Google DeepMind introduced Gemini 4 Neon, a fast frontier model for long software and enterprise jobs with a cyber variant built to write secure code, find flaws and test exploits.
Can a cheaper, faster model hold up on security work? Google DeepMind says yes, and priced it to prove it.
Neon costs $3 per million input tokens and $15 per million output tokens, with cached input at $0.15, according to Google DeepMind. It ships with a 1 million-token context window. That window lets it read spreadsheets, PDFs, charts, audio and video together and return structured work.
Security teams are the ones who have to change a workflow here. Neon for cyber scored 76% on the hard tier of AutoBaxBench, a test of generating secure code from the start, which Google DeepMind called state of the art. It hit 51.26% on WebExploitBench v1.2 for penetration testing at half the cost of large competing models, per the company.
The hardening numbers are tighter. Neon reached 90.5% on SEC-bench Pro for finding vulnerabilities at roughly half the cost per challenge, according to Google DeepMind. It matched frontier performance on ExploitGym while using half the tokens of Carbon. It reached 70.7% on Collinear’s CWE-bench v2 for patching, a 3.4% lift over Carbon.
What does that mean in production? A defender can run find, validate and patch in one faster loop instead of splitting cheap triage from expensive review. Google Cloud documentation now lists Neon in the Enterprise Agent Platform as a standalone worker for multi-file engineering or as an orchestrator directing parallel Hydrogen subagents, which tells platform teams where Google DeepMind wants it deployed.
General coding still matters because insecure helpers ship bugs. Neon ranked No. 1 on the Artificial Analysis Coding Agent Index at 58.6, with 65.2% on Terminal-Bench 4.1 and 46.2% on FrontierCode 2.0 at default high thinking, per Google DeepMind. On SWE-sweep it jumped from 12.8% on a mini harness to 39.0% inside Google Antigravity.
So what stops it from going rogue with tools and shell access? Prompt injection, which is malicious instructions hidden in web pages, emails, documents or code an agent reads, is the sharp edge. Google DeepMind said Neon with real-time safeguards at its highest thinking level resisted every attack in Gray Swan’s IPI benchmark, a 0.0% attack success rate matched only by other Gemini models.
There is friction. Neon for cyber is not open yet. It rolls to trusted defenders in the coming days through the Fairwind Program, according to Google DeepMind. The best intelligence score needs xhigh thinking, which costs more per task than high, though Google DeepMind said it still runs 36% cheaper than rivals at top settings.
Watch whether Fairwind teams keep Neon for WebExploit and patch loops or kick hard targets back to Carbon.





