemini 4 Neon is accessible today in the Gemini API through Google AI Studio, Gemini for business, and for Pro and Ultra subscribers in Google Antigravity.
Try it in 2 minutes
curl -X POST "https://generativelanguage.googleapis.com/v1beta/models/gemini-4-neon:generateContent?key=${GEMINI_API_KEY}" \
-H 'Content-Type: application/json' \
-d '{"contents": [{"parts": [{"text": "Refactor this repository module for optimal caching."}]}]}'When to use
- Use for complex multi-file repository refactoring, automated bug fixing, and long-running agent swarms.
- Deploy in agent systems reading large codebases or documents that benefit from the 1M window and cached tokens.
When NOT to use
- Do not use for simple single-file scripts where small local models or sub-dollar tiers offer zero-latency responses.
- Avoid if your workflow does not reuse system prompts or schemas, missing the 95% caching cost benefit.
What to do today
- Test Gemini 4 Neon in Google AI Studio or update your API endpoint configuration.
- Structure agent system prompts and static codebase files to leverage the 95% prompt caching discount.
- Evaluate the four thinking levels to balance token spend against reasoning depth.
Step names are taken from What to do today.
Step names are taken from What to do today. Test Gemini 4 Neon in Google AI Studio or update your API endpoint configuration.. Structure agent system prompts and static codebase files to leverage the 95% prompt caching discount.. Evaluate the four thinking levels to balance token spend against reasoning depth.
- Test Gemini 4 Neon in Google AI Studio or update your API endpoint configuration.
- Structure agent system prompts and static codebase files to leverage the 95% prompt caching discount.
- Evaluate the four thinking levels to balance token spend against reasoning depth.
What the community says
So basically: haiku -> hydrogen opus -> neon fable -> carbon No mention of Argon in that page though.
I don't really use AI for that sort of thing but even for technical stuff I still find it's writing style a real turn off. Gemini still has lots of AIisms but it's a lot more tolerable.
