Title card reading Gemini 4 Neon Is Here under the line The newest Gemini 4 costs more per token than the flagship, above two slim price cards — Gemini 4 Neon at $3.00 in / $15.00 out and Gemini 4 Carbon at $2.00 in / $10.00 out — with the caption Flat rate, no free tier.
Guides & Insights

Gemini 4 Neon Is Here, and It Costs More Per Token Than the Flagship It Was Measured Against

Author

jinhao.song

Date Published

Latest models · 20View all models →
Benchmarks: Artificial Analysis · updated daily
Back to all posts

Gem​ini 4 Neon went generally available on 18 November 2026 at $3.00 per million input tokens and $15.00 per million output tokens. Gemini 4 Carbon, the model Goog​le benchmarks it against in almost every paragraph of the announcement, is $2.00 and $10.00. For thirteen days — until Carbon's introductory rate expires on 30 November — the newest, fastest-tier Gem​ini 4 model is the most expensive per-token option in the family, and it is priced that way deliberately. Neon is not meant to be the cheap one. It is meant to be the one that finishes.

That inversion is the whole story, and it is easy to miss because every other Gem​ini 4 launch this autumn ran the other direction. Argon shipped in October as the efficient frontier model. Carbon shipped in November as the strongest reasoning model, gated first to cyber defenders and then opened to everyone on 10 November. Hydrogen shipped two days after that as the everyday, free-tier-having, $1.50/$7.50 workhorse. Each of those names tells you where it sits: heavier, or lighter. Neon tells you nothing, and the price says "heavier" while the positioning says "faster."

What $3 and $15 actually buys

The rate card is flat. There is no introductory window, no scheduled increase, no free tier. Goog​le's pricing page lists $3.00 input, $15.00 output, $0.15 cached input — the 95% cache discount the whole family carries — and no date at which any of those numbers move. Batch and flex tiers run at half rate, $1.50 and $7.50. Priority runs at 1.8 times standard, $5.40 and $27.00. Implicit caching only; explicit caching is not supported on any Gem​ini 4 model.

Against that, Carbon and Argon both run $2.00/$10.00 with a footnote: through 30 November 2026, then $4.00/$20.00. Which produces the odd situation where the price ranking of the family changes on a date that has nothing to do with Neon's launch.

• Through 30 November 2026 — Neon $3.00/$15.00 is the most expensive Gem​ini 4 tier; Carbon and Argon are $2.00/$10.00, Hydrogen is $1.50/$7.50.

• From 1 December 2026 — Carbon and Argon step to $4.00/$20.00; Neon stays at $3.00/$15.00 and becomes 25% cheaper than both, without any change to its own rate card.

• Cache pricing follows the same shape — Neon $0.15, Carbon and Argon $0.10 then $0.20, Hydrogen $0.075.

• Batch and flex — Neon $1.50/$7.50, Carbon and Argon $1.00/$5.00 then $2.00/$10.00, Hydrogen $0.75/$3.75.

• Free tier — Hydrogen only. Neon, Carbon and Argon are paid-tier models with no free quota at all.

So the honest way to read Neon's price is not "expensive" or "cheap" but "the only Gem​ini 4 rate card that does not move in December." If you are budgeting a workload that runs past the end of the month, the comparison you should run is Neon against what Carbon will cost you in January, not what Carbon costs you today.

Four thinking levels, and only two models have the fourth

Neon defaults to high and supports low, medium, high and xhigh. That is the same ladder Hydrogen has. Carbon and Argon stop at high and default to medium. Goog​le's own docs are explicit that high is "maximum reasoning depth" on Carbon and Argon but merely the recommended default on Neon and Hydrogen, with xhigh reserved above it for hard multi-step coding and agentic problems.

This matters more than it looks, because most of Neon's headline benchmark numbers are quoted at xhigh while Carbon's are quoted at its own default. The Artificial Analysis Intelligence Index figure Goog​le cites for Neon — 61, third overall — is an xhigh number. Neon's $3/$15 rate does not include any allowance for the extra tokens xhigh burns. Goog​le's own documentation warns that xhigh "can cause overthinking or exhaust max_output_tokens on simpler prompts such as text or JSON extraction," and recommends running evaluations at all four levels and picking the lowest one that clears your quality bar. That advice is worth more than the benchmark table, because the gap between Neon at high and Neon at xhigh is a gap you pay for twice: once in tokens, once in latency.

The benchmark claims, and what kind of number each one is

Goog​le's announcement is unusually dense with numbers, and almost none of them are the kind a buyer can re-run. It is worth sorting them before quoting any of them internally.

Third-party leaderboards Goog​le is citing, which you can go and read. The Vals Index at 70.9% and rank #2, on the Pareto frontier. A debut at #2 globally on Arena's new Agent Arena Index with 1336 across 4,200+ real-world agent sessions, and #2 on that index's steerability measure. First place on Zapier's AutomationBench at 56.3% and on Andon Labs' Vending-Bench 2 at $16,843. Agents' Last Exam v1.1 at 75.8%, described as on par with leading models rather than ahead of them.

Goog​le's own reported numbers, which nobody outside has reproduced. Chartography 73.5%. Terminal-Bench Science 70.5% and RiemannBench 84.0% and HLE-Diamond 71.3%, each called best in class. Terminal-Bench 4.1 at 65.2%, FrontierCode 2.0 at 46.2% on the default high setting, and first overall on the Artificial Analysis Coding Agent Index at 58.6. On SWE-sweep, 12.8% under the mini-SWE harness and 39.0% inside Goog​le's own Antigravity harness — a threefold difference that says more about the harness than about the model.

