News · 2026-11-19
Google releases Gemini 4 Neon with a lower-cost frontier pitch
Google has released Gemini 4 Neon, a proprietary AI model adapted from Gemini 4 Carbon for frequent complex work. Independent tests place it near the top of one index at a lower test cost than the leader, although those results do not establish that it is faster or cheaper for every user.
Key facts
- Google announced Neon on November 18, 2026.
- Vals measured 70.87% and $7.40 per full index run; the leading model scored 71.31% at $17.94.
- The paid API lists $3 per million input tokens and $15 per million output tokens.
- Primary sources: Google's launch announcement, model card and pricing.
A model's sticker price is only one part of the bill. How much text it reads, how much internal reasoning it generates, and whether repeated input can be reused all matter. Neon supports implicit caching: previously processed input may be reused at the listed cached-input price. Explicit caching is not supported. Google counts thinking tokens in output billing, and grounding services can incur separate charges. The existing prompt caching lesson explains why a repeat visit can cost less than the first.
Google calls Neon adapted from Carbon. Its model card does not give the adaptation recipe or a parameter count, so more specific claims about distillation or model size are unsupported. Google's API documentation lists text, image, audio, video and PDF input and text output, along with tools such as code execution and search grounding. The API Free tier does not include the model. Google separately describes access for eligible Pro and Ultra subscribers in its applications; the product entitlements should not be confused.
The one-million-token context limit is combined input and output, while ordinary synchronous output is capped at 65,536 tokens. Google documents a background long-decode route for longer output. Think of context as the size of a desk holding both the source material and the answer, not the guaranteed length of a single immediate reply.
Vals's independent index places Neon close to the leader with overlapping uncertainty bands. Vals calls its sector-weighting scheme a “vast oversimplification,” a useful warning against treating a single blended score as a definitive ranking. Its cost figure applies to that suite, its prompts and settings. Artificial Analysis reports 21.5 minutes per index task and 56.51 seconds to the first answer token at xhigh effort; these are benchmark measurements, not typical chat latency.
Google's broader claim that Neon matches Carbon at three times the speed and 60% of the cost remains a company comparison, not a generally replicated finding. Latent Space and the Hacker News discussion show early enthusiasm alongside capacity anecdotes and questions about model positioning. The practical test is a user's own workload, with the same task, effort setting, cache behavior and quality threshold for each candidate.
The model card also sets an important safety-evaluation boundary. Google says Neon increased general capability over Argon under its own material-capability-change assessment but fell below the threshold for a full critical-capability evaluation suite. It therefore carries forward Argon's relevant determinations. The Neon safety report was marked forthcoming when the dossier's researcher checked. This describes the scope of Google's process; it is not evidence that Neon failed a particular safety test.
An early field report cited by Latent Space recorded 13 sessions and 57 capacity errors, with four stalled runs. It was a small, self-selected sample rather than an availability study. In practice, a buyer may care as much about capacity and successful completion as the price of a token. A measured score under one harness also includes how that harness uses tools, sets effort and handles errors; changing those details can change the result.
Key questions
Is Gemini 4 Neon free in the API?
Is Neon a distilled version of Carbon?
How much does Neon cost?
Cite this
APA
Ground Truth. (2026, November 19). Google releases Gemini 4 Neon with a lower-cost frontier pitch. Ground Truth. https://groundtruth.day/news/google-gemini-4-neon-paid-api-cost-and-benchmarks.html
BibTeX
@misc{groundtruth:google-gemini-4-neon-paid-api-cost-and-benchmarks,
title = {Google releases Gemini 4 Neon with a lower-cost frontier pitch},
author = {{Ground Truth}},
year = {2026},
month = {nov},
url = {https://groundtruth.day/news/google-gemini-4-neon-paid-api-cost-and-benchmarks.html}
}
Comments are replies to this story on Bluesky — reply with any Bluesky account to join in.