Between depth and speed, Gemini 4 Neon targets complex work
Designed for coding, analysis, and science, Gemini 4 Neon combines four reasoning levels with a one-million-token context window.
Between the long-running projects assigned to Carbon and the everyday tasks intended for Hydrogen, a new position is opening up in the Gemini 4 family. Adapted from Carbon, the Neon model targets complex, frequently performed work: software development, document analysis, and scientific research. In its announcement, Google emphasizes the time needed to complete an entire task, with settings that adjust the effort devoted to reasoning. Adjusting Reasoning Effort to the Task
Four levels are available: low, medium, high, and xhigh, with high as the default. Inputs can combine text, images, audio, video, and documents within a one-million-token context window. This capacity supports cross-referencing files and producing reports, summaries, or code. The rollout begins in Gemini and Antigravity for Google AI Pro and Ultra subscribers, in Gemini for business, and through the Gemini API’s paid tier, accessible from AI Studio and Android Studio. What the Comparisons With Carbon Measure
External evaluations provide some reference points for this positioning. In its launch analysis, Artificial Analysis records an Intelligence Index score of 61 for Neon at xhigh, compared with 60 for Carbon at high. Under that evaluation protocol, Neon takes approximately 21 minutes per task versus 36 for Carbon, at an average cost of $6.44 versus $7.76. The calculation uses Carbon’s standard pricing. Meanwhile, Vals AI ranks Neon second on its index with 70.87%, and first on its finance and legal indexes. These results describe different configurations and test suites.
API pricing is set at $3 per million input tokens, $15 per million output tokens, and $0.15 per million cached input tokens. Comparing it with Carbon requires distinguishing Carbon’s standard input and output prices of $4 and $20 from its promotional prices of $2 and $10, announced through November 30. Neon’s tokens therefore cost less than Carbon’s at standard pricing, but more during the promotion. The final cost of a task also depends on the volume generated: Artificial Analysis observes that Neon uses more output tokens in its xhigh comparison. Agents at Work in Antigravity
For software development, the emphasis is on long sessions and tasks involving multiple tools. In Antigravity, the model can direct Hydrogen subagents in parallel, respond to user corrections, and continue working after accumulated context has been condensed. The published results also show the importance of the execution environment: on SWE-sweep, Google reports 12.8% with mini-SWE and 39% in Antigravity. The model card distinguishes these configurations; the second figure combines the model’s capabilities with those of the tools and their orchestration.
Reliability is another focus of the announcement. Google reports its lowest rate of fabricated results to date in internal Antigravity testing and emphasizes adherence to multistep instructions. These evaluations examine agent behavior, shortcuts, and actions; they do not measure a general accuracy rate applicable to every response. A cybersecurity version is also scheduled to become available in the coming days to eligible defenders in the Fairwind Program, supporting vulnerability detection, validation, and patching.
Creative demonstrations illustrate this execution across multiple tools. By controlling Blender through MCP and the /swarm command in Antigravity, the model built a 3D laboratory bench connecting members of the model family with the properties of their namesake elements. Another demonstration reconstructs Alan Turing’s study in Cambridge using archival photographs. These examples presented by Google show a sequence of research, programming, and scene construction, with external software producing the visual output.