One-line summary
Neon is the mid-tier workhorse of Gemini 4: $3 input, $15 output, 1M context, roughly 3x Carbon's speed. It is not a Carbon replacement; it compresses Carbon-class capability into a more commonly used price band.
Key figures
| Metric | Value |
|---|---|
| Release | November 19, 2026 (Beijing time) |
| Input price | $3 per 1M tokens |
| Cached input price | $0.15 per 1M tokens |
| Output price | $15 per 1M tokens |
| Context window | 1M tokens |
| Reasoning depth | four adjustable levels |
| Intelligence index | 61, third globally |
| Availability | AI Studio, Gemini app, enterprise API, Antigravity |
Ten questions
Q1 Where does Neon sit in the Gemini 4 family? Mid-tier workhorse. Carbon is the flagship for long-horizon, hardest tasks; Hydrogen targets high-frequency, latency-sensitive work; Neon sits between them, aimed at complex coding, analytical reasoning and production enterprise workflows. Demis Hassabis called it "Alchemy at its finest."
Q2 When did it ship, and can I use it now? Released November 19, 2026 Beijing time and live the same day in Google AI Studio, the Gemini app, the enterprise API and Google Antigravity. No waitlist.
Q3 How exactly is it priced? $3 per million input tokens, $0.15 for cached input, $15 per million output tokens. Cached input is billed at 5% of the headline rate, a 95% discount.
Q4 Is it cheaper or more expensive than Carbon? Both, depending on timing. Carbon's standard rate is $4/$20, so Neon at $3/$15 is 25% cheaper. But Carbon currently carries a promotional $2/$10 through November 30, 2026, making it cheaper until then. From December 1, Neon becomes the better-value tier.
Q5 How much fits in a 1M-token context? Roughly 700,000 to 750,000 English words, or about 1.5-2 million Chinese characters depending on the tokenizer. Google's emphasis is less on raw capacity than on holding spreadsheets, PDFs, charts and audio/video in one window and returning structured deliverables.
Q6 How should the four reasoning levels be used? They trade reasoning depth for tokens: use low levels for extraction and format conversion, high levels for complex analysis and multi-step reasoning. In practice, pin the level per task type rather than running everything at the top.
Q7 When can I use the cyber-security variant? Google says it will open over the coming days through the Fairwind program to vetted cyber defenders — not general availability. Published scores: AutoBaxBench hard tier 76%, WebExploitBench v1.2 51.26%, SEC-Bench Pro 90.5%, CWE-Bench v2 70.7%.
Q8 Is Neon related to Genie 4? No, they only shipped on the same day. Genie 4 is an interactive world-generation capability for AI Ultra subscribers — generate interactive scenes, change environments with text during exploration, and hear responsive audio. It is a product feature, not a base model.
Q9 Where does 61 rank globally? Third on the Artificial Analysis composite index. GPT-6.2 Astra scored 62 and GPT-6.2 Sol 57 in the same week. Index versions can change, so cross-version comparisons are not valid.
Q10 Should I migrate from Carbon to Neon? Three checks: whether the work genuinely needs Carbon's long-horizon capability; whether the post-November-30 price difference justifies switching; and whether 3x speed converts into real concurrency. High-volume, mid-complexity coding and analysis favor Neon; long-horizon autonomous tasks with expensive failures favor staying on Carbon.
Common misreadings
First, treating "3x the speed of Carbon" as uniformly faster. It is a claim about specific scenarios; the speedup varies by task.
Second, comparing 61 directly with this site's October Carbon figure of 59. Those may come from different index versions; confirm the methodology first.
Third, assuming same-day availability means stable enterprise readiness. General availability removes gating, not quota, regional availability or contract terms.
A worked cost example
Suppose a team sends 200 million requests a month, averaging 2,000 input tokens and 500 output tokens — 400 million input tokens and 100 million output tokens in total.
On Neon's $3/$15: about $1,200 for input plus about $1,500 for output, roughly $2,700. On Sol's $2/$10: about $800 plus $1,000, roughly $1,800. On Astra's $8/$40: about $3,200 plus $4,000, roughly $7,200.
The absolute gaps are $900 and $4,500 respectively. At moderate volume, headline unit prices produce limited absolute differences; cache hit rates and retry ratios move the bill far more than the sticker price does. That is why "cost per successful task," which requires running your own benchmark, is a better selection criterion than "price per million tokens."
What to verify before committing a workload
Three things are worth testing on real traffic before migrating anything: the actual cache hit rate on your prompts, the rework rate on your hardest task type, and the latency at the reasoning level you intend to use. Vendor benchmarks rarely match a specific production trace, and the price difference between tiers is often smaller than the cost of one failed batch.
Boundary
- Pricing, context, reasoning levels and channels come from Google and reporting by multiple outlets;
- Cyber-security scores are Google's own evaluations;
- Token-to-character conversions are commonly used ranges and depend on tokenization;
- Index scores are third-party measurements and version-sensitive;
- This article is not procurement advice.
