Google DeepMind Releases Gemini 4 Neon with High Efficiency and Agentic Capabilities
Google DeepMind has launched Gemini 4 Neon, a model balancing frontier reasoning with lower token costs across software engineering, analysis, and cybersecurity.

Google DeepMind has expanded its model family with the launch of Gemini 4 Neon, a model designed to deliver frontier-level intelligence with fast end-to-end task execution across software engineering, enterprise analysis, and scientific research. Positioned between the high-capacity Gemini 4 Carbon and the lightweight Hydrogen, Neon provides a balance of deep reasoning capability and reduced operational costs for production workflows.
The model features a 1 million-token context window and introduces four distinct thinking levels, enabling engineering teams to fine-tune intelligence and token spend based on specific workload needs.
Performance and Cost-Efficiency Across Workflows
Evaluations on public benchmarks highlight Neon's capabilities relative to existing frontier models. On the Artificial Analysis Intelligence Index, Neon scored 61 at its "xhigh" thinking setting, placing third overall and ranking just behind larger, slower systems. It secured second place on the Vals Index with a 70.9% score and achieved top rankings on specialized evaluations, including the Legal Research Bench, Tax Agent Bench, and Zapier's AutomationBench (56.3%). In scientific reasoning, the model recorded scores of 70.5% on Terminal-Bench Science, 84.0% on RiemannBench, and 71.3% on HLE-Diamond.
Neon's architecture emphasizes resource efficiency without sacrificing accuracy. Pricing for the model is set at $3 per million input tokens—with cached input tokens discounted by 95% to $0.15 per million—and $15 per million output tokens. At its default "high" thinking level, Neon delivers a 32% lower cost per task compared to Gemini 4 Carbon, while offering a 50% to 80% cost reduction per task relative to competing top-tier frontier models on the Vals Index and Agent Arena. At the maximum "xhigh" setting, DeepMind reports that Neon maintains a cost advantage of more than 36% per task against rival models running at their highest settings.
Coding Capabilities and Cybersecurity Protections
In software engineering tasks, Neon is engineered to operate as an orchestrator directing subagents or running standalone across long-running developer workflows. It ranked first on the Artificial Analysis Coding Agent Index with a score of 58.6, achieved 65.2% on Terminal-Bench 4.1, and registered 46.2% on FrontierCode 2.0 at its default high setting. When evaluated on repository-scale multi-file editing and debugging using SWE-sweep, Neon achieved 12.8% on the mini-SWE harness, which expanded to a state-of-the-art 39.0% within the Google Antigravity harness.
DeepMind has also tuned Neon for defensive and offensive cybersecurity applications through its "Neon for cyber" implementation. The model achieved a 76% score on the hard tier of AutoBaxBench, 90.5% on SEC-bench Pro for vulnerability identification, and 70.7% on Collinear's CWE-bench v2 for vulnerability patching. On WebExploitBench v1.2, it recorded a 51.26% success rate at half the token cost of larger competing models. To defend against prompt injection attacks, Neon resisted 100% of exploit attempts on the Gray Swan indirect prompt injection (IPI) benchmark when combined with real-time safeguards.
What it means for developers
For development teams building agentic software and enterprise integrations, Gemini 4 Neon offers a flexible mid-tier alternative that cuts operational costs while maintaining strong context handling. Its ability to act as an orchestrator directing parallel Hydrogen subagents in Google Antigravity reduces redundant file reads and manages context compaction during long coding sessions.
Developers can access Gemini 4 Neon starting today through the Gemini API in Google AI Studio and Android Studio, as well as via Google Antigravity for paid Google AI Pro and Ultra subscribers. Enterprise access is available through Gemini for business API endpoints. For engineering teams comparing multiple model APIs, developers can try top AI models cheaply through one API at https://apixoai.online.
Safety metrics indicate that Neon achieves lower rates of fabrication compared to previous releases. Internal Google Antigravity testing recorded its lowest fabrication rate yet, alongside a 2.45% shortcut rate on coding evaluations and a 0.09% misaligned action rate on agentic safety suites. Task refusal rates for legitimate technical requests remained low, refusing only 1.98% of tasks on Terminal-Bench 4.0 and 0% on Terminal-Bench Science.
Source: Introducing Gemini 4 Neon — Google DeepMind. Written by the Apixo team from that report.
One key for Claude, GPT, GLM, DeepSeek and more. Pay per token with crypto.
Get your API key

