Google's Gemini 4 Neon Beats Carbon on Coding Tasks While Costing 23% Less
Google's new middle-tier Gemini 4 Neon is live in Cursor and the Gemini API. On CursorBench it beats Carbon, Google's flagship, at a lower cost per task, which makes it a strong candidate for your default coding agent.
Takeaways
- Gemini 4 Neon is live in Cursor, scoring 63.7% on CursorBench 4.1 at its default effort.
- That beats Hydrogen's 57.3% and Carbon High's 61.5%, at a lower cost per task.
- It costs $3/$15 per million tokens, with cached input at $0.15. In Cursor it's billed outside plan usage.
- It ranks first on the Artificial Analysis Coding Agent Index and third on the Intelligence Index, behind Fable 5.2 and Astra.
- It trails Opus 5.5 slightly on Terminal-Bench 4.1 and FrontierCode 2.0. Extra High costs more than Carbon High.
- It's available in AI Studio, Antigravity, the Gemini app and Gemini for business. Google's announcement
Cursor has added gemini-4-neon to its model picker. At the default High effort level, Neon scores 63.7% on CursorBench 4.1, compared with 61.5% for Gemini 4 Carbon High and 57.3% for Gemini 4 Hydrogen. Cursor estimates that Neon High costs $6.54 per task, 23% less than Carbon High’s $8.52. Google released Neon across its own products in the same rollout, making it the fourth Gemini 4 model introduced in roughly a month.
Neon slots between speed and depth
Google positions its three general-purpose Gemini 4 tiers around different workloads. Carbon targets long, difficult projects, while Hydrogen favors low-latency everyday work. Neon occupies the middle tier. Google’s launch post says it can match or exceed Carbon’s overall quality while running three times as fast and costing 60% as much under Google’s testing.
Leaderboard results at launch
- Artificial Analysis Intelligence Index: 61 at the highest thinking level, placing third overall and one point above Carbon.
- Artificial Analysis Coding Agent Index: first overall with a score of 58.6.
- Vals Index: second overall at 70.9%, with first-place results on Legal Research Bench and Tax Agent Bench.
- Agent Arena: second globally with a score of 1,336 across more than 4,200 agent sessions. Neon also placed second for steerability, which measures how reliably a model follows corrections during a task.
- Science and mathematics: 70.5% on Terminal-Bench Science, 84.0% on RiemannBench and 71.3% on HLE-Diamond.
Anthropic’s Claude Fable 5.2 scores 65 on the Intelligence Index, while OpenAI’s top Astra model scores 62. Neon’s 61 therefore ranks third, with its strongest case resting on the combination of quality, latency and completed-task cost.
Completed-task cost favors High
Neon’s API list price is $3 per million input tokens and $15 per million output tokens, twice Hydrogen’s $1.50 and $7.50 rates. Carbon currently carries introductory pricing of $2 and $10, scheduled to rise to $4 and $20 after the promotion. Neon will then cost 25% less per token than Carbon.
Token efficiency drives the immediate savings in Cursor. Neon High uses an average of 88,596 tokens per CursorBench task, compared with 114,518 for Carbon High. CursorBench uses ambiguous, multi-file coding tasks derived from real Cursor sessions, although it remains Cursor’s own benchmark rather than a fully independent evaluation.
| Model and effort | Score | Estimated task cost | Average tokens |
|---|---|---|---|
| Neon Low | 45.5% | $1.26 | 19,580 |
| Neon Medium | 57.7% | $3.07 | 47,059 |
| Neon High, default | 63.7% | $6.54 | 88,596 |
| Neon Extra High | 65.4% | $10.67 | 136,273 |
| Carbon High | 61.5% | $8.52 | 114,518 |
Neon Extra High adds 1.7 percentage points over High while increasing estimated cost by 63%. Within Cursor’s results, High delivers the strongest balance for routine agentic coding. Actual costs will vary with input-output ratios, cache hits, retries and tool-call patterns.
The engineering choices developers will notice
Google’s public documents describe Neon’s lineage differently: the Neon model card says it was adapted from Carbon, while Google Cloud’s model page says it builds on Hydrogen with deeper reasoning. Google has not published the training recipe or model weights.
- Four reasoning levels: Low, Medium, High and Extra High let each request trade latency and token use for additional reasoning.
