Title card reading Gemini 4 Neon vs Gemini 4 Carbon under the line Slower to be trusted, faster to finish, above two cards — Cost per token, Neon higher, and Cost per task, Neon lower, per Google — with the caption One number prices tokens. The other prices the work.
Guides & Insights

Gemini 4 Neon vs Gemini 4 Carbon: The Newer Model Is Slower to Be Trusted and Faster to Finish

Author

jinhao.song

Date Published

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

Put Gem​ini 4 Neon and Gemini 4 Carbon side by side on a rate card and Neon loses: $3.00 in and $15.00 out against Carbon's $2.00 and $10.00. Put them side by side on a finished task and Goog​le says the opposite, claiming Neon costs 32% less per task at its default thinking level. Both statements are true, and the gap between them is the entire decision. One number prices tokens. The other prices the work those tokens do — and when a family ships four models inside eight weeks, the second number is the only one that tells you which model to put in production.

Both are generally available. Carbon since 10 November 2026, after two weeks gated to cyber defenders through Goog​le's Fairwind Program. Neon since 18 November 2026, open to any paid Gemini API developer from day one. They share a 1M-token unified context, the same five input modalities, the same 64K default output ceiling with up to 1M available, the same implicit-only caching, and the same absence of a free tier. Almost everything that distinguishes them is behaviour, not specification.

Cost per token, cost per task, and the thirteen days that invert them

The rate cards first, because the reversal on 1 December changes the answer to "which is cheaper" without either model changing at all.

• Today, 19 November 2026 — Carbon $2.00 input / $10.00 output / $0.10 cached; Neon $3.00 / $15.00 / $0.15. Carbon is 33% cheaper on every line.

• From 1 December 2026 — Carbon steps to $4.00 / $20.00 / $0.20 on Goog​le's published schedule; Neon does not move. Neon becomes 25% cheaper on every line.

• Batch and flex — Carbon $1.00 / $5.00 today, $2.00 / $10.00 after the step; Neon flat at $1.50 / $7.50.

• Priority — the 1.8x tier: Carbon $3.60 / $18.00 today, $7.20 / $36.00 after; Neon flat at $5.40 / $27.00.

• Artificial Analysis's own Carbon page, read on 19 November, lists Carbon at $4.00 and $20.00 — the December rate, not the one Goog​le is currently billing. If you are pulling pricing from a leaderboard rather than from the vendor's rate card, you may already be comparing against a price nobody is paying yet.

Now the per-task number, which is where Neon's case lives. Goog​le's announcement states that at Neon's default high thinking level it delivers "a 32% lower cost per task than Carbon." That is not a pricing claim, it is a verbosity claim wearing a pricing claim's clothes. A model that reasons for twice as many tokens at half the rate costs the same per task; Goog​le is saying Neon emits materially fewer tokens to reach a comparable answer. It is also the one claim in the announcement that maps directly onto your own invoice, which means it is the one you should re-measure on your own traffic rather than accept.

There is independent evidence for the direction of it. Artificial Analysis measures Carbon at 205M output tokens across the Intelligence Index, which it describes as very verbose against a median of 65.1M, at 36.0 minutes per task and $7.76 per task, with 57 tokens/second output speed and a 75-second time to first answer token. Whatever Neon's own numbers turn out to be, Carbon's profile is established: slow to start, heavy on tokens, expensive per task at its list price. Neon's entire positioning is built on being the opposite of that profile.

The comparison Goog​le is actually making is not at matched settings

This is the trap in every Neon-versus-Carbon table circulating this week, including Goog​le's own chart. The two models do not have the same thinking ladder, and their headline numbers are not measured at the same rung.

• Thinking levels — Carbon: low, medium, high, defaulting to medium. Neon: low, medium, high, xhigh, defaulting to high.

• What "high" means — on Carbon it is the maximum; on Neon it is the recommended default, with a deeper rung above it.

• The index figure Goog​le cites for Neon — 61 on the Artificial Analysis Intelligence Index, third overall — is an xhigh number. The same announcement quotes Neon's cost-per-task advantage at high. Those are two different configurations being described in adjacent sentences.

• Carbon's index figure — 59 on index v5.0 at launch, ranked first at the time; Artificial Analysis's own page now shows 60 at high, seventh of 92 models, as newer releases moved above it.

So the honest comparison is not 61 versus 60. It is: Neon at xhigh beats Carbon at high on the index by roughly a point; Neon at high is the configuration whose economics Goog​le is claiming; and Carbon has no xhigh to retreat to. Goog​le's own documentation advises running evaluations at all four Neon levels and picking the lowest one that clears your bar, warning that xhigh "can cause overthinking or exhaust max_output_tokens on simpler prompts." A model with a higher ceiling and a lower default is a model you have to tune, not one you can drop in.

A two-column scoreboard titled Gemini 4 Neon vs Gemini 4 Carbon — the scoreboard. Gemini 4 Neon reads price today $3.00 / $15.00, price from 1 Dec $3.00 / $15.00, cached input $0.15 flat, thinking levels low to xhigh, default thinking high, AA Index 61 at xhigh; Gemini 4 Carbon reads $2.00 / $10.00, $4.00 / $20.00, cached input $0.10 then $0.20, thinking levels low to high, default thinking medium, AA Index 60 at high. The footer says Neon figures are vendor-reported and Carbon's Index of 60 is per Artificial Analysis, 19 November 2026.

