Skip to main content
Gemini 4 Neon is Google’s model for complex software engineering, enterprise analysis, and scientific workflows that need frontier reasoning without waiting on a slower model. It sits between Gemini 4 Hydrogen and Gemini 4 Carbon, and Google designates it as the target upgrade for workloads running on Gemini 4 Argon, Gemini 3.1 Pro, or demanding jobs moving up from Gemini 3.8 Flash. The identifier on this page is gemini-4-neon. The 1,048,576-token window is unified across input, thinking, function calls, and the final reply. A single foreground segment generates up to 65,536 tokens for both streaming and non-streaming requests. With long decode (SDK automatic continuation, a continuation_token loop, or background=true), generation can extend up to 1,048,576 tokens minus your input. On Gemini Enterprise Agent Platform, the provider also accepts Word, PowerPoint, Excel, and OpenDocument files; agentic video understanding is available and off by default. This page does not claim that the Tess picker exposes audio or video.

Where Neon fits in the Gemini 4 family

While Hydrogen handles fast, high-volume everyday execution and Carbon takes on open-ended long-horizon projects, Neon balances frontier reasoning depth with nearly double Carbon’s output speed for recurring complex workflows.
Reach for Neon when Argon felt slow or capped out and Hydrogen does not go deep enough. Keep high-volume, routine agent steps on Hydrogen.

What Google measured

Google highlights wins across enterprise agent workflows, legal and tax research, scientific reasoning, and mid-task steerability.
  • On the Vals Index v2.2, Neon ranked 2nd overall at 70.9%, taking 1st place on both Legal Research Bench and Tax Agent Bench.
  • On Zapier’s AutomationBench, it ranked 1st overall at 56.3%, and also placed 1st on Vending-Bench 2 (US$16,843). On the Agent Arena Index, it debuted 2nd globally (1,336 across 4,200+ real agent sessions) and ranked 2nd on steerability (adapting cleanly to mid-task user corrections).
  • On science and math, it scored 70.5% on Terminal-Bench-Science 0.1, 84.0% on RiemannBench, and 71.3% on HLE-Diamond.
  • On production coding, it ranked 1st on the Coding Agent Index cited in the announcement (58.6), scoring 65.2% on Terminal-Bench 4.1 and 46.2% on FrontierCode 2.0 at default high. On SWE-sweep, it scored 12.8% on the mini-SWE harness and 39.0% in Google Antigravity.
  • Across multimodal and long-context evaluations, it posted 73.5% on Chartography, 75.8% on Agents’ Last Exam v1.1, and 95.2% on GraphWalks (256K to 1M tokens). Google also reported its lowest internal fabrication rate to date and a 99.2% pass rate on agent honesty.
Multi-model orchestration that uses Neon as a planner and hands subagent steps to Hydrogen (previous_interaction_id or /swarm in Google Antigravity) is a Google platform feature, not a Tess picker control.

What Artificial Analysis measured

At high, the default setting on gemini-4-neon, Neon scores 60 on the Intelligence Index, matching Carbon (high, 60) in well under half the time per task. Artificial Analysis also evaluated Google’s API xhigh tier at 61 (3rd overall behind GPT-6.2 Astra and Claude Fable 5.2); in Tess, Neon runs on LOW, MEDIUM, and HIGH. On dense professional PDFs (GDP.pdf), Neon (high) scored 33.4%, 5 points ahead of Carbon (high, 28.4%). On AA-Briefcase v1.1, high reached 1,632 Elo, also ahead of Carbon (high, 1,611). On FrontierCode-AA 2.0, high scored 43.3% against Carbon’s 40.6%.

