The launch marks the fourth new Gemini 4 model released in just four weeks, a pace that underscores the intense competition in the AI sector. Neon is positioned strategically between Gemini 4 Hydrogen, which handles routine queries, and Gemini 4 Carbon, reserved for the most difficult, long-horizon projects. It is available immediately through the Gemini API in Google AI Studio, Gemini for business, and the consumer-facing Gemini app, as well as in Android Studio for developers. For enterprise customers, the model is documented for use in Vertex AI across global, US, and EU regions.
At its core, Neon is a multimodal model capable of processing text, images, video, audio, and PDFs within a context window of one million tokens. It supports advanced capabilities such as Computer Use, code execution, file search, and grounding via Google Search and Maps. However, it does not support image or audio generation. The model operates with four thinking levels, with “high” set as the default, allowing users to balance computational depth against speed.
The commercial implications of Neon are significant. Priced at $3 per million input tokens and $15 per million output tokens, it is more expensive than its predecessors Carbon and Argon, which started at $2/$10, but significantly cheaper than the top-tier alternatives. Cached input is offered at a steep discount of $0.15 per million tokens. This pricing structure suggests Google is targeting developers and businesses that require high contextual understanding and complex reasoning without incurring the costs associated with the largest, slowest models.

Google’s internal benchmarks place Neon in a strong position. On the Artificial Analysis Intelligence Index, Neon scores 61 at its “xhigh” thinking level, placing it third overall, just behind larger, slower models. It outperforms Anthropic’s Claude Opus 5.5 (score 59) and OpenAI’s GPT-6.2 Sol (score 57) on this specific metric. Furthermore, Neon ranks first on the Vals Index with a score of 70.9% and leads on the Legal Research Bench and Tax Agent Bench. However, the model is not universally dominant; it trails Opus 5.5 on Terminal-Bench 4.1 and FrontierCode 2.0, and falls behind GPT-6.2 Sol on Chartography.
A critical aspect of the launch involves the performance of coding agents. Google reported that Neon resolved 12.8% of the SWE-sweep coding benchmark when using a standard “mini-SWE” harness. However, when run inside Google’s Antigravity environment, that score jumped to 39.0%. This threefold improvement highlights a crucial distinction in AI evaluation: the performance of a coding agent is heavily influenced by the harness—the software framework that manages tools, context, and execution—rather than solely by the underlying model. Antigravity’s harness was co-trained with Gemini, and Neon has been updated to reduce unnecessary file reads and manage context more efficiently during long sessions.
This finding contrasts with results from other providers. In the same comparison, Claude Opus 5.5 saw its SWE-sweep score climb from 3.6% to 10.1% when moved from the mini-SWE harness to Claude Code. In contrast, GPT-6.2 Sol’s performance declined from 11.6% to 8.3% when moved from mini-SWE to Codex. These discrepancies suggest that the synergy between a model and its proprietary agent environment is a major determinant of real-world utility. Other models, such as those tested in the FrontierHarness benchmark, showed that running the same open model could cost up to five times more depending on the harness used, further complicating cost-performance comparisons.

Despite the strong benchmark numbers, limitations remain. Neon is not a substitute for the most powerful models in every scenario; it is designed for a specific niche of high-efficiency, complex work. The model’s ability to synthesize across diverse inputs like spreadsheets, PDFs, and audio in a single pass is a key differentiator for enterprise workflows, where data is rarely confined to text alone. Google claims that Neon records its lowest fabrication rate yet in internal testing, a critical factor for trust in business applications.
For users, the arrival of Neon means a more granular choice of tools. The naming scheme, which now includes Hydrogen, Neon, Argon, and Carbon, can be confusing, but the functional differences are clear: Hydrogen for speed, Neon for balanced intelligence and cost, and Carbon for maximum capability. As the AI market matures, the focus is shifting from raw benchmark dominance to practical usability, cost-efficiency, and the seamless integration of models into specific professional workflows. Neon represents Google’s bet that a well-tuned mid-tier model, paired with a robust agent harness, can deliver value that rivals the most expensive frontier systems.
The next confirmed development will be the full rollout of Neon to all paid subscribers and the release of independent third-party benchmarks that may verify or challenge Google’s internal assessments. Until then, Neon stands as a testament to the evolving strategy of AI labs: not just to build the smartest model, but to build the most useful one for the price.



