Models#Benchmark#Model inference

Google ships Gemini 4 Neon at $3/$15 with a 61 on Artificial Analysis

Google shipped Gemini 4 Neon on Nov 18: $3 per million input tokens, $15 output, a 1M-token context and a reported 61 on Artificial Analysis, live the same day.

Google blog header: the Gemini 4 Neon wordmark on a purple gradient

Three weeks after Gemini 4 Carbon, Google released another mainline model on November 18: Gemini 4 Neon. Unlike Carbon’s staged rollout, Neon shipped the same day across the Gemini API in Google AI Studio, Gemini for business, Google Antigravity and the Gemini app, available to paying AI Pro and Ultra subscribers. Its position is explicit: between the lighter Hydrogen and the strongest Carbon, built for software engineering, knowledge work and science workflows at a lower everyday price.

Key points

  • Price and specs: $3 per million input tokens, $0.15 per million cached input tokens (a 95% discount), $15 per million output tokens, a 1 million-token context window and four thinking levels (low, medium, default high and xhigh). Against Carbon’s standard $4/$20, Neon is about a quarter cheaper.
  • Same-day availability: the Gemini API and AI Studio, Gemini for business (API and model dropdown), Antigravity and the Gemini app for paid Pro and Ultra subscribers. Chinese outlet ifanr notes Google skipped the waitlist-style launch this time.
  • The benchmarks are vendor-reported: Google says Neon scores 61 on the Artificial Analysis Intelligence Index at its xhigh level, third overall, ranks #2 on the Vals Index and on its Pareto frontier, and recorded its lowest fabrication rate yet in internal Antigravity testing. None of this has independent replication yet; selection should rest on your own measured tasks.

Google blog header: the Gemini 4 Neon wordmark on a purple gradient

Where Neon sits

Google’s blog lays out the family with an element-table graphic: Hydrogen for daily coding and light agent work, Carbon for the hardest long-horizon problems, Neon in the middle. One detail stands out: Neon can act as a high-judgment orchestrator directing parallel Hydrogen subagents, cutting redundant file reads, sharing progress and handling context compaction in long sessions.

That positioning explains the price. A coding agent’s main loop rereads files, runs commands and compacts history, so token consumption far exceeds single-turn Q&A. Making the middle tier $3/$15 with a 1M-token context moves much of that high-frequency spend off Carbon, which stays reserved for multi-day jobs where a wrong step is costly. ifanr also reports that Ultra subscribers got access to Project Genie and the Genie 4 world model in the same window.

The pricing reflects a shift in usage shape. Google’s post highlights stable context compaction in long sessions and Neon’s role orchestrating parallel Hydrogen subagents — exactly the work coding agents repeat daily. For Pro and Ultra subscribers, Neon is already in the app and Antigravity with no extra application; for API teams, the cache-hit price is what decides the bill.

Analysis

  • Don’t treat 61 as the whole story: the Artificial Analysis index is one fixed suite. Neon trails Carbon by a single point, but their thinking levels, latency and prices differ entirely; the real comparison is tokens per task times unit price.
  • Cached input is the hidden variable: $0.15 per million cached tokens matters enormously for coding agents that resend the same context; the bill can land far below list price if your requests hit the cache.
  • The cadence is accelerating: a new flagship every three weeks — Carbon’s promotional pricing runs to November 30 while Neon already succeeds it. Teams planning long-term capacity around a single model carry growing risk; see our earlier analysis of Gemini 4 Carbon.

Summary

If your workload is coding and knowledge work, Neon belongs in the candidate pool: run a small set of real tasks against the Opus 5.5, GPT-6.2 Sol or Carbon you use today, record tokens, cache hit rate and the final bill, then pick a primary model. The benchmark and “cost-performance” claims come from the vendor and secondary reports; the migration case has to be your own controlled test.