cd /news/large-language-models/gemini-4-neon-the-balance-point-betw… · home › topics › large-language-models › article
[ARTICLE · art-172258] src=dev.to ↗ pub= topic=large-language-models verified=true sentiment=↑ positive

Gemini 4 Neon: The Balance Point Between Carbon and Hydrogen — Benchmarks, API Pricing, and Orchestrator Workflows

Google released Gemini 4 Neon on November 18, 2026, a mid-tier model priced at a flat $3 input / $15 output per 1M tokens with a unified 1M-token context window and thinking levels up to xhigh. The release lands two days before introductory pricing on Gemini 4 Carbon and Argon doubles to $4 / $20 on December 1, making Neon cheaper than Carbon for high-frequency workloads like code review and agent loops. Neon ranks #3 on the Artificial Analysis Intelligence Index (61) but #1 on its Coding Agent Index (58.6), Zapier AutomationBench (56.3%), and Andon Labs Vending-Bench 2 ($16,843).

key takeaway

Gemini 4 Neon's flat $3/$15 per-1M pricing undercuts Carbon's post-November-30 rate of $4/$20, and because all Gemini 4 models share one request schema, model IDs can be swapped without code changes.

by read10 min views4 publishedNov 19, 2026

On November 18, 2026, Google introduced Gemini 4 Neon — the fourth axis of the Gemini 4 family after Argon (October 13), Carbon (November 10), and Hydrogen (November 12). Google's one-line definition is "frontier-level intelligence with fast end-to-end task completion." Not the smartest model, but a model that carries near-frontier judgment to completion at a cost and latency you can repeat all day.

The reason this matters in practice is not the benchmark ranking but the price curve. According to Google's published pricing, Carbon and Argon carry an introductory price ($2 input / $10 output per 1M tokens) that expires on November 30, 2026, after which they become $4 / $20 — double, with cached input moving from $0.10 to $0.20. Neon has no introductory price: a flat $3 / $15.

The crossover happens in two days. Any high-frequency workload you run on Carbon today (code review, refactoring, document extraction, agent loops) doubles in cost on December 1 if left untouched; moving the same work to Neon flips its position from 1.5x the introductory Carbon rate to 0.75x the new Carbon rate.

Rather than restating the launch numbers, this article breaks adoption down into the four axes an engineer actually has to decide on: token accounting, thinking-level defaults, harness integration, and safety/honesty metrics.

All three models share the same request schema, a unified 1M-token context window, and long decode up to 1M tokens. That means you can switch model IDs without changing your code — the precondition for the mixed-model workflows in section 5.

| Item | Hydrogen | Neon | Carbon |

|---|---|---|---|
| Model ID | gemini-4-hydrogen | gemini-4-neon | gemini-4-carbon | 

| Speed | Fast | Moderate | Deliberate | | Context | 1,048,576 unified | 1,048,576 unified | 1,048,576 unified | | Max output | 65,536 default / 1,048,576 long decode | Same | Same | | Thinking levels | low, medium, high, xhigh (default high) | low, medium, high, xhigh (default high) | low, medium, high (default medium), no xhigh | | Input per 1M | $1.50 | $3.00 | $2.00 → $4.00 on Dec 1 | | Cached input per 1M | $0.075 | $0.15 | $0.10 → $0.20 on Dec 1 | | Output per 1M (incl. thinking) | $7.50 | $15.00 | $10.00 → $20.00 on Dec 1 | | Free tier | Available | None | None |

Three details are easy to miss. First, Carbon has no xhigh and defaults to medium, so always set thinking_level explicitly during a migration. Second, all Gemini 4 models use flat pricing across the full 1M window — 500K tokens costs the same per million as 5K. Third, only Hydrogen has a free tier; free-tier input and output are used to improve Google's products.

Table 2. Official benchmark results

Benchmark Result Rank / condition
Artificial Analysis Intelligence Index 61 #3 overall at xhigh
Artificial Analysis Coding Agent Index 58.6 #1 overall
Vals Index 70.9% #2, on the Pareto frontier
Vals Legal Research / Tax Agent Bench - #1 on each
Arena Agent Arena Index 1,336 #2 over 4,200+ real agent sessions
Zapier AutomationBench 56.3% #1
Andon Labs Vending-Bench 2 $16,843 #1
Agents' Last Exam v1.1 75.8% On par with the leaders in its class
Chartography 73.5% -
Terminal-Bench Science 70.5% Best in class
RiemannBench 84.0% Best in class
HLE-Diamond 71.3% Best in class
Terminal-Bench 4.1 65.2% -
FrontierCode 2.0 46.2% At the default high setting

| SWE-sweep | 12.8% → 39.0% | mini-SWE harness → Antigravity harness (SOTA) | Read the table in this order. Start with general intelligence (61, #3) as "not first": Google itself says it lands "just behind larger, slower models," and the score is at xhigh. When you genuinely need maximum intelligence, Carbon or a competitor's largest model still wins; Neon's value is how cheaply it completes the repetitive work beneath that tier.

