Google launches Gemini 4 Neon at $3 per million input tokens

Google launches Gemini 4 Neon at $3 per million input tokens

Google announced Gemini 4 Neon on its blog. It is a model in the Gemini 4 family, positioned between the larger Carbon, which Google says tackles the hardest long-horizon projects, and Hydrogen, which handles fast everyday execution. Google describes Neon as bringing frontier-level intelligence with fast end-to-end task completion across software engineering, knowledge work and science workflows.

Pricing is $3 per million input tokens, $0.15 per million cached input tokens (a 95% discount) and $15 per million output tokens, with a 1 million-token context window. Rollout begins the day of the post in the Gemini API in Google AI Studio, in Gemini for business (via API and model dropdown), in Google Antigravity for paid Google AI Pro and Ultra subscribers, and in the Gemini app for Pro and Ultra subscribers. Tasks in Gemini can now run on Neon, which Google says works through multi-step jobs more quickly and efficiently than larger models. A separate cyber variant, Gemini 4 Neon for cyber, is due for trusted defenders in the coming days through the Fairwind Program; eligible members get access over the next few days, and access is requested through a Cloud account team.

On general intelligence, Google says Neon ranks #3 overall with a score of 61 on the Artificial Analysis Intelligence Index at its xhigh thinking level, just behind larger, slower models. It ranks #2 on the Vals Index (70.9%) and lands on the Pareto frontier there. It ranks #2 on steerability, meaning how cleanly it adapts to mid-task user corrections, on the new Agent Arena Index, and debuts #2 globally with 1336 on that index across 4,200+ real-world agent sessions. Google also says Neon records its lowest fabrication rate yet in internal Google Antigravity testing.

For enterprise and analytical work, Google says Neon can synthesize spreadsheets, PDFs, charts, audio and video in a single context window and turn them into structured deliverables. It reaches #1 on the Legal Research Bench and Tax Agent Bench, and #1 on most Surge AI benchmarks across finance, accounting, spreadsheet modeling and complex constraint following. Other results: 73.5% on Chartography, 75.8% on Agents' Last Exam v1.1 (on par with leading models), #1 on Zapier's AutomationBench (56.3%) and #1 on Andon Labs' Vending-Bench 2 ($16,843), a long-horizon simulation of running a business. In science and maths, Google calls Neon best in class on Terminal-Bench Science (70.5%), RiemannBench (84.0%) and HLE-Diamond (71.3%).

For coding, Google says Neon ranks #1 overall on the Artificial Analysis Coding Agent Index (58.6), scores 65.2% on Terminal-Bench 4.1 and 46.2% on FrontierCode 2.0 at its default high setting. On SWE-sweep it scores 12.8% on the mini-SWE harness and 39.0% in Google Antigravity, which Google calls state-of-the-art and says is more than triple the mini-SWE figure. The post describes Neon as suited to long-running agent workflows, either standalone or as an orchestrator directing parallel Hydrogen subagents in Antigravity, with fewer redundant file reads and reliable context compaction in long sessions.

Google stresses cost and control. Neon has four thinking levels. At the default high level, Google says it has a 32% lower cost per task than Carbon, and roughly 50% to 80% lower cost per task than top competing frontier models, sitting on the Pareto frontier of both the Vals Index and Agent Arena. At xhigh, where it shows its highest intelligence score, Google says it still costs 36%+ less per task than competing models at their highest setting.

The cyber variant gets its own results. Google says it produces more secure code than Carbon across the easy, medium and hard tiers of AutoBaxBench, with a state-of-the-art 76% on the hard tier. It scores a state-of-the-art 51.26% on WebExploitBench v1.2 at half the cost of competing large models, and 90.5% on SEC-bench Pro at roughly half the cost per challenge. On ExploitGym it reaches frontier-level performance using half the tokens of Carbon. On Collinear's CWE-bench v2, which covers patching, it reaches 70.7%, a 3.4% uplift over Carbon.

