Starts Nov 23Live Cohort 3 of Vibe Coding & AI Agents: 3 weeks, 9 live hands-on sessionsCohort 3 (Nov 23) →
ReleaseAPI / Product · ActionableDeepMind Blog→ original

Google releases Gemini 4 Neon: 3rd place in AI ranking and tasks 32% cheaper

Google has released Gemini 4 Neon, a model that nearly matches the largest AI systems but runs faster and costs less. In the independent Artificial Analysis ranking it scored 61 points and took third place, while a single task costs 32% less than on the larger Gemini 4 Carbon. Input costs $3 per million tokens and output costs $15. If your agent fixes code on an expensive flagship model, it is worth running its tasks on Neon.

AI-processed from DeepMind Blog; edited by Hamidun News
Google releases Gemini 4 Neon: 3rd place in AI ranking and tasks 32% cheaper
Source: DeepMind Blog. Collage: Hamidun News.
Listen to article
Article Fact Sheet & Key Figures
ParameterDetails
Primary sourceDeepMind Blog ↗
Editor-in-Chief
SWE-sweep39.0% in Google Antigravity versus 12.8% in mini-SWE
Key metric #2Cached input costs $0.15 per million tokens, a 95% discount
Key metric #3Context window is 1 million tokens; Vals Index: second place (70.9%)

Google DeepMind has introduced Gemini 4 Neon, a model that nearly matches the largest AI systems but runs faster and costs less. In the independent Artificial Analysis Intelligence Index (a composite score across a set of reasoning and knowledge tests), it scored 61 points at the maximum xhigh thinking level and took third place, just behind larger and slower models. At the high thinking level, which is enabled by default, a single task costs 32% less than on Gemini 4 Carbon, the flagship of the family.

Picture an agent that fixes bugs in a repository every day: each run consumes thousands of tokens, and the per-task price difference quickly turns into real money. API pricing is $3 per million input tokens and $15 per million output tokens.

What the model is and what it costs

Each model in the family has a different role. Carbon takes on the heaviest long-running projects, Hydrogen handles fast everyday execution, and Neon balances depth and efficiency. The context window is one million tokens, so a single request can hold large spreadsheets, PDFs, charts, audio and video. Cached input (when the same part of a request is repeated) costs $0.15 per million tokens, a 95% discount.

Neon has four "thinking" levels, so teams can choose how much to pay for reasoning depth. At the high level, the model costs 50–80% less per task than leading competing models and is among the best in price-to-quality terms (on Vals Index and Agent Arena). At the maximum xhigh level it posts its best score and still costs at least 36% less than competitors running in their maximum modes.

Code, agents and office work

The main bet is on long agentic workflows. Neon ranks first in the Artificial Analysis Coding Agent Index (58.6), scores 65.2% on Terminal-Bench 4.1 and 46.2% on FrontierCode 2.0. On SWE-sweep it gets 12.8% in the lightweight mini-SWE harness, but that rises to 39.0% inside Google Antigravity. The gap shows that the tooling around a model affects results as much as the model itself. Google also says Neon rereads the same files less often, makes targeted edits across several files at once, and compresses context carefully in long sessions.

For business tasks, the numbers are as follows: second place in Vals Index (70.9%), first place in Legal Research Bench and Tax Agent Bench, and 56.3% and first place in Zapier's AutomationBench. In Vending-Bench 2, a long business-management simulation, the model earned $16 843. In the new Agent Arena Index, Neon has 1336 points and second place based on more than 4200 real sessions. It is also second in steerability (how well a model accepts user corrections in the middle of a task).

In science, Google calls Neon best in class on Terminal-Bench Science (70.5%), RiemannBench (84.0%) and HLE-Diamond (71.3%).

Security and cyber defense

The Neon for cyber version set records on several tests: 76% on the hard level of AutoBaxBench (secure code generation), 51.26% on WebExploitBench v1.2 at half the cost of larger competitors, and 90.5% on SEC-bench Pro (vulnerability discovery). On vulnerability patching in CWE-bench v2 it scores 70.7%, 3.4% higher than Carbon. Access will open to vetted defense professionals through the Fairwind program in the coming days.

In Google's own tests, the model cheats less often: 0.0% on Terminal-Bench 4.0 and 2.45% on a test of the tendency to cut corners. The share of incorrect actions in the agentic safety suite is 0.09%, and the share of honesty checks passed is 99.2%. Against injections through web pages, emails and documents on the Gray Swan benchmark, with defenses enabled and maximum thinking, Neon repelled all attacks.

What to do in practice

The model is available in the Gemini API through Google AI Studio, in Gemini for business, in Google Antigravity for paid Pro and Ultra subscribers, and in the Gemini app. If your agent currently runs on a flagship model and the bill keeps growing, it makes sense to run your typical tasks on Neon and compare the cost of a successful completion, not just answer quality.

Caveat: almost all the figures come from Google's announcement. Some of them are internal tests, such as the cheating and honesty checks, and the 39.0% on SWE-sweep was achieved in Antigravity's own harness. You will have to verify on your own data: the model that wins benchmarks is not always the one that works best in your repository.

Story

Key Questions Answered

What the model is and what it costs?

Each model in the family has a different role. Carbon takes on the heaviest long-running projects, Hydrogen handles fast everyday execution, and Neon balances depth and efficiency. The context window is one million tokens, so a single request can hold large spreadsheets, PDFs, charts, audio and video. Cached input (when the same part of a request is repeated) costs $0.15 per million tokens, a 95% discount. Neon has four "thinking" levels, so teams can choose how much to pay for reasoning depth.

What are the key details on Code, agents and office work?

The main bet is on long agentic workflows. Neon ranks first in the Artificial Analysis Coding Agent Index (58.6), scores 65.2% on Terminal-Bench 4.1 and 46.2% on FrontierCode 2.0. On SWE-sweep it gets 12.8% in the lightweight mini-SWE harness, but that rises to 39.0% inside Google Antigravity. The gap shows that the tooling around a model affects results as much as the model itself.

What are the key details on Security and cyber defense?

The Neon for cyber version set records on several tests: 76% on the hard level of AutoBaxBench (secure code generation), 51.26% on WebExploitBench v1.2 at half the cost of larger competitors, and 90.5% on SEC-bench Pro (vulnerability discovery). On vulnerability patching in CWE-bench v2 it scores 70.7%, 3.4% higher than Carbon. Access will open to vetted defense professionals through the Fairwind program in the coming days.

ZK
Zhemal Khamidun — CPO at Alpina Digital, product manager of AlpinaGPT. Building AI products since 2019. Hamidun News is a pet project that grew into a production aggregator with 6,000+ readers.
✦ Unified AI Workspace · Nomad

Test Frontier LLMs & Autonomous Agents in One Workspace

Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.

Try Nomad AI →Launch Nomad Agent
What do you think?
Loading comments…