The cyber numbers, which ship separately. AutoBaxBench hard tier 76%, WebExploitBench v1.2 at 51.26%, SEC-bench Pro 90.5%, and CWE-bench v2 at 70.7% — 3.4 points over Carbon. Those come from the "Gem​ini 4 Neon for cyber" variant, which Goog​le says reaches trusted defenders through the Fairwind Program "over the following days." That access gate has not opened yet as of 19 November, so the security numbers are claims about a model almost nobody outside Goog​le can call.

The efficiency claim, which is the one to check against your own bill. Goog​le says Neon at default high costs 32% less per task than Carbon, and 36% or more less per task than competing frontier models at their highest thinking setting. Per-task cost is a function of tokens consumed, and Neon's verbosity is not disclosed anywhere in the announcement. The number is plausible — Carbon is documented as a very verbose model — but it is a vendor figure about a vendor benchmark.

One more thing worth knowing if you plan to cite the index number. Artificial Analysis had no Gem​ini 4 Neon model page at all when we checked on 19 November: its model index lists Gemini 4 Carbon and Gem​ini 4 Hydrogen and nothing else from the family. The 61 and the #3 are Goog​le citing Artificial Analysis, not Artificial Analysis publishing a Neon entry you can inspect. That will change; until it does, treat the figure as attributed rather than verified.

A single-column scoreboard titled Gemini 4 Neon — the scoreboard, with six rows: price $3.00 in / $15.00 out, cached input $0.15, free tier none, thinking levels low, medium, high and xhigh, default thinking high, and AA Index 61 at xhigh Google-cited. The footer notes Neon figures are vendor-reported and that the AA Index of 61 is Google citing Artificial Analysis, which had no Neon page on 19 November 2026.

What Neon is actually built to do

The clearest signal in the announcement is not a benchmark. It is that Goog​le describes Neon as an orchestrator: "acting as a fast, high-judgment Orchestrator directing parallel Hydrogen subagents in Goog​le Antigravity." That is a specific architectural role, and it explains the price. An orchestrator issues many turns, holds a lot of context, and delegates the bulk work downward to a cheaper model. If your agent loop is a planner that spends its tokens on coordination and hands execution to a $1.50/$7.50 worker, paying $3.00 for the planner is rational. If your loop is a single model doing everything, Neon is strictly more expensive than Hydrogen for the same job, and more expensive than Carbon until 1 December.

Two other capabilities are worth noting because they are family-wide rather than Neon-specific, and marketing copy tends to imply otherwise. Every Gem​ini 4 model has the same unified 1M-token context and decodes up to 64K output tokens by default with up to 1M available; Neon is not buying you a bigger window than Hydrogen. And every Gem​ini 4 model accepts text, image, audio, video and PDF. What Neon adds is not modalities or context — it is the combination of the xhigh ceiling, the orchestrator positioning, and a rate card that does not expire.

The concrete demos Goog​le shows all run inside Antigravity with the /swarm feature: a WebGL2 ski simulator over real Mont Blanc elevation data, an interactive 3D reconstruction of Turing's 1936 rooms at King's College, and a single image converted into a 100,000-cell spreadsheet driven by 500,000 live formulas. Those are demonstrations of a harness plus a model, not of a model alone, and none of them are reproducible from the API today.

What you can call right now

Gem​ini 4 Neon is available on the Gemini API paid tier, in Goog​le AI Studio and Android Studio, in Gemini for business via API and model dropdown, in Goog​le Antigravity for paid AI Pro and Ultra subscribers, and in the Gemini app for Pro and Ultra subscribers. It is not in the free tier anywhere.

On our side, the catalogue currently carries google/gemini-4-carbon and google/gemini-4-argon. Neon is not there yet, so anything you read here about calling it applies to those two, not to Neon — we do not serve a model we do not serve, and we are not going to pretend a rate card is an endpoint. When Neon lands in the catalogue it will inherit the same terms as the rest: provider list price passed through with no markup, so Goog​le's numbers are our numbers, and the December step on Carbon and Argon appears in your bill the day it appears in Goog​le's.

What the router is useful for in the meantime is the thing Neon is designed to be part of. If you are building the orchestrator pattern Goog​le is describing — a strong planner calling cheaper workers — you do not want that structure to depend on one vendor's harness reaching general availability on schedule. A fallback chain configured against Carbon with Argon behind it keeps a production path alive while you evaluate whether Neon's per-task economics hold up on your traffic rather than on a benchmark suite.

Google's blog post Introducing Gemini 4 Neon, with the headline, the subtitle about frontier-level intelligence with fast end-to-end task completion across software engineering, knowledge work and science workflows, and the byline Koray Kavukcuoglu.

The decision, stated plainly

If your work is high-frequency frontier reasoning — dense documents, multi-file debugging, tool-heavy agents that need to finish rather than to explore — Neon is the model Goog​le built for that, and after 1 December it will also be the cheapest way to get Gem​ini 4-level reasoning per token. If your work is the hardest long-horizon problem you have, Carbon is still the model Goog​le points at, and its price step is the thing to plan around rather than Neon's launch. And if your work is everyday execution at volume, none of this concerns you: Hydrogen has been doing that since 12 November at half Neon's rate, with a free tier Neon will never have.

The open questions to watch are narrow and checkable. Whether Artificial Analysis publishes its own Neon entry, and what it scores outside Goog​le's citation. Whether the Fairwind gate on Neon for cyber actually opens in the "coming days" Goog​le promised. And whether the 1 December step on Carbon and Argon survives contact with demand — a 100% increase on the two models most production traffic currently routes to is the largest scheduled price change in this family, and it lands thirteen days after a model launched at a price that only looks high until it does.

The OrcaRouter models catalogue, showing the filter rail across input modalities, context length, input price, status, access tier, series, supported parameters, privacy and inference region, above the models list.