On safety, Google says Neon keeps Carbon-level protections against severe misuse and prompt injection, with real-time safeguards against cyber and biosecurity misuse while declining less legitimate work: it refused 1.98% of tasks on Terminal-Bench 4.0 and 0% on Terminal-Bench-Science. In internal coding use, alerts from its chain-of-thought and action misalignment monitor stayed below Carbon's low baselines, which Google says suggests improved alignment propensity. Neon matched Carbon with 0.0% cheating on Terminal-Bench 4.0, took shortcuts on 2.45% of tasks on Google's coding cheating evaluation (the lowest of any Gemini 4 model), had a 0.09% misaligned action rate on its agentic safety suite and a 99.2% pass rate on its agent honesty evaluation. Against Carbon's baseline on its agentic behavior evaluation, Google says under-service fell by 43% to 56% and overreach by 15% to 35% across effort levels. On Gray Swan's indirect prompt injection benchmark, with real-time safeguards at its highest thinking level, Neon resisted every attack, matching the 0.0% attack success rate reached only by other Gemini models.

In the Gemini app, Google pitches Neon for clearer writing, more precise reading of uploads such as spreadsheets, PDFs and photos, pulling in details from Gmail and Docs, and research that draws on current web sources. Its examples include analyzing billing statements to see why a utility charge spiked and cross-checking receipts to negotiate a refund. Google points readers to the Gemini 4 Neon model card for more detail.

Key facts

  • Gemini 4 Neon sits between Gemini 4 Carbon (hardest, long-horizon work) and Hydrogen (fast everyday execution), with a 1 million-token context window.
  • Pricing: $3 per million input tokens, $0.15 per million cached input tokens (a 95% discount) and $15 per million output tokens.
  • Google says Neon ranks #3 on the Artificial Analysis Intelligence Index (61, at xhigh), #2 on the Vals Index (70.9%) and #1 on its Coding Agent Index (58.6).
  • Google claims a 32% lower cost per task than Carbon at the default high level, and roughly 50% to 80% lower than top competing frontier models.
  • Rollout starts the same day in the Gemini API, Google AI Studio, Gemini for business, Google Antigravity and the Gemini app; a cyber variant follows for trusted defenders via the Fairwind Program.

Why it matters

Google is positioning Neon as a frontier-level model that is faster and cheaper than its larger sibling Carbon. The pitch rests on cost per task as much as raw scores: a 32% lower cost than Carbon at the default high level, and 36%+ lower than competing models at their highest setting. Google says that at xhigh Neon lands just behind larger, slower models on the Artificial Analysis Intelligence Index. Its four thinking levels let teams trade intelligence against token spend.

Who it affects

Developers building on the Gemini API in Google AI Studio and Android Studio, enterprises using Gemini for business, and Google AI Pro and Ultra subscribers in the Gemini app and Google Antigravity. Teams running long agent workflows, including orchestrator and subagent setups with Hydrogen in Antigravity, are a stated target. Security defenders who are eligible Fairwind Program members will get the cyber variant.

How to use it

Developers can call Neon through the paid tier of the Gemini API in Google AI Studio and Android Studio, at $3 per million input tokens, $0.15 per million cached input tokens and $15 per million output tokens. Paid Google AI Pro or Ultra subscribers can try it in Google Antigravity. Pro and Ultra subscribers can pick it from the model dropdown in the Gemini app. Enterprises can use it in Gemini for business via the API or model dropdown. Defenders who want the cyber variant should contact their Cloud account team about the Fairwind Program. Google says Neon's thinking levels let you tune cost: high is the default, and xhigh adds deeper multi-step reasoning for the hardest workloads.

How solid is it

The post is Google's own announcement, and every figure is Google's account of third-party and internal evaluations. The post names no specific competitor models in its comparisons, using phrases such as "comparable models" and "top competing frontier models". The #3 rank and score of 61 on the Artificial Analysis Intelligence Index apply to the xhigh level only. The 3.4% CWE-bench v2 uplift over Carbon is not specified as percentage points or relative. The availability wording is also inconsistent: one passage says all Pro and Ultra subscribers can try it in AI Studio, while another ties Antigravity access to paid Pro and Ultra subscribers.

Risks and caveats

Safety and alignment figures, such as the 2.45% shortcut rate, 0.09% misaligned action rate and 99.2% honesty pass rate, come from Google's own evaluations. The monitor-based alignment improvement is described by Google with hedged wording ("suggesting improved alignment propensity"). The post does not give prices for the cyber variant or for Gemini app subscriptions, and it does not state parameter count, weights release or knowledge cutoff. No regional availability is given. The cyber variant is restricted to trusted defenders, and access is still rolling out over the following days.

“While Carbon tackles your hardest, long-horizon projects and Hydrogen handles fast everyday execution, Neon balances depth and efficiency.”

— Google, Gemini 4 Neon announcement