Google says it launched Gemini 4 Neon on November 18, opening a model built for complex coding, analytical reasoning and knowledge work through its paid API and selected subscription products.
For small-business owners, the question isn’t benchmark rank but whether Neon produces better customer-facing work at an acceptable finished-task cost.
Google says Neon is rolling out in the Gemini app for Google AI Pro and Ultra subscribers, alongside access through Google AI Studio, Gemini for business and Google Antigravity.
API pricing is $3 per million input tokens, $15 per million output tokens and $0.15 per million cached input tokens. Google DeepMind describes Neon as adapted from Carbon and positioned between Hydrogen’s everyday execution and Carbon’s hardest, long-running projects.
Its model card lists a 1 million-token combined input-output context window, up to 64,000 default synchronous output tokens and up to 1 million output tokens through background long decode.
Price depends on the date and setting
According to 9to5Google, Argon and Carbon have temporary API rates of $2 per million input tokens and $10 per million output tokens through November 30, below Neon’s launch rates. Their scheduled rates then rise to $4 and $20 respectively, making Neon’s rates 25% lower from December 1.
The outlet also reports that Neon offers four thinking levels: Low, Medium, High and Extra High.
Google’s launch announcement reports 32% lower cost per task than Carbon on the Artificial Analysis index at Neon’s default high setting.
VentureBeat reports that the advantage narrows to about 17% at xhigh against Carbon at high.
Benchmarks aren’t customer outcomes
Google’s announcement and model card report:
- Third place, scoring 61 at the highest thinking level, on the Artificial Analysis Intelligence Index v5.0.1; second place at 70.9% on Vals Index v2.2; first on Legal Research Bench and Tax Agent Bench.
- AutomationBench: 56.3% versus GPT-6.2 Sol’s 41.3%. Chartography’s chart-reading test: 73.5% versus Sol’s 75.1%.
- Agents’ Last Exam v1.1: 75.8%, against Claude Opus 5.5’s 71.6% and Sol’s 73.0%.
- Agent Arena Index steerability: second place, scoring 1336 across more than 4,200 real-world agent sessions. Google’s internal agent honesty evaluation: 99.2% pass rate.
These are results published in Google’s announcement and model card, not evidence of improved customer conversion. The card also includes hallucination cautions, a March 2026 knowledge cutoff and a Safety & Security Report listed as forthcoming.
Coding results underline how much setup matters: SWE-sweep scores range from 12.8% with the mini-SWE harness to 39.0% in Antigravity for the same model. On FrontierCode 2.0, Neon scores 46.2% at default high versus Opus 5.5’s 47.3%. Neither measures marketing quality.
A bounded, human-reviewed comparison offers a more relevant test: draft one service description or inquiry-response template with Neon and the current tool. Track editing time, factual corrections and total cost per acceptable finished draft, including review. The same principle applies to a lower-cost AI content comparison.
Wider use should depend on better usable output, less correction work or an acceptable total-cost trade-off, not launch rankings.
Sources
- https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-neon/
- https://deepmind.google/models/model-cards/gemini-4-neon/
- https://9to5google.com/2026/11/18/gemini-4-neon-announcement/?extended-comments=1
- https://venturebeat.com/technology/google-launches-gemini-4-neon-with-opus-5-5-level-coding-performance-at-25-lower-api-token-prices
- https://www.androidauthority.com/gemini-4-neon-3736931/