Gemini 4 Neon Released: A Balance Between Frontier Intelligence and Efficient Task Execution

Gemini 4 Neon is released, emphasizing a balance between frontier intelligence and cost efficiency, with standout performance in coding and agent scenarios.
Google DeepMind has released Gemini 4 Neon, positioned between the deep-reasoning Carbon and the lightweight Hydrogen, filling the gap for "high-frequency, complex, yet not too expensive" workloads. It ranks third on the Artificial Analysis Intelligence Index with a score of 61, and second on the Vals Index while sitting on the Pareto frontier. Pricing is $3 per million input tokens, $0.15 for cached input, and $15 for output, with a 1 million token context window; at the default high level, cost per task is 32% lower than Carbon. It performs outstandingly in coding and agent scenarios, ranking first on the Coding Agent Index, and reaching 39.0% on SWE-sweep in Antigravity. Safety alignment data is impressive, with a refusal rate of 1.98% and a cheating rate of 0.0%. There is also a Neon for cyber version for trusted defenders.
Google DeepMind has officially released Gemini 4 Neon, a model positioned between "frontier intelligence + fast end-to-end task completion." Unlike Carbon, which focuses on ultra-long-horizon complex projects in the same family, or Hydrogen, which emphasizes fast everyday execution, Neon attempts to find a new balance between deep reasoning and cost efficiency, covering complex software engineering, knowledge work, and scientific workflows.

