ML in Production · 3 min read

Google Ships Gemini 4 Neon: Mid-Tier Model at $3/$15 Per Million Tokens

Google launches Gemini 4 Neon at $3/$15 per million tokens, claiming 3x speed over Carbon and 60% cost cuts, with a 1M-token context and four thinking levels.

By Tom Whitfield643 words

Key takeaways

  1. Gemini 4 Neon costs USD 3 per million input tokens and USD 15 per million output tokens
  2. Google claims Neon is about 3x faster than Gemini 4 Carbon with roughly 60% lower costs
  3. Cached input tokens cost USD 0.15 per million, a 95% discount
  4. The model supports a 1 million-token context window and four thinking levels
  5. Available via Gemini API, Google AI Studio, Gemini for Business, and paid AI Pro/Ultra tiers
Google Launches Gemini 4 Neon to Speed up Complex AI Tasks
Google Launches Gemini 4 Neon to Speed up Complex AI Tasks — Pexels / pixabay

Google has launched Gemini 4 Neon at USD 3 per million input tokens and USD 15 per million output tokens. The company claims the model runs roughly three times faster than Gemini 4 Carbon while cutting overall costs by about 60%. Those figures come from Google itself. No third-party benchmark accompanies the launch.

Where does Neon sit in the lineup?

Neon occupies the middle position between two existing models. Gemini 4 Carbon targets demanding, long-duration projects. Gemini 4 Hydrogen handles quick, everyday tasks. Neon aims at work that is complex but time-sensitive — software engineering, knowledge work, and scientific workflows. Google frames this as a middle path, sparing customers a binary choice between heavyweight and lightweight models. That framing is convenient. It also positions Google to upsell subscriptions across tiers.

Availability is broad from day one:

  • Gemini API in Google AI Studio
  • Gemini for Business
  • Paid Google AI Pro and Ultra subscriptions
  • Google Antigravity and the Gemini app

What can the model actually do?

Google says Neon combines advanced reasoning with fast end-to-end task completion (the full pipeline from prompt to usable output). For developers, the claimed capability is debugging and refactoring code across multiple files while making targeted edits. For knowledge workers and researchers, the model is designed to follow multi-step instructions and synthesize large volumes of information.

These are claims, not verified results. Google has published no benchmark suite alongside the announcement. Data teams should treat the capability statements as vendor positioning until independent evaluations appear.

How large is the context window?

Neon supports a 1 million-token context window (the span of text the model can process in a single session). Google says this accommodates lengthy documents, large codebases, tables, and charts in one pass. That matches the context sizes already standard across the Gemini 4 family, so it is table stakes rather than a differentiator.

The model also ships with four thinking levels: low, medium, high, and extra high. Users adjust the balance between speed and depth through these settings. This mirrors a broader industry pattern — reasoning-effort controls now appear in most frontier model APIs — and it gives teams a cost lever they previously lacked within a single model.

What does it cost in practice?

The pricing structure rewards caching:

  • Input: USD 3 per million tokens
  • Output: USD 15 per million tokens
  • Cached input: USD 0.15 per million tokens, a 95% discount on the standard input price

For workloads with repeated context — a large codebase queried many times, for instance — the cache discount matters more than the headline input price. Teams running multi-file refactoring loops should model their cache hit rate before comparing Neon against Carbon or Hydrogen on list price alone. Google's 60% cost-reduction claim against Carbon is a rough aggregate; actual savings depend heavily on token mix and cache behavior.

What should data teams do now?

Three practical steps follow from the announcement. First, benchmark Neon on your own tasks rather than trusting the speed claim — the three-times figure has no published methodology. Second, exploit the four thinking levels to match spend to task complexity; defaulting to the highest level wastes the model's cost advantage. Third, instrument caching early, since the 95% discount on cached input is where the economics tilt in Google's favor and yours.

The strategic signal is clear. Google said the launch reflects its effort to give enterprise and professional users more control over AI performance, cost, and speed. Translation: the tier war is now about granularity of control, not just raw capability. Expect the middle tier to become the default enterprise seat. Whether Neon's reasoning holds up under sustained agentic use (multi-step autonomous task execution) remains unproven — Google did not address failure rates, latency distributions, or throughput limits in the announcement, and those omissions are where skepticism belongs.

Analytics Insight (Source)

Price tags

Share this article:

More from Tom Whitfield

Tom Whitfield

Show full bio

Staff writer covering regulation, policy and their impact on Analytics Digest is a professional daily at Analytics Digest.

16 articles

Read also

« Previous articleNext article »