- One shared token budget: the 1 million-token limit covers input, internal reasoning, function calls and the final response. Gemini 3 models applied separate limits to those stages.
- Long-session tuning: Google says Neon performs fewer redundant file reads and handles context compaction reliably. Compaction summarizes older session history as the context limit approaches.
- Orchestrator mode: Google’s Antigravity IDE can use Neon to direct several lower-cost Hydrogen agents in parallel.
- Fixed sampling: temperature remains locked at 1.0, and the API ignores custom temperature values.
Agent models can edit files, execute shell commands and operate browsers, so benchmark shortcuts and fabricated completion claims can obscure whether a task was completed correctly. Google reports 0.0% cheating on Terminal-Bench 4.0 and shortcut behavior on 2.45% of tasks in its internal coding evaluation, the lowest rate among Gemini 4 models. Google also reports its lowest fabrication rate to date in internal Antigravity testing, without publishing a comparable external score.
Harnesses can reorder the leaderboard
- Terminal-Bench 4.1: Claude Opus 5.5 leads Neon by roughly one point; Neon scores 65.2%.
- FrontierCode 2.0: Opus 5.5 also leads by about one point; Neon scores 46.2%.
- Chartography: GPT-6.2 Sol scores higher on this benchmark for interpreting complex charts.
- SWE-sweep: Neon scores 12.8% with the generic mini-SWE harness and 39.0% inside Google’s Antigravity environment. The gap shows how strongly tools, prompts and agent scaffolding can affect results.
- Source mix: many launch figures come from Google, and broader independent testing remains limited.
Cursor cautions that small score differences may lack statistical significance. A one-point lead offers weak grounds for choosing a model without confidence intervals or repeated trials. Larger gaps, including Neon’s 65.4% versus Opus 5.5’s 58.1% on CursorBench, still require validation against the repositories, tools and task types used in production.
Run Neon in Cursor or through the API
Cursor users can select gemini-4-neon from the model picker. According to the Cursor documentation, sessions use a 200,000-token context window by default and can extend to 1 million tokens. The model supports Cursor’s shell, browser, file-editing and web-search tools.
Neon counts as a third-party model in Cursor, so standard plan usage does not cover it. Current plans charge prepaid credit, while legacy plans draw from the Other Models pool. Cached input costs $0.15 per million tokens, a 95% reduction from the regular input rate, which can lower the cost of agent loops that repeatedly read the same files and prompts.
| Product | Access |
|---|---|
| Cursor | Model picker with prepaid or legacy Other Models billing |
| Google AI Studio | Gemini API paid tier |
| Antigravity | Paid AI Pro and Ultra subscriptions |
| Gemini for business | Business accounts |
| Gemini app | Supported consumer accounts |
| Google Cloud | Verified enterprise customers in selected regions |
With the google-genai package installed and GEMINI_API_KEY set, a Python request can select Neon and its reasoning level directly:
from google import genai
from google.genai import types
client = genai.Client()
response = client.models.generate_content(
model="gemini-4-neon",
contents="Refactor auth.py to use dependency injection.",
config=types.GenerateContentConfig(
thinking_config=types.ThinkingConfig(
thinking_level="high"
)
),
)
print(response.text)
Google identifies Neon as the migration target for Gemini 3.1 Pro. It also recommends Neon for coding and enterprise reasoning workloads previously assigned to Argon, along with complex coding tasks moved from Gemini 3.8 Flash.
Choose by completed-task cost
Neon extends beyond coding in Google’s launch results, scoring 56.3% on Zapier’s AutomationBench and leading Vending-Bench 2, a long-running business simulation. Google also reports first-place results in Surge AI’s finance and accounting evaluations. A specialized cybersecurity version, Neon for cyber, is available to vetted security teams through the Fairwind Program and scores 90.5% on SEC-bench Pro, which measures vulnerability discovery.
Model selection depends on the cost and reliability of a completed task, including retries, tool calls and generated tokens. Cursor’s figures support Neon High for multi-file refactors, debugging and tool-heavy agent sessions. Medium, Low or Hydrogen suit higher-volume workloads with tighter latency budgets, while Extra High, Carbon and competing frontier models remain candidates for difficult long-horizon work.
Production evaluations should track task success, wall-clock latency, total token cost, cache use, tool failures and human correction time on representative repositories. Neon's large SWE-sweep variance shows that the surrounding harness can influence results as much as the model name in the picker.