Next, note the three consecutive #1s on the agentic axis (Coding Agent 58.6, AutomationBench 56.3%, Vending-Bench 2 at $16,843). They share one trait: they are not single-answer tasks but multi-step execution with tool calls, file reads, state retention, and long-horizon capital management. Conversely, the knowledge and science axis is "best in class," which Google's own wording signals as within error of peers.

The most practically useful line is the last one: 12.8% on the mini-SWE harness versus 39.0% in Antigravity — the model is identical, the harness tripled the score.

Methodological caveats: ranks are among "comparable models"; each score carries a different thinking level; third-party indices weight cost and latency; Agent Arena is sensitive to sample composition; honesty and cheating metrics are Google-internal and cannot be reproduced; public benchmarks risk contamination.

Neon costs $3 input, $0.15 cached input (95% discount), and $15 output per 1M tokens. The critical detail: output pricing includes thinking tokens. Raising the thinking level directly increases tokens billed at the output rate, and max_output_tokens caps thinking plus output together. Your first cost lever is the thinking level, not prompt length.

Table 3. API pricing, caching, and savings (per 1M tokens, USD)

Item Neon Carbon (now → Dec 1) Hydrogen
Input $3.00 $2.00 → $4.00 $1.50
Cached input (95% off) $0.15 $0.10 → $0.20 $0.075
Output (incl. thinking) $15.00 $10.00 → $20.00 $7.50
Caching mode Implicit only Implicit only Implicit only
Long-context surcharge None None None
Cost per task (default high) 32% lower than Carbon Baseline -

Reverse-engineering the 32%. Neon's token price is 1.5x Carbon's introductory price, yet cost per task is 32% lower. Simple accounting implies Neon bills roughly 45% of Carbon's tokens for the same task (0.68 ÷ 1.5 ≈ 0.45). This is my calculation from those two published figures, not a Google disclosure, and it assumes both cost-per-task numbers were measured on the same benchmark and task set. The direction is unambiguous: Neon's advantage comes from token efficiency, not from the rate card.

The December 1 scenario. Carbon and Argon input, output, and cached-input prices all double; Neon does not change. Run your budget twice — before and after December 1.

A cache accounting example (my calculation). Assume one task uses 200K input tokens (150K cache hits) and 4K output tokens including thinking. Total cost: Neon $0.2325, Carbon (through Nov 30) $0.155, Carbon (from Dec 1) $0.310. When 75% of input is cached, input cost falls to roughly a fifth of the total and output dominates.

The Batch API also gives a 50% reduction, and Flex and Priority inference are supported. Hydrogen costs half of Neon at $1.50 / $7.50, and Gemini 3.8 Flash moves to the same price on January 1, 2027 — so the Flash-to-Hydrogen decision becomes "the same price from 2027 anyway, so move now."

Orchestrator: Neon plans, Hydrogen executes. Google Antigravity's Orchestrator can now pair Neon as planner with Hydrogen as executor in a single session. Neon handles planning, architecture, and complex reasoning; Hydrogen takes routine edits, package updates, and test runs. If Hydrogen hits friction, it hands back to Neon to inspect state, repair the plan, and resume. Google reports the analogous Carbon + Hydrogen pairing held near-Carbon success rates on Terminal-Bench 4.0 while consuming 24% less quota and finishing 19% faster.

Where you can use it. Neon is available in the Google Antigravity desktop app, Antigravity CLI, and IDE extensions for paid Google AI Pro and Ultra subscribers; enterprises reach it through Gemini for business. On the Gemini API, call gemini-4-neon on the paid tier (Google AI Studio, Android Studio).

GitHub Copilot availability. Neon is generally available in Copilot for Pro+, Max, Business, and Enterprise plans, selectable across VS Code, Visual Studio, Copilot CLI, Copilot coding agent, the Copilot app, github.com, GitHub Mobile, JetBrains IDEs, Xcode, and Eclipse. Billing is usage-based at provider list pricing, separate from premium request allowances, and Enterprise and Business administrators manage access through the model policy — new models are enabled by default unless an admin disables them.

