Gemini 4 Neon positions itself between speed and reasoning

Google is rolling out Gemini 4 Neon in its Gemini app, API, and several professional tools. The model targets code, search, and complex tasks, with pricing and performance designed for broader use.

Google has launched Gemini 4 Neon, an AI model presented as suitable for complex coding, research uses, and scientific tasks. Available to Google AI Pro and Ultra subscribers, it already replaces Argon in the Gemini app and extends to the API, Google AI Studio, Android Studio, and Gemini for business. The positioning is clear: to offer a more versatile model than Hydrogen, without tipping into the heaviness of Carbon.

An intermediate model, but not lukewarm

Neon occupies an intermediate place in the Gemini 4 range. Google places it between Carbon, reserved for the most ambitious projects, and Hydrogen, designed for faster everyday assistance. In the Gemini interface, it appears under the "Deeper analysis" tab, while Hydrogen remains under "Everyday help." The message is clear: Neon targets uses that require reasoning, but without making the user wait like a poorly awakened server.

Technically, the model accepts a context window of 1 million tokens, with four levels of reasoning and a "high" mode enabled by default. On the API side, Google announces a price of 3 dollars per million input tokens, 15 dollars per million output tokens, and 0.15 dollar for cached inputs. These rates place it in a more accessible zone than the most expensive frontier models, while maintaining a high level of ambition.

What Google promises for use cases

In the Gemini app, Google presents Neon as a useful tool for writing, conducting research, and handling everyday tasks considered tedious: reading a stack of receipts, checking an electricity bill, or cross-referencing multiple files. The model can work on spreadsheets, PDFs, charts, audio, and video in a single request. This multi-format capability is as important as raw performance, as it brings AI closer to a work assistant rather than a simple text generator.

Neon also connects to Gemini tasks and can leverage contexts from Gmail and Docs. For Pro and Ultra subscribers, Google also integrates it into Google Antigravity, with an Orchestrator mode that combines Neon’s planning and Hydrogen’s fast execution. Verified companies can find it in the Vertex AI documentation for global, US, and European regions. Again, the deployment goes beyond a simple product test.

Scores that put Google in the race

Google claims that Neon outperforms Claude Opus 5.5 and GPT-6.2 Sol on the Artificial Analysis Intelligence Index, with a score of 61 versus 59 and 57. The model also leads on the Vals Index, at 70.9%, and on AutomationBench, at 56.3%. However, it remains behind Opus 5.5 on Terminal-Bench 4.1 and FrontierCode 2.0, as well as behind GPT-6.2 Sol on Chartography. These figures come from Google, so should be taken as an internal comparison baseline [to be verified].

At its default reasoning level, Google says Neon achieves a result close to Carbon, with a cost per task 32% lower and a cost 50% to 80% lower than leading frontier models. This is where the challenge becomes concrete for developers and companies: if these gaps are confirmed outside of in-house benchmarks, Neon could become an interesting compromise between quality, latency, and cost. A rarity in generative AI, where the bill often reminds us that it also knows how to reason very well.

Key points

  • Gemini 4 Neon is launched in the Gemini app, API, and pro tools.
  • Google places it between Carbon and Hydrogen in its range.
  • The context window reaches 1 million tokens.
  • The announced API price is 3 dollars per million inputs.
  • Google cites a score of 61 on the Artificial Analysis Intelligence Index.
  • Google AI Pro and Ultra subscribers have priority access.

By the numbers

  • 1 million tokens — context window announced by Google.
  • $3 / million tokens — API input price, according to Google.
  • $15 / million tokens — API output price, according to Google.
  • 61 — score shown on the Artificial Analysis Intelligence Index.
  • 70.9% — claimed score on the Vals Index.

To read