Where it does not lead

  • On the overall Intelligence Index, high at 60 sits behind Claude Fable 5.2 (max, 65) and GPT-6.2 Astra (xhigh/max, 62).
  • On two Artificial Analysis evaluations, Neon (high) trailed Carbon (high): CritPt v2 (Hard Subset) at 59.0% vs. Carbon’s 69.0%, and Terminal-Bench 4.0 at 61.6% vs. Carbon’s 65.2%.
  • While Google’s provider documentation lists XHIGH, the thinking levels available in Tess are LOW, MEDIUM, and HIGH.
  • Passing thinking_budget, thinking_level="MINIMAL", frequency_penalty, presence_penalty, or candidate_count > 1 returns an HTTP 400 error. temperature, top_p, and top_k are ignored.
  • Offensive cyberdefense configurations are restricted to Google’s Fairwind Program; this page does not cover vulnerability exploitation. The model card knowledge cutoff is March 2026.

Where it creates more value

Multi-file debugging and refactoring

Neon cuts down on redundant file reads, survives context compaction in long tool-heavy sessions, and adapts cleanly when you steer it mid-task. Expected result: surgical edits across the right modules, concise progress updates, and verified test runs. Combine contracts, statements, spreadsheets, and dense charts in one 1M-token window to produce structured deliverables, playing to Neon’s strengths on GDP.pdf (high), Legal Research, Tax Agent, and AA-Briefcase presentation quality. Expected result: a structured report or table with every figure tied to its source page or tab, and missing inputs flagged rather than guessed.

Scientific and quantitative workflows

Applied math, experimental data interpretation, and multi-step analytical scripts that need deeper reasoning than Hydrogen provides, finished much faster than on Carbon. Expected result: checked calculations, runnable verification scripts, and a clear boundary between proven results and open questions.

How to apply it in Tess

1

Choose between LOW, MEDIUM, and HIGH

HIGH is the default, matching Carbon’s overall index score in less than half the time while leading on dense PDFs. Use MEDIUM to balance latency on day-to-day runs and LOW for routing, classification, and fast extraction.
2

Steer mid-conversation instead of restarting

If an agent run heads down the wrong path, send a direct correction in the next turn: Neon handles mid-task course changes cleanly without losing prior context.
3

Front-load static context

Explicit caching is unsupported, while durable implicit caching triggers automatically on shared prompt prefixes of 4,096 tokens or more.
4

Leave headroom in the unified window

Because input, thinking, tool calls, and output share one 1,048,576-token budget, trim unneeded attachments when you expect a long reasoning trace or deliverable.
5

Scope tools and human approvals

Enable only the tools required for the workflow and require confirmation before mutating external records, sending messages, or publishing files.

Three playbooks to start

1. Surgical multi-file refactor

Input: target modules, public contracts that must hold, and test command.
Tools: code search, file read/edit, and terminal.
Done when: only the necessary files are modified, tests pass, and any out-of-scope decisions are surfaced. Input: PDF reports, supporting spreadsheets/charts, and brief outline.
Tools: document and image reading.
Done when: the brief is formatted for immediate review, every figure traces to a source, and conflicts are isolated.

3. Reproducible quantitative or scientific check

Input: dataset or technical problem statement and validation criteria.
Tools: code/script execution and file reading.
Done when: the numbers reproduce from the returned script and data limits are stated plainly.

Limits and good practice

  • Default thinking_level on gemini-4-neon is HIGH (up from MEDIUM on Argon and Carbon), so migrating latency-sensitive turns without setting MEDIUM explicitly will spend more thinking tokens.
  • In direct API integrations with parallel function calls, only the first functionCall part in a turn carries thought_signature, and signatures must be preserved across prior turns to avoid 400 errors or degraded tool calls.
  • Foreground single-segment output still caps at 65,536 tokens (including thinking); longer completions require v1beta1 continuation.
  • Check Models and Costs for Tess credit pricing once the slug is in the catalog, and keep human sign-off on financial, legal, clinical, or production changes.
See also: Gemini 4 Carbon · Gemini 4 Hydrogen · Gemini 4 Argon · Models and Costs · Introducing Gemini 4 Neon · What’s new · Model card · Artificial Analysis.