Where Carbon is still the better model, and it is not a small list

Goog​le's positioning sentence is that "Carbon tackles your hardest, long-horizon projects and Hydrogen handles fast everyday execution, while Neon balances depth and efficiency." Read literally, that puts Neon between the two, not above Carbon — and the benchmark list supports the reading in the places that matter most for long-horizon work.

• FrontierCode 2.0 — Neon 46.2% at its default high. Goog​le does not publish Carbon's figure on the same benchmark in the Neon announcement, which is the kind of omission worth noticing in a document that is otherwise generous with side-by-side numbers.

• SWE-sweep — Neon 12.8% on the mini-SWE harness, 39.0% inside Antigravity. The threefold spread is Goog​le telling you that its best coding number depends on its own agent harness, and the harness-free number is the one a normal API caller would recognise.

• Multimodal knowledge work — Carbon launched with a rank on the new Multimodal Knowledge Agent benchmark (second, at more than 6x lower cost per task than the leader). Neon's announcement does not report an MKA figure.

• Web development — Carbon jumps 89 Elo points to fourth on WebDev Arena at 1768. No Neon WebDev Arena number is published.

• Cyber — Carbon's SEC-bench Pro 88.2%, ExploitGym 54.8% and CWE-bench 71.0% are the numbers Neon is measured against, and Neon's better figures (90.5%, and 70.7% on CWE-bench v2) come from the separate "Neon for cyber" variant that is still gated to Fairwind members.

The pattern is consistent: Neon's wins are on benchmarks that reward finishing a task efficiently under a fixed session, and Carbon's wins are on benchmarks that reward grinding a hard problem for a long time. That is not a marketing distinction, it is a different shape of compute spend, and it decides which one belongs in your agent loop.

What the two models are each built to be

Goog​le describes Neon as an orchestrator: "a fast, high-judgment Orchestrator directing parallel Hydrogen subagents in Goog​le Antigravity." That single sentence is more useful than any of the percentages. An orchestrator spends its life reading state, deciding what to do next, and delegating — many turns, modest output per turn, high sensitivity to latency. Carbon's documented profile is the opposite: a model that "run[s] for days to persist through dead ends," using more tokens than Argon "to explore thoroughly, reason deeply, and run extended verification loops."

If you are choosing between them for a single-shot hard problem — a subtle bug across a large codebase, an open-ended research question, a security review where missing the flaw is the failure mode — Carbon is the model whose stated design matches the job, and the December price step is a budgeting problem rather than a reason to switch.

If you are choosing for a pipeline that runs thousands of times a day and needs each run to close, Neon's design matches that job, and its flat rate card is a real advantage over a model whose price doubles in twelve days. The expensive mistake is treating "newer" as "better" and moving a long-horizon workload onto a model Goog​le itself places below Carbon on depth.

The OrcaRouter model page for Gemini 4 Carbon, showing the identifier google/gemini-4-carbon, a 1,048,576-token context window, multimodal input across text, image, audio, video and documents, and the pricing panel reading $2.00 input and $10.00 output per million tokens with a 10.00 second p50 time to first token.

Running them together rather than choosing

The realistic answer for most production systems is both, and the reason is that the two models fail differently. Carbon fails by grinding: it will spend minutes and hundreds of thousands of tokens before it answers, which is correct on a hard problem and wasteful on an easy one. Neon fails by climbing: push xhigh at a task that does not need it and you get overthinking, truncated JSON and a bill that looks like Carbon's.

That is a routing problem before it is a model problem. In our catalogue, google/gemini-4-carbon is available today at provider list price with no markup — $2.00 and $10.00 per million tokens while the introductory rate holds, stepping to $4.00 and $20.00 on 1 December, with the step appearing in your bill the day it appears in Goog​le's. google/gemini-4-argon is there on the same terms. Neon is not in the catalogue yet, and this piece is not going to describe a fallback chain to a model we cannot serve; when it lands, it lands on the same pass-through terms.

What the routing layer buys you today is the ability to write the decision once instead of hardcoding it: a rule that sends classification and extraction to the cheap tier, sends anything that fails a first pass to Carbon, and caps thinking depth per route rather than per model. When Neon becomes an option, that rule gains a branch instead of a migration. And because the router fails over automatically, a step-change in one provider's price or availability does not take the pipeline down with it.

Which one to pick on 19 November

Pick Carbon if the work is hard and infrequent, if you need the published multimodal-knowledge or web-development numbers, or if you already have Carbon tuned and the December step is absorbable. Pick Neon if the work is high-frequency, tool-heavy, and measured in tasks completed rather than problems solved — and if you are willing to benchmark your own traffic at high before trusting a per-task claim from the model's own announcement. Pick neither, yet, if your workload is everyday execution at volume: Gem​ini 4 Hydrogen has been doing that since 12 November at $1.50 and $7.50 with a free tier, and nothing in either of these two launches changes that arithmetic.

The thing to actually watch is not the index gap, which is a point or two either way. It is whether Goog​le's 32%-cheaper-per-task claim survives independent measurement — and whether the 1 December doubling on Carbon pushes production traffic toward Neon faster than anyone planned for a model that launched thirteen days before the increase.

The OrcaRouter models catalogue filter rail, with controls for input modalities, context length, input price, status, access tier, series, supported parameters, privacy and inference region.