1M multimodal and long decode. Neon ingests text, images, video, audio, and PDFs in one 1M window. Synchronous responses cap at 65,536 tokens, so for longer output set background: true together with max_output_tokens (131,072–262,144). Hard prompts can take minutes; set HTTP timeouts to at least 15 minutes.

What the /swarm demos prove. A 3D lab bench driving Blender MCP, Alan Turing's 1936 study reconstructed in interactive 3D, a 100,000-cell Google Sheet with 500,000 live formulas from a single image, and a first-person WebGL2 ski simulator with 10,200 lines of custom shaders — all built with zero pre-made assets. The takeaway: harness design, more than model choice, sets the ceiling of what you ship.

Code-level migration checklist

Cyber defense (Neon for cyber). A variant tuned for defensive workflows, rolling out to trusted defenders through the Fairwind Program — not an offensive capability. Secure code generation is state-of-the-art on the hard AutoBaxBench tier (76%), vulnerability discovery hits 90.5% on SEC-bench Pro at roughly half the cost per challenge, ExploitGym validation runs at frontier level with half the tokens of Carbon, patching reaches 70.7% on CWE-bench v2 (a 3.4-point gain over Carbon), and WebExploitBench v1.2 lands at 51.26% SOTA.

Safety and honesty. Legitimate refusals are 1.98% on Terminal-Bench 4.0 and 0% on Terminal-Bench-Science; cheating and shortcuts are 0.0% on Terminal-Bench 4.0 and 2.45% on Google's coding cheating evaluation; the misaligned action rate is 0.09% with a 99.2% agent honesty pass rate; under-service is down 43–56% and overreach down 15–35%.

Separate the conditions: refusal rates, honesty, and misalignment are Google-internal evaluations, and the 0.0% IPI figure holds at the highest thinking level with safeguards enabled. If your system feeds external web pages, emails, or documents into an agent, do not treat vendor numbers as a waiver — run your own injection red team on a schedule. The 1.98% refusal rate does indicate the model is not over-blocking legitimate security work, which is the most common failure mode in internal security automation.

1. Choose the tier by frequency × difficulty. High frequency with moderate difficulty (multi-file refactoring, code review, document analysis, agent loops) means Neon. Low frequency with extreme difficulty (architecture, cross-subsystem root-cause work, formal proofs, security audits) means Carbon. High frequency with low difficulty (translation, summarization, JSON extraction at volume) means Hydrogen or Flash-Lite. For a solo developer: interactive IDE work on Neon, overnight automation on Hydrogen batch.

2. Model your budget twice, at December 1 prices. Pull the monthly cost of your current Carbon traffic and place three scenarios side by side: (a) keep it as is, (b) move it to Neon with an explicit cache hit-rate assumption, (c) keep Carbon for the hardest paths only and split the rest.

3. Pin thinking levels with an evaluation set, and gate xhigh. Run low through xhigh and take the lowest level that clears your quality bar as the default. Keep max_output_tokens at 65,536 or more during evaluation.

4. Fix the harness and billing integration first. On the direct API, apply the checklist above. On Antigravity, enable the Neon + Hydrogen Orchestrator and measure quota. On Copilot, confirm the plan, rollout status, billing limits, and model policy. Finally, log tokens, cache hit rate, and retry rate.

Gemini 4 Neon is not "the smartest model." Third overall in general intelligence, best in class on science and knowledge, first on agentic coding and automation — together that describes a model that keeps judgment where it matters while lowering the unit cost of repeated execution. With Carbon and Argon doubling on December 1, Neon's real value is not the $3 / $15 rate card but its operating position: high-frequency complex work at 32% lower cost per task than Carbon, with a gap that widens after December.

Decide in this order: frequency × difficulty → December price recomputation → pinned thinking levels → harness and billing integration — not by the leaderboard.

Originally published at uulab-ai.blogspot.com

── more in #large-language-models 4 stories · sorted by recency
── more on @google 3 stories trending now
sponsored brought to you by zahid.host 4,200+ EU-deployed projects
reading about agents? ship yours in a single git push.

Run your AI side-project on zahid.host

EU-based hosting, git-push deploys, automatic HTTPS, no cold starts. Free tier with a custom domain — perfect for shipping the agent you just read about.

$git push zahid main
→ Live at https://your-agent.zahid.host ✓
Get free account → Pricing
from €0/mo · no card required
LIVE [news/gemini-4-neon-the-ba…] indexed:0 read:10min 2026-11-19 · —