Hot AI News

Gemini 4 Neon costs more per token than Carbon, but Google says tasks cost less

Google's new Gemini 4 Neon is priced at $3 and $15 per million tokens, above Carbon, yet Google says it is 32% cheaper per task at its default thinking level.

Gemini 4 Neon costs more per token than Carbon, but Google says tasks cost less
Image: Google

Google has introduced Gemini 4 Neon, a model it positions between Gemini 4 Carbon and Hydrogen. It costs $3 per million input tokens and $15 per million output tokens, and it is rolling out from November 18.

Who can use Neon today, and at what price

Google lists a 1 million-token context window and cached input at $0.15 per million tokens, which it describes as a 95% discount. Neon is rolling out in the Gemini API in Google AI Studio, in Gemini for business (through the API and the model dropdown), and in Android Studio. Paid Google AI Pro and Ultra subscribers can use it in Google Antigravity, and Pro and Ultra subscribers can pick it in the Gemini app.

Chart of Artificial Analysis cost per task for Gemini 4 Neon
Image: Google

A separate version, Gemini 4 Neon for cyber, is not open to everyone. Google says eligible members of its Fairwind Program will get access over the next few days, and that others should contact their Cloud account team. The announcement does not say anything about regional availability or free-tier access.

In the Gemini app, Google says Tasks can now run on Neon, and that it handles multi-step jobs faster than larger models while stretching usage limits further.

Chart showing where Neon sits among the Gemini 4 models
Image: Google
Advertisements

Per-token price is higher, cost per task is not

The pricing looks odd next to the rest of the family. According to our API pricing tracker, which uses OpenRouter’s listed rates, Gemini 4 Carbon (released November 10) is priced at $2 input and $10 output per million tokens. Hydrogen (November 12) is $1.50 and $7.50. Neon’s $3 and $15 therefore sit above both, even though Google describes it as the faster model.

Google’s answer is that tokens are not the right unit. The company says that at the default high thinking level, Neon posts a strong Artificial Analysis Intelligence Index result while a task costs 32% less to run than on Carbon. It also claims roughly 50% to 80% lower cost per task than top competing frontier models on the Vals Index and Agent Arena, and at the highest xhigh setting, 36% or more below rivals at their top settings. Those are Google’s figures, built on third-party benchmarks, and the announcement does not publish the underlying per-task costs in text.

Hero image for Gemini 4 Neon
Image: Google

For comparison, the same tracker lists Claude Opus 5.5 at $4 and $20 and Claude Sonnet 5.5 at $2 and $10. Neon is cheaper than the first and pricier than the second per token. Whether it wins in practice depends on how many tokens a given job burns at each thinking level, which is the variable Google says it is tuning with four levels.

Neon scores 61 on the Artificial Analysis Intelligence Index at xhigh, which Google says ranks #3 overall.

Benchmarks Google chose to highlight

Google says Neon scores 61 on the Artificial Analysis Intelligence Index at xhigh thinking, third overall and just behind larger, slower models. It reports #2 on the Vals Index (70.9%), #1 on the Artificial Analysis Coding Agent Index (58.6), 65.2% on Terminal-Bench 4.1 and 46.2% on FrontierCode 2.0 at its default setting. On SWE-sweep it reports 12.8% on the mini-SWE harness and 39.0% inside Antigravity, which it uses to argue that the harness matters as much as the model.

For science and math, Google cites 70.5% on Terminal-Bench Science, 84.0% on RiemannBench and 71.3% on HLE-Diamond. In business tasks it reports #1 on Zapier’s AutomationBench (56.3%) and on Andon Labs’ Vending-Bench 2 ($16,843), plus a #2 debut on Arena’s new Agent Arena Index with 1336 points across more than 4,200 agent sessions.

The cyber results are the most specific. Google reports 90.5% on SEC-bench Pro for finding vulnerabilities, 51.26% on WebExploitBench v1.2 and 70.7% on CWE-bench v2 for patching, a 3.4% gain over Carbon. It also says Neon refused only 1.98% of tasks on Terminal-Bench 4.0 and none on Terminal-Bench-Science, a sign that it is trying to cut unnecessary refusals while keeping safeguards against cyber and biosecurity misuse.

Advertisements

What the announcement does not settle

Nearly every number above comes from Google’s own post. Some are from third-party leaderboards, but the company chose which to show, and the safety figures (a 0.09% misaligned action rate, a 99.2% pass rate on agent honesty) come from Google’s internal evaluations. Cost-per-task claims depend on the benchmark mix and on each rival’s chosen settings.

Neon also sits in a crowded week. Our tracker shows OpenAI’s GPT-6.2 Astra and Sol both listed on November 17, at $8 and $40 and at $2 and $10 per million tokens respectively. Neon’s case rests on per-task efficiency rather than sticker price, and independent tests will show whether that holds outside Google’s chosen benchmarks. For now, developers on the paid API tier and Pro or Ultra subscribers can try it directly.

Sign Up for the Daily AI Pulse

One email a day. All the stories that matter.

Alex CarterAI News & Big Tech Correspondent

Alex Carter writes for Aiholics, keeping readers updated on the fast-paced world of AI and Big Tech.

Leave a Reply

Your email address will not be published. Required fields are marked *

You may also like