Google releases Gemini 4 Neon for coding and analysis at $3/$15 API rates
The rollout spans developer tools, business products and the Gemini app. Google’s efficiency claims depend on reasoning settings and the software running the model.
Loading page…
The rollout spans developer tools, business products and the Gemini app. Google’s efficiency claims depend on reasoning settings and the software running the model.
Listen to this story
Google’s staged rollout of Gemini 4 Neon adds a model aimed at frequent, demanding coding and analysis work, positioned between the Gemini 4 family’s Carbon and Hydrogen models. Developers and businesses can access it through Google’s API and Antigravity, while eligible paid subscribers can select it in the Gemini app. Its one-million-token context and adjustable reasoning effort broaden its use, but the economics depend on settings: Google’s task-cost comparisons are benchmark claims, distinct from its published API rates.
At high thinking, Google claims Neon costs 32% less per task than Carbon on Artificial Analysis’s intelligence evaluation; this is not a discount to its API rates.
At xhigh, Google reports a 61-point, third-place Artificial Analysis Intelligence Index result and a first-place Coding Agent Index score of 58.6.
Neon accepts text, images, video, audio, and PDFs, but does not generate images or audio.
Google began rolling out Gemini 4 Neon on November 18, 2026, giving developers, businesses and paid Gemini subscribers another model for complex coding and analysis. Its launch pitch centers on completing demanding work faster and more efficiently—not just answering harder questions—with API pricing set at $3 per million input tokens and $15 per million output tokens.
Google positions Neon between two existing roles in its Gemini 4 family: Carbon handles the hardest, long-running projects, while Hydrogen targets fast everyday execution. Neon is intended for complex work repeated frequently, including software engineering, analytical reasoning and business workflows. The company is distributing it across consumer and professional products in the same staged release.
Neon has a 1-million-token context window for working across documents and media. It accepts text, images, video, audio and PDFs. The Decoder also lists computer use, code execution, file search and grounding through Google Search and Maps among its supported tools. Those input capabilities do not make it a media generator: image and audio generation are not supported.
Google gives Neon four thinking levels, letting teams adjust reasoning effort and token spending to the job. High is the default; xhigh adds deeper, multi-step reasoning for harder engineering, analytical and scientific tasks. That distinction runs through the launch benchmarks: Google’s highest cited intelligence score uses xhigh, while several cost comparisons use high.
At high, Google claims a 32% lower cost per task than Carbon on Artificial Analysis’s intelligence evaluation. It also cites roughly 50%–80% lower task costs than top competing frontier models on Vals and Agent Arena. These are Google-presented benchmark comparisons, not a blanket discount on API rates; its published pricing separately includes a reduced rate for cached inputs.
Google’s standard input rate.
A 95% discount from the standard input rate.
Google’s published output rate.
For capability, Google reports a score of 61 and third place overall on the Artificial Analysis Intelligence Index at xhigh. It also places Neon first on the Artificial Analysis Coding Agent Index, with 58.6. Those results support Google’s positioning of Neon as a capable working model, but they measure different tasks rather than one universal level of performance.
The surrounding software also figures prominently in Google’s coding pitch. Google reports a SWE-sweep score of 12.8% with the mini-SWE harness, versus 39.0% in Antigravity. A harness is the software setup running the model through a task. Google says Neon can work alone or direct parallel Hydrogen agents in Antigravity, while reducing redundant file reads and handling long sessions.
For long-running work, Google says Neon retains Carbon-level protections against severe misuse and malicious instructions hidden in outside content. It also deploys a monitor for potentially misaligned reasoning and actions across external products. In internal evaluations, Google reports a 99.2% pass rate on agent honesty and shortcuts on 2.45% of tasks in its coding cheating test. These are company evaluation results, not guarantees for every workflow.
The next announced access milestone is the specialized cyber version. Google says eligible defenders in its Fairwind Program will begin receiving Neon for cyber over the coming days. Organizations seeking that access are directed to their Cloud account teams, keeping this security-focused rollout on a separate path from the general model.
Loading discussion...
Join the conversation
Explain which comparison would help you choose more confidently.
Be the first to share a perspective or an experience.
Reader comments
Newest comments first. Replies stay oldest first.