TL;DR
Google’s Gemini 4 Neon targets demanding agent workflows with lower measured task cost than its larger sibling Carbon. Its reported coding scores depend on the harness, and production economics require more than a token-price comparison.
Introduction
Google positions Gemini 4 Neon between Carbon for harder, longer projects and Hydrogen for faster everyday work. That placement matters for agents: a coding session may combine complex planning with many routine tool calls, and neither the strongest model nor the cheapest model is necessarily optimal for every step. (Source : Google — Introducing Gemini 4 Neon)
The launch extends the model-routing question explored in our analysis of Gemini 4 Carbon and agent economics. The interesting question is not simply whether Neon leads a leaderboard, but how much useful work it completes for the total cost of an agent run.
The token price is only the starting point
Google lists Neon at $3 per million input tokens, $0.15 per million cached input tokens and $15 per million output tokens. The announced context window is one million tokens. (Source : Google — Introducing Gemini 4 Neon)
For illustration, a run billed for 800,000 uncached input tokens, 200,000 cached input tokens and 100,000 output tokens would cost $3.93 at those rates. This is a calculation, not a measured workload or a quote: it excludes cache storage or write charges if applicable, tooling, infrastructure and human review. The output component alone accounts for $1.50, so a model that generates long reasoning traces or repeated patches may cost more than its headline input price suggests.
Google says Neon costs 32% less per task than Carbon at the default high reasoning setting in Artificial Analysis Intelligence Index evaluations. This is a benchmark-specific comparison of API cost per task, not proof of a 32% reduction in every production bill. (Source : InfoWorld — Google launches Gemini 4 Neon)
Artificial Analysis says its coding-agent cost metric includes applicable input, cached-input and output prices but not the full operational cost of production, such as engineering and supervision. A team should therefore compare cost per accepted change, including failed attempts and review time, rather than cost per generated answer alone. (Source : Artificial Analysis — Coding agent benchmarks)
Why the harness changes the result
Google reports Neon at 58.6 on the Artificial Analysis Coding Agent Index and 65.2% on Terminal-Bench 4.1. It also reports 12.8% on SWE-sweep using the mini-SWE harness versus 39.0% using Google Antigravity. (Source : Google — Built for production-ready code)
The same model’s reported result changes substantially across those SWE-sweep setups. That does not isolate the cause: prompts, tools, settings and execution environment can all matter. It does show why treating a score as a model-only property is risky. The independent Artificial Analysis index is itself a composite of three benchmarks, equally weighted, with scores averaged over three attempts per task. (Source : Artificial Analysis — Coding agent methodology)
For a purchasing decision, reproduce the comparison on a fixed set of real issues. Hold the repository, test suite, retry policy and approval rules constant; then report task completion, elapsed time, token mix, regressions and reviewer effort. Otherwise a model switch may actually be a harness switch in disguise.
Planner and executor: a useful split with overhead
Google says Antigravity’s Orchestrator can use Neon for planning and complex reasoning while delegating routine edits, package updates and test runs to Hydrogen. When the executor encounters difficulty, it can hand the task back to Neon. (Source : Google Antigravity — Gemini 4 Neon)
The design makes sense when expensive judgment is concentrated at decision points. But delegation also adds context transfer, coordination and potentially conflicting edits. Multiple agents are not automatically cheaper: compare the combined token use and failure recovery of the full workflow with a single-agent baseline. InfoWorld notes that analysts see human oversight, regression testing and coordination as continuing production challenges. (Source : InfoWorld — Google launches Gemini 4 Neon)
Safety claims are bounded by the test
Google reports that Neon, with real-time safeguards at its highest thinking level, resisted every attack in its Gray Swan indirect prompt-injection benchmark. That is a result for a specific benchmark and configuration, not a guarantee against malicious instructions in arbitrary documents or repositories. Restrict tool permissions, protect secrets and require human approval for consequential actions even if the model scores well. (Source : Google — Safe, dependable execution)
FAQ
Is Neon a replacement for Carbon and Hydrogen?
Google positions it between them; whether it should replace either depends on measured quality and total task cost for the specific workflow.
Does a one-million-token window make context management unnecessary?
No. A larger window does not make repeated input free, and relevant information still needs to be selected and verified.
Does the SWE-sweep comparison prove Antigravity is always better?
No. It shows different results in the reported configurations, not which component caused the difference or whether it generalizes to another codebase.
Where is Neon available?
Google says it is rolling out through the paid Gemini API, Gemini for business, and selected paid consumer and Antigravity offerings; access depends on the product and plan. (Source : Google — Start building)
Further Reading
- Google — Introducing Gemini 4 Neon
- Google Antigravity — Neon and Orchestrator
- Artificial Analysis — Coding agent benchmarks and methodology
- InfoWorld — Enterprise coding and deployment considerations
— The Agent Report