Google releases Gemini 4 Neon, priced at $3/$15 per million tokens with a 1M-token context window, rolling out across API, Antigravity and the Gemini app
Edited by Elliot Marsh, Editor · written with AI from the linked sources (how)
Google began rolling out Gemini 4 Neon on 18 Nov 2026 in the Gemini API (Google AI Studio), Gemini for business, Google Antigravity (paid AI Pro and Ultra subscribers) and the Gemini app (Pro and Ultra subscribers). Pricing is $3 per million input tokens, $0.15 per million cached input tokens and $15 per million output tokens, with a 1M-token context window and four thinking levels (default is high). A cyber-specialised variant, Gemini 4 Neon for cyber, is due to reach trusted defenders in the coming days through the Fairwind Program. Google reports benchmark results: #3 on the Artificial Analysis Intelligence Index (61 at xhigh), #2 on the Vals Index (70.9%), #1 on the Artificial Analysis Coding Agent Index (58.6), and #2 on Arena's Agent Arena Index (1336). It also reports 65.2% on Terminal-Bench 4.1 and 39.0% on SWE-sweep within Antigravity. Google says cost per task is 32% lower than Gemini 4 Carbon at the default setting.
Entities: Google, Google DeepMind, Koray Kavukcuoglu, Gemini 4 Neon, Gemini 4 Carbon, Google Antigravity
1 primary
What happened
On 18 Nov 2026 Google began rolling out Gemini 4 Neon in the Gemini API (AI Studio), Gemini for business, Antigravity (paid AI Pro and Ultra) and the Gemini app (Pro and Ultra). API pricing is $3 per million input tokens, $0.15 cached input and $15 output, with a 1M-token context window and four thinking levels (default high). Google says it ranks #3 on the Artificial Analysis Intelligence Index, #2 on the Vals Index, and #1 on the Artificial Analysis Coding Agent Index, with cost per task 32% below Gemini 4 Carbon at default settings. A cyber-specialised variant is due for vetted defenders via the Fairwind Program in the coming days.
Why it matters
Developers and enterprises already on Gemini get a new model at a published price and can test it today, particularly for coding agents and long-context work. The pricing and cost-per-task claims matter most to teams running high volumes, because they could change which vendor is cheapest for a given workload. The rankings show Google competitive at the frontier but not clearly ahead: #3 overall on one index and #2 on another. Whether it is worth switching depends on your own tasks, not on a leaderboard.
What is noise
Benchmark ranks, the 50-80% cheaper-than-rivals figure, "lowest fabrication rate yet" and "state-of-the-art cyber defence" are all Google's own claims, with no independent verification yet and no links to the underlying evidence in the extraction. Cost per task depends on how many tokens a model spends thinking, so a low per-token price does not guarantee a low bill. Statements like "frontier-level intelligence" are marketing, and the customer names listed (Surge AI, Zapier, Andon Labs) are endorsements, not tests.
Watch next
- 01Independent re-runs of the Artificial Analysis, Vals and Arena results, plus third-party cost-per-task comparisons against rival frontier models on real coding and agent workloads.
- 02Actual API bills and token usage from early adopters at the default high thinking level, including rate limits and any later price changes or tier restrictions.
- 03Fairwind Program access terms and the first independent evaluations of Gemini 4 Neon for cyber, including any misuse reports or safety findings.
Coverage
4 stories- Introducing Gemini 4 NeonGoogle Gemini Blog · primary · 18 Nov 2026Tier 1
- Google ships Gemini 4 Neon, its fourth new Gemini model in a monthThe Decoder · 18 Nov 2026Tier 2
- llm-gemini 0.38Simon Willison · 18 Nov 2026Tier 3
- Google's latest AI model is faster, cheaper and smarter than its flagship predecessor from just weeks agoThe Decoder · 19 Nov 2026Tier 2
Newsletter
Get the no-hype AI news straight to your inbox
Weekly. No spam, unsubscribe any time. Privacy
More capability signals
Full feed →- Deepseek releases V4.1-Flash, an open-source model that sharply cuts KV cache memory and input-processing compute for AI agents10 Sept 202682
- Mistral releases Mistral Large 4, a ~1T-parameter open-weight multimodal MoE model, with day-0 Hugging Face support29 Oct 202680
- Meta releases Muse Spark Open, an open-weight ~816B-parameter multimodal mixture-of-experts model, with Hugging Face support15 Oct 202680
- OpenAI discloses sandbox-escape and credential-leak incidents, confirms pause on tool-use for its most capable models26 Sept 202680