ByteBulletin

[models] · · 4 min read

Google launches Gemini 4 Neon for agentic coding workflows

The new model targets the middle ground between heavy reasoning and fast execution, claiming state-of-the-art results in secure code generation and enterprise automation at a lower cost per task.

By ByteBulletin Editor · Editor

Google launches Gemini 4 Neon for agentic coding workflows

AI-generated illustration · Z-Image-Turbo, self-hosted


Google has released Gemini 4 Neon, a new frontier model designed to balance high-level reasoning with fast, cost-efficient execution for complex software engineering and knowledge work. According to the official announcement on the Google blog, the model is positioned as a versatile workhorse that sits between the heavy-duty Gemini 4 Carbon and the lightweight Gemini 4 Hydrogen, aiming to handle long-running agent workflows without the latency or expense of larger models.

The launch marks a significant shift in Google’s strategy for agentic coding, emphasizing not just raw intelligence but also steerability and reliability in multi-step tasks. Google states that Neon ranks #3 overall on the Artificial Analysis Intelligence Index with a score of 61 at its xhigh thinking level, landing just behind larger, slower models. More notably for developers, it ranks #1 on the Artificial Analysis Coding Agent Index with a score of 58.6, signaling a specific optimization for code generation and debugging tasks.

Performance and Benchmarks

Google’s release notes highlight a suite of benchmark results that position Neon as a top-tier competitor in both general intelligence and specialized domains. On the Vals Index, Neon ranks #2 with a score of 70.9%, placing it on the Pareto frontier for efficiency versus capability. In enterprise-specific evaluations, the model achieves #1 on both the Legal Research Bench and Tax Agent Bench, while also scoring 73.5% on Chartography for reading dense professional charts.

For scientific and mathematical reasoning, Neon reports best-in-class results on Terminal-Bench Science (70.5%), RiemannBench (84.0%), and HLE-Diamond (71.3%). In the realm of secure coding, the model claims state-of-the-art performance on the hard tier of AutoBaxBench (76%) and leads on WebExploitBench v1.2 with a score of 51.26%, achieving these results at half the cost of competing large models. Google also notes that Neon records its lowest fabrication rate yet in internal testing, a critical metric for agents that must follow complex instructions without hallucinating capabilities or data.

Pricing and Availability

Gemini 4 Neon is priced at $3 per million input tokens and $15 per million output tokens, with a 1 million-token context window. A key feature for high-volume workloads is the 95% discount on cached input tokens, priced at $0.15 per million. This pricing structure is designed to make long-context, iterative agent workflows economically viable for enterprise users.

The model is rolling out immediately across several platforms:

  • Gemini API in Google AI Studio.
  • Gemini for business via API and model dropdown.
  • Google Antigravity for paid Google AI Pro and Ultra subscribers.
  • Gemini app for Pro and Ultra subscribers.

A specialized variant, Gemini 4 Neon for cyber, is also being rolled out to trusted defenders under the Fairwind Program, focusing on vulnerability finding, validation, and patching.

Developer Workflows and Capabilities

Google emphasizes that Neon is engineered for long-running agent workflows, sustaining performance through complex, tool-heavy tasks. In Google Antigravity, the model can act as an Orchestrator directing parallel Hydrogen subagents, reducing redundant file reads and handling context compaction reliably in long sessions. This architecture is designed to improve the efficiency of multi-agent systems where a single model might struggle with context limits or cost.

The release notes include several demonstration cases that illustrate the model’s capabilities in creative and technical domains:

  • 3D Visualization: Using the Blender MCP in Google Antigravity, Neon built a photorealistic 3D lab bench visualizing model names with elemental accuracy.
  • Game Development: The model wrote a first-person WebGL2 freeride ski simulator from scratch, featuring 10,200 lines of custom shaders and 120 Hz ski physics across six alpine lines.
  • Historical Reconstruction: An agent swarm recreated Alan Turing’s 1936 study at King’s College, Cambridge in interactive 3D, including a working Universal Turing Machine.
  • Data Processing: Neon turned a single image into a 100,000-cell Google Sheet using over 500,000 live spreadsheet formulas to run edge detection and map 3D depth.

These examples highlight the model’s ability to synthesize across multiple input types—text, code, images, and data—within a single context window to produce structured deliverables.

What It Means for Developers

For developers, Gemini 4 Neon offers a compelling alternative to larger, slower models for tasks that require high judgment but do not need the maximum reasoning depth of Carbon. The 1 million-token context window and aggressive caching discounts make it suitable for processing large codebases or document sets without breaking the budget. The model’s focus on secure code generation and low fabrication rates is particularly relevant for teams deploying agents in production environments where reliability and safety are paramount.

However, developers should be aware that Neon is positioned as a middle-ground option. For the most complex, long-horizon projects, Google still recommends Carbon, while Hydrogen remains the choice for fast, everyday execution. The choice between these models will depend on the specific trade-offs between latency, cost, and reasoning depth required for each task.

What to Watch

  • Real-World Agent Performance: While benchmarks are strong, the true test will be how Neon performs in sustained, multi-agent workflows in production environments.
  • Cyber Variant Rollout: The expansion of the Fairwind Program and the availability of the cyber-specific variant will be key indicators of Google’s commitment to enterprise security.
  • Competitor Responses: How other providers adjust their pricing and model offerings in response to Neon’s aggressive cost-per-task positioning.
  • Long-Term Reliability: Monitoring for any issues with context compaction or steerability in extended sessions, as these are critical for agentic reliability.

Get the signal, not the noise.

One short email when it matters. No recaps of recaps.

SHARE

← All stories