1. Performance Positioning: Breaking into the Top Tier, but Focused on "Fast Enough and Cheap Enough"
According to Koray Kavukcuoglu, Senior Vice President and Chief AI Architect at Google DeepMind, Neon scores 61 on the Artificial Analysis Intelligence Index at the xhigh thinking level, ranking third overall, behind only larger and slower models. On the Vals Index it ranks second (70.9%) and sits on the Pareto frontier—meaning that on the "intelligence vs. cost" trade-off curve, it has no obvious positional disadvantage.
Notably, Neon performs well on "steerability." On the new Agent Arena Index, it ranks second and can adapt well to mid-task user corrections. In Google's internal Antigravity tests, Neon also set Google's lowest fabrication rate to date and performed outstandingly on multi-step instruction following.
On pricing, Neon costs $3 per million input tokens (cached input at $0.15 per million tokens, a 95% discount) and $15 per million output tokens, with a 1 million token context window.
Pareto frontier is a core concept in multi-objective optimization: when no solution can improve one metric without sacrificing another, the set of such solutions forms the Pareto frontier. In the context of AI model evaluation, the horizontal axis is usually cost per task and the vertical axis is intelligence score; sitting on the frontier means the model has the highest intelligence for a given cost, or the lowest cost for a given intelligence—no other model can dominate it on both dimensions simultaneously. The article's claim that Neon "sits on the Pareto frontier" emphasizes that it has no obvious positional disadvantage, rather than simply saying it has the highest score on any single metric.
Artificial Analysis Intelligence Index is a composite intelligence index maintained by the independent evaluation organization Artificial Analysis. It aggregates multiple highly difficult benchmarks (reasoning, coding, math, knowledge, etc.) into a single score for easy cross-model comparison. The so-called "xhigh thinking level" refers to the model's configurable internal reasoning token budget tier; the higher the level, the more explicit thinking the model does before giving a final answer, usually yielding higher scores but also higher latency and cost. Therefore, seeing a score of 61 is only meaningful when compared with other models at the same thinking level.
2. Enterprise and Scientific Scenarios: From a Spreadsheet to a Laboratory
One of Neon's core selling points is "synthesizing information scattered across spreadsheets, PDFs, charts, audio, and video into structured deliverables within a single 1 million token context."
In independent evaluations, Neon ranks second on the Vals Index (70.9%) and tops both Legal Research Bench and Tax Agent Bench. In Surge AI's evaluations, it ranks first in multiple categories including finance, accounting, spreadsheet modeling, and complex constraint following; it scores 73.5% on Chartography and 75.8% on Agents' Last Exam v1.1, on par with leading models at the same level. It also takes overall first place on Zapier's AutomationBench (56.3%) and Andon Labs' Vending-Bench 2 ($16,843, a long-horizon business simulation), and ranks second globally with 1336 points on Arena's new Agent Arena Index, which covers more than 4,200 real agent sessions.
In scientific reasoning, Neon leads on Terminal-Bench Science (70.5%), RiemannBench (84.0%), and HLE-Diamond (71.3%).
3. Production-Grade Code and Agent Workflows
Neon is explicitly designed for long-running agent workflows. It can run independently or serve as a "high-speed judgment-oriented Orchestrator" in Google Antigravity, directing parallel Hydrogen sub-agents. Google says it reduces redundant file reads, shares progress while working, and reliably handles context compaction in long sessions.
In core developer workflows, Neon ranks first 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 under the default high setting. For production-grade code, it can perform "surgical" targeted debugging and refactoring in multi-file repositories. On SWE-sweep, it scores 12.8% under the mini-SWE harness, but jumps to a current best of 39.0% in Google Antigravity—which Google uses to illustrate the role of the harness itself in unlocking model capabilities.
Google also provides several demonstrative cases: using /swarm in Antigravity to drive Blender MCP to build a realistic 3D laboratory bench with visualized model naming; writing from scratch a first-person WebGL2 ski simulator based on real elevation terrain around Mont Blanc and Chamonix, including tens of thousands of lines of custom shaders and 120 Hz ski physics; using /swarm to pull archival photos and interactively reconstruct in 3D Turing's study at King's College, Cambridge in 1936, even making the universal Turing machine on the desk actually run; and converting a single image into a Google Sheet with 100,000 cells and more than 500,000 live formulas, completing edge detection, 3D depth mapping, and relighting.
Harness (evaluation framework/scaffolding) refers to the entire runtime environment built around a model, including tool-calling interfaces, file system operations, context management, retry and error-handling logic, and so on. The same model can score several times differently under different harnesses, because the harness determines whether the model can effectively use tools and maintain state across multiple turns. The article's example of SWE-sweep scoring only 12.8% under the mini-SWE harness but jumping to 39.0% in Google Antigravity illustrates that "model capability" and "the ability released by the engineering framework" are two independent variables—which is why Google emphasizes the role of the harness.
Orchestrator / sub-agent architecture is an organizational approach for long-horizon agent workflows: a master model is responsible for decomposing tasks, dispatching subtasks, and aggregating results, while multiple sub-agents execute specific work in parallel. Neon plays the role of a "high-speed judgment-oriented orchestrator," directing parallel Hydrogen sub-agents. The benefit is that the master model does not have to do all the heavy lifting itself, thereby reducing per-task cost and improving throughput. Accordingly, such architectures place high reliability demands on context compaction, progress synchronization, and error recovery; the article's mention that Neon "reliably handles context compaction in long sessions" addresses exactly this point.
4. Efficiency and Thinking Levels: Four Adjustable Tiers with Clear Cost Advantages
