What Google released

Google released Gemini 4 Neon on 18 November 2026. The Google Cloud model page lists it as generally available under the model ID gemini-4-neon, with a knowledge cutoff of March 2026. Neon is the fourth Gemini 4 model in about two months. 9to5Google reports that Carbon arrived on 10 November and Hydrogen on 12 November, with Argon released before both. Google places Neon between Hydrogen, the fast everyday model, and Carbon, which handles the hardest long-horizon work. Google's launch post pitches Neon at complex, high-frequency workflows where cost per task matters.

The release also changes the consumer app. Per 9to5Google, Neon takes Argon's place in the Gemini app's model picker under the label "Deeper analysis", and Google is retiring Gemini 3.1 Pro. Builders on Argon or 3.1 Pro should treat Neon as the migration target, and Google's developer guide is written with that in mind.

Specs and how the model behaves

Neon takes text, image, audio and video as input and returns only text. The context window is 1,048,576 tokens, and input and output share that one budget. A single response segment tops out at 65,536 output tokens. The developer guide describes "long decodes" for longer output, where the SDK or a server-side background job chains further segments together until the shared limit is reached. Long decodes need the v1beta1 API and minimum SDK versions, and the OpenAI compatibility layer can't retrieve them. Google bills each completed segment when that segment finishes, so a failure late in a long job still costs you the earlier segments.

Reasoning is controlled with a thinking_level parameter set to LOW, MEDIUM, HIGH (the default) or XHIGH. Thinking tokens are the hidden reasoning a model writes before its answer. They're billed as output and they count against max_output_tokens at the higher levels. The guide matches each level to a kind of work: LOW for classification, routing and extraction, and XHIGH for proofs and multi-file engineering.

Several familiar controls are gone. Neon ignores temperature, top_p and top_k. Frequency and presence penalties, thinking_budget, prefilled model turns and candidate_count values above 1 all return HTTP 400. Explicit context caching through cached_content isn't supported either. Neon relies on implicit caching, which applies automatically to shared prompt prefixes of 4,096 tokens or more. Google's guide recommends thinking_level and response_schema for steering output, so any pipeline that depended on low temperature for repeatable results will need a structured-output schema and testing.

The guide also describes a planner pattern where Neon plans and checks the work while cheaper Hydrogen subagents carry out individual steps, linked through previous_interaction_id. The pattern fits the pricing gap between the two models, covered below.

Pricing

The AI Studio pricing page sets Neon's standard rates at US$3 per million input tokens, US$15 per million output tokens and US$0.15 per million cached input tokens. Batch and Flex tiers halve those rates to US$1.50 and US$7.50. Priority costs US$5.40 input and US$27 output. Neon has no free tier, and the page lists no separate surcharge for prompts over 200K tokens. Search grounding comes with 5,000 free requests a month, shared across Gemini 3 and newer models, then costs US$14 per 1,000.

Comparing Neon with its siblings depends on the date. Hydrogen costs US$1.50 input and US$7.50 output, so Neon is exactly double. Carbon is on a promotional US$2/US$10 until the end of November, and from 1 December 2026 the pricing page shows Carbon rising to US$4/US$20. After that change, Neon's per-token price will be 25% below Carbon's. Google's own figure puts Neon's cost per task at default HIGH thinking at 32% below Carbon's.

Per-token price tells only part of the cost story. Artificial Analysis measured 283 million output tokens for Neon at XHIGH to complete the Intelligence Index, against a median of 65.4 million for comparable models, and calculated a cost of US$6.44 per index task. A model that writes four times as much reasoning can cost more than a model with higher rates per token.

Benchmarks and their caveats

Google's launch post quotes a long list of results. The strongest claims are first place on the Artificial Analysis Coding Agent Index at 58.6, 65.2% on Terminal-Bench 4.1 and 46.2% on FrontierCode 2.0 at the default setting. Google also reports a score of 61 on the Artificial Analysis Intelligence Index at XHIGH, ranked third overall when the post was written.

We'd read those numbers with three caveats in mind. Some of the coding results depend on the harness, meaning the agent scaffolding around the model. Google's SWE-sweep figure is 12.8% in the neutral mini-SWE harness and 39.0% inside Google's own Antigravity tool, so your results will depend on your own agent loop. Many of the comparisons in the post are against unnamed "competing" models. Several of the benchmarks are new enough that few independent results exist to compare against.

The independent Artificial Analysis page confirms the score of 61 at XHIGH, but when we checked it ranked Neon sixth of 95 models. The page measured output speed at 110 tokens per second, which ranked 71st of 94, and 56.5 seconds before the first answer token. Dropping to MEDIUM cut the score to 55 and the first-token wait to about 20 seconds. Google describes Neon as running at Flash-level speeds, but at the higher thinking levels the independent measurements don't support that claim for interactive use.

Access and limits for New Zealand builders

The simplest route is the paid tier of the Gemini API in Google AI Studio, according to Google's launch post. Neon is also available in Antigravity and the Gemini app for Google AI Pro and Ultra subscribers. A cyber-defence variant is going only to vetted organisations.

The enterprise route has a gap for anyone with data residency requirements. On Google's Gemini Enterprise Agent Platform, Neon runs from the global endpoint or from the us and eu multi-regions, and ML processing happens only in the US or EU. The model page lists no Australian or Asian region, and it says the model is "available to verified enterprise customers in select regions." NZ teams that must keep processing in Australia can't use Neon there yet. Those teams will need to wait or keep that workload on another model. The platform does support customer-managed encryption keys (CMEK) and VPC Service Controls for teams that can accept US or EU processing.

Where Neon fits

We think Neon makes most sense as a planner or reviewer in batch and background agent work. In that role its benchmark strength counts, while its slow first token matters less. The batch tier at US$1.50/US$7.50 brings the price close to Hydrogen's standard rate. Teams running Argon or Gemini 3.1 Pro should test Neon soon, before Carbon's price rises on 1 December, and should fix the thinking level explicitly in code, since the default changed to HIGH. Run your own evaluation with your own harness before you rely on Google's numbers, and track output tokens per task alongside accuracy. If you build something with Neon, the society's projects page is open for members to share their results.

Sources

Build with people working on the same things

Join the Auckland Society for AI: NZ$599 a year with a ForgeBox and the members' projects area. Our Auckland meetups are open to everyone.

Become a member