Gemini 4 Neon: Breakthrough AI Unlocks 1M Token Context

Gemini 4 Neon: A New Frontier in AI Intelligence

Google DeepMind has officially launched Gemini 4 Neon, a significant advancement in artificial intelligence designed for rapid, end-to-end task completion across demanding software engineering, knowledge work, and scientific domains. This new model demonstrates frontier-level intelligence, positioning itself as a top performer with remarkable efficiency. Gemini 4 Neon is engineered to handle complex, multi-step instructions and exhibits a notable reduction in fabrication rates, according to internal Google Antigravity testing.

The model achieves an impressive score of 61 on the Artificial Analysis Intelligence Index at its xhigh thinking level, placing it just behind larger, slower counterparts but ahead of comparable models. Furthermore, Neon secures the second position on the Vals Index, landing squarely on the Pareto frontier. Its steerability, a critical factor for adapting to mid-task user corrections, ranks second on the new Agent Arena Index, highlighting its user-friendly nature.

Architecture, Performance, and Unprecedented Context Window

Gemini 4 Neon’s architecture is optimized for both depth and speed, offering a massive 1 million-token context window. This expansive window allows the model to synthesize information from diverse inputs such as spreadsheets, PDFs, audio, and video in a single session, a capability crucial for complex enterprise and analytical workflows. Independent evaluations confirm Neon’s leading position, with top rankings on benchmarks like the Vals Index (70.9%) and achieving #1 on the Legal Research Bench and Tax Agent Bench.

Performance metrics are equally compelling. Neon ranks #1 overall on Zapier’s AutomationBench (56.3%) and demonstrates exceptional chart comprehension with a 73.5% score on Chartography. For scientific and mathematical reasoning, it leads in its class on Terminal-Bench Science (70.5%), RiemannBench (84.0%), and HLE-Diamond (71.3%). This robust performance is achieved while maintaining cost-efficiency, with pricing set at $3 per million input tokens and $15 per million output tokens, representing a significant discount for cached input tokens.

Engineering Trade-offs and Enhanced Safety Protocols

While Gemini 4 Neon pushes the boundaries of AI capability, Google DeepMind has focused on refining its thinking levels to provide finer control and a smoother progression for users. Neon offers four distinct levels, allowing teams to precisely manage intelligence and token expenditure based on workload requirements. This tiered approach ensures that users pay for the intelligence they need, offering high performance at a significantly lower cost per task compared to previous models like Carbon.

Safety and alignment remain paramount. Gemini 4 Neon builds upon the Gemini 4 safety and security stack, maintaining Carbon-level protections against severe misuse and prompt injection. It demonstrates improved alignment propensity, with alert rates from misalignment monitors remaining below Carbon’s already low baselines. The model is trained to avoid fabricating results or overstepping user boundaries, achieving a 0.0% cheating rate on Terminal-Bench 4.0 and a minimal 2.45% shortcut rate on coding evaluations.

Production Readiness and Cybersecurity Applications

Gemini 4 Neon is engineered for production-ready code and long-running agent workflows, sustaining performance through complex, tool-heavy tasks. It excels in multi-file debugging and refactoring, offering targeted edits across repositories. For cybersecurity, Gemini 4 Neon for cyber produces more secure code than Carbon, achieving state-of-the-art results on AutoBaxBench, particularly on hard challenges (76%).

Its capabilities extend to proactive threat detection. On WebExploitBench v1.2, Neon achieves state-of-the-art performance in penetration testing at half the cost of competing large models. It also fortifies existing code by finding, validating, and patching vulnerabilities efficiently. Neon achieves 90.5% on SEC-bench Pro for vulnerability finding and demonstrates frontier-level performance on ExploitGym for validating risk through exploitation. For patching, it reaches 70.7% on Collinear’s CWE-bench v2, a notable improvement over Carbon. Gemini 4 Neon for cyber is rolling out to trusted defenders via the Fairwind Program.

Gemini 4 Neon is now available through the Gemini API in Google AI Studio, Gemini for business, and for paid Google AI Pro and Ultra subscribers in Google Antigravity and the Gemini app. This release marks a significant step forward in making advanced AI accessible, efficient, and reliable for a wide range of professional applications.

Source: Google DeepMind Blog