Neon offers four thinking levels, allowing teams to adjust intelligence level and token consumption according to task needs. Google says that at the default high level, its cost per task is 32% lower than Carbon's; on the Vals Index and Agent Arena it sits on the Pareto frontier, with cost per task about 50% to 80% lower than top competitors. The xhigh level targets the hardest engineering, analysis, and scientific workloads, with the highest intelligence score while still costing more than 36% less per task than competitors' highest settings.
5. Cybersecurity: From Writing Secure Code to Proactively Finding Vulnerabilities
Neon for cyber is a specialized version for trusted defenders. On AutoBaxBench, it generates safer code than Carbon at all three levels—easy, medium, and hard—reaching a current best of 76% on the hard level. On WebExploitBench v1.2, it achieves a current best of 51.26% at half the cost of competing large models. Across the entire code-hardening lifecycle, it shows frontier performance in vulnerability discovery (SEC-bench Pro 90.5%), exploitability validation (ExploitGym, using half the tokens of Carbon), and patch remediation (CWE-bench v2 70.7%, a 3.4% improvement over Carbon). This version will be made available to trusted defenders through the Fairwind Program.
Prompt injection is one of the most typical security threats in the agent era: attackers hide malicious instructions in content that the model will read (web pages, documents, emails, tool return results), inducing the model to deviate from its original task and execute the attacker's intent. Indirect prompt injection specifically refers to malicious content that is not directly input by the user but comes from external data passively obtained by the model during task processing, making it harder to defend against. Gray Swan's benchmark specifically measures the model's success rate under such attacks; 0.0% means that among the attack samples covered by the benchmark, the model was never compromised—but note that this does not mean it is immune to all real-world attacks.
CWE-bench, SEC-bench Pro, and ExploitGym correspond to different stages of the vulnerability lifecycle: CWE-bench measures the model's ability to fix known types of vulnerabilities (CWE stands for Common Weakness Enumeration), SEC-bench Pro measures the ability to discover vulnerabilities, and ExploitGym measures the ability to determine whether a vulnerability can actually be exploited. Together, the three form a "discover—validate—fix" loop, which is what the term "code-hardening lifecycle" means in the article.
6. Safety and Alignment: Fewer Refusals, Fewer Oversteps
Neon continues the Gemini 4 safety stack, maintaining Carbon-level severe abuse protection and prompt injection defense while continuing to advance model alignment and agent behavior.
Official data shows: Neon's refusal rate on legitimate tasks is 1.98% on Terminal-Bench 4.0 and 0% on Terminal-Bench-Science; in internal coding use, the alert rate for chain-of-thought and action misalignment monitoring is below Carbon's already very low baseline; the cheating rate on Terminal-Bench 4.0 is 0.0%, and only 2.45% of tasks took shortcuts in coding cheating evaluations, the lowest in the Gemini 4 series; the misaligned action rate on the agent safety suite is 0.09%, and the agent honesty evaluation pass rate is 99.2%. In agent behavior evaluations, Neon reduces "underserving" by 43% to 56% and "overstepping" by 15% to 35%. On Gray Swan's indirect prompt injection benchmark, with real-time protection at the highest thinking level, Neon resisted all attacks within that benchmark, achieving the same 0.0% attack success rate as other Gemini models.
7. Where to Use It
Gemini 4 Neon is available starting today on the paid tier of the Gemini API (Google AI Studio, Android Studio); Google AI Pro and Ultra subscribers can try it in Google Antigravity; enterprise users can access it through the Gemini for business API or model dropdown menu; consumers can select Neon from the dropdown menu in the Gemini app. Cybersecurity defenders need to apply for access through the Fairwind Program.
In terms of positioning, Neon is not meant to replace Carbon's depth or compete with Hydrogen's lightness, but to fill the gap for "high-frequency, complex, yet not too expensive" workloads. For teams that need to run agent pipelines over the long term, this may be a more realistic choice.
One-line Summary
Gemini 4 Neon is released, emphasizing a balance between frontier intelligence and cost efficiency, with standout performance in coding and agent scenarios.
Paragraph Summary
Google DeepMind has released Gemini 4 Neon, positioned between the deep-reasoning Carbon and the lightweight Hydrogen, filling the gap for "high-frequency, complex, yet not too expensive" workloads. It ranks third on the Artificial Analysis Intelligence Index with a score of 61, and second on the Vals Index while sitting on the Pareto frontier. Pricing is $3 per million input tokens, $0.15 for cached input, and $15 for output, with a 1 million token context window; at the default high level, cost per task is 32% lower than Carbon. It performs outstandingly in coding and agent scenarios, ranking first on the Coding Agent Index, and reaching 39.0% on SWE-sweep in Antigravity. Safety alignment data is impressive, with a refusal rate of 1.98% and a cheating rate of 0.0%. There is also a Neon for cyber version for trusted defenders.
Related articles

Is AI the Culprit Behind Global Outages? The Trust Crisis Behind the NSA and FAA Failures
Fourth major failure in five days hits NSA classified network. Microsoft blames a software bug, FAA denies its new AI system, NSA says routine maintenance. Senator Warner suggests AI accelerates vulnerability discovery, leading to rushed fixes — while 39% of Americans already view AI negatively, exposing a deepening trust crisis.

Robot Developer Builds Custom B-rep Kernel with 3,487 Tests
A robot developer builds a custom B-rep geometry kernel with 3,487 tests, exploring why robotics needs dedicated CAD tools and the barriers to building one.

Russia Reportedly Planned Mass-Casualty Attack in Poland to Be Blamed on Ukraine
Yahoo News reports Russia allegedly planned a mass-casualty attack in Poland to blame Ukraine. This article covers the report's key points, Poland's strategic position, the intent behind the false flag operation, and its potential impact on the region, while analyzing the allegation's credibility and discussion traction.