DEV Community

Cover image for Gemini 4 Neon vs Carbon: is the faster model good enough?
AIHubMix
AIHubMix

Posted on Edited on

Gemini 4 Neon vs Carbon: is the faster model good enough?

On the Artificial Analysis leaderboard, Gemini 4 Neon at its default high setting scores 59.6 on the Intelligence Index and finishes an average task in 14.8 minutes. Carbon High scores 60.0 and takes 36.6. If long Carbon reasoning runs are what hit your client timeouts or hold up your queue, Neon High matches Carbon High on the index in 40% of the time.

Which one costs less depends on the date. Neon is $3/$15 per million input/output tokens, 50% more per token than Carbon's introductory $2/$10. On December 1 Carbon's rates double to $4/$20, and Neon becomes 25% cheaper per token. Per completed task at high, the two cost about the same today, and Neon costs about half from December 1.

Carbon still leads on research-level physics and has the lower hallucination rate. Google keeps it for the cases where "a faster model gets stuck".

What changes when you swap the model ID

Google's model card states the relationship in one line: "Gemini 4 Neon is adapted from Gemini 4 Carbon." The model ID is gemini-4-neon, the same on AIHubMix. Google lists it as stable, released November 18, 2026, with a March 2026 knowledge cutoff, a 1,048,576-token context window shared by input and output, and text, image, video, audio and PDF input. Neither model is on Google's free tier.

Item Neon Carbon
Price, input / output $3 / $15 $2 / $10, then $4 / $20
Thinking levels low, medium, high, xhigh low, medium, high
Default level high medium
Output per request 65,536 65,536
Native non-streaming request 65,536 32,768
Output speed (AA) about 100 tok/s about 55 tok/s

Prices are Google's, in USD per million tokens; Carbon's $2/$10 runs through November 30, 2026, and $4/$20 applies from December 1. The non-streaming row is the output a single non-streaming REST call returns on Google's own API before it needs a continuation. Output speed is Artificial Analysis's measurement of Google's API at the high setting.

The default level is one step higher. Carbon defaults to medium; Neon defaults to high. A Carbon integration that never set a level gets more thinking per turn after the swap. On quality that's the comparable setting, but it isn't the same token budget.

Non-streaming answers can run twice as long. A direct non-streaming call to Carbon returns at most 32,768 output tokens, and Google's long-decode guide says that setting max_output_tokens above 32,768 on such a call "returns a 400 error". Neon allows 65,536 either way. AIHubMix's parameter schema records the same split: Carbon at 65,536 streaming and 32,768 non-streaming, Neon at 65,536 for both. A Carbon integration that kept non-streaming calls under 32K needs no change, and after the move it can raise the cap.

There is a fourth level. xhigh is new for this tier; Carbon tops out at high. It only applies when a request asks for it, and Google's advice is to "reserve xhigh for hard, multi-step problems where high falls short."

Everything else is the same family design: one unified 1M window, implicit caching only, ignored sampling parameters and the same request schema. Google's checklist for moving from another Gemini 4 model has three lines: change the model ID, check the thinking level, use a paid project.

Google places Neon between the fast and the deep model. The AI Studio notes say it "sits between Gemini 4 Hydrogen and Gemini 4 Carbon for high-frequency complex tasks." The model-selection guide says Neon "finishes faster with fewer output tokens than Carbon while handling harder reasoning than Hydrogen," and still calls Carbon "