· via dev.to (home feed)
Google DeepMind launches Gemini 4 Neon, a $3 frontier model undercutting rivals on cost
Google DeepMind's Gemini 4 Neon brings frontier-tier reasoning to $3 per million input tokens, with Google claiming 50-80% lower cost per task than rival frontier models in some settings.

Google DeepMind has started rolling out Gemini 4 Neon, a lower-cost frontier AI model priced at $3 per million input tokens, in a move meant to make high-end reasoning affordable for a far broader set of developers, businesses and researchers. According to a Fathom News report published on dev.to, the model is available now through the Gemini API, Gemini for business, Google Antigravity and the Gemini app, with a cybersecurity-focused variant arriving for selected defenders within days.
Koray Kavukcuoglu, senior vice president at Google DeepMind and Google's chief AI architect, introduced Neon as a model aimed at demanding software engineering, knowledge work and scientific workflows.
Where Neon sits in the Gemini 4 family
Google places Neon between two existing siblings: Gemini 4 Carbon, built for the hardest long-horizon projects, and Gemini 4 Hydrogen, tuned for fast everyday execution. Neon is positioned as the middle option, combining depth with efficiency for complex, high-frequency work.
The model offers a 1 million-token context window and four thinking levels, letting teams trade intelligence against token spend. At its xhigh thinking level, Neon scored 61 on an intelligence index, ranking third overall behind larger, slower models. It placed second on a benchmark for analytical and enterprise work and second on steerability, which Google described as how cleanly a model adapts when a user corrects it mid-task. Google also said Neon posted its lowest fabrication rate to date in internal Antigravity testing.
Neon is designed to handle the kinds of files that office and research work already produce. Google said it can synthesize spreadsheets, PDFs, charts, audio and video within a single context window and turn them into structured deliverables. In outside evaluations summarized by the company, Neon ranked first on legal research and tax work, led most finance, accounting, spreadsheet-modeling and constraint-following tests, scored 73.5% on dense chart reading and reached 75.8% on an agents exam.
Coding and agentic credentials
Coding is a central part of the pitch. Google said Neon is engineered for long-running agent workflows, reduces redundant file reads, shares progress as it works and handles context compaction reliably over extended sessions. On Google's reported numbers, Neon topped a coding agent index with 58.6, scored 65.2% on Terminal-Bench 4.1 and hit 46.2% on FrontierCode 2.0 at its default high setting. On a mini-SWE harness, it managed 12.8% on its own and 39.0% inside Google Antigravity.
Pricing and availability
Input tokens cost $3 per million, output tokens $15 per million, and cached input tokens $0.15 per million, a discount Google put at 95%. That cached rate is notable for repeated workflows where the same background material is reused across many tasks.
Google also claims a sharp cost advantage per completed task. At the default high setting, Neon is said to be 32% cheaper per task than Carbon. At that same setting, Google said Neon sits on the Pareto frontier in two separate evaluations, with roughly 50% to 80% lower cost per task than leading competing frontier models. At xhigh, Neon still costs 36% or more less per task than competitors running at their highest setting.
Access, however, is tiered rather than universal:
- Developers: the paid tier of the Gemini API in Google AI Studio and Android Studio, plus AI Studio and Antigravity for eligible Google AI Pro and Ultra subscribers
- Enterprises: Gemini for business, via the API or model dropdown, and Google Antigravity
- Consumers: the model dropdown in the Gemini app for Google AI Pro and Ultra subscribers
- Cybersecurity defenders: eligible Fairwind Program members, with access beginning in the coming days
Cyber variant and safety claims
The cyber version is the next test. Google said it produces more secure code than Carbon across easy, medium and hard challenges, scoring 76% on the hard tier of one security benchmark, 51.26% on a web exploitation benchmark at half the cost of competing large models, 90.5% on vulnerability finding and 70.7% on patching, a 3.4% improvement over Carbon.
Safety claims accompany the agent push. Neon reportedly matches Carbon's defenses against severe misuse and prompt injection while improving alignment and agentic behavior. Google said it recorded 0.0% cheating on Terminal-Bench 4.0, a 0.09% misaligned action rate on its agentic safety suite and a 99.2% pass rate on an agent honesty evaluation, and that with real-time safeguards at its highest thinking level it resisted every attack in an indirect prompt injection benchmark. Google gave no broader date for the cyber rollout.
Why it matters
Frontier-level reasoning has been among the most expensive capabilities in AI, and Neon's pricing puts direct pressure on rival model providers, particularly for high-volume agentic and coding workloads where cost per task, not cost per token, decides adoption. The steep cached-input discount could especially change the economics of agents that repeatedly process the same documents. That said, most figures here come from Google's own announcements as relayed in a single report, and access is gated behind paid subscriptions and invitation-only programs, so the widening of access is real but relative. The Fairwind cyber rollout, and whether independent evaluations confirm the benchmark claims, will show how durable the price advantage is.
- #gemini
- #google-deepmind
- #llm
- #ai-pricing
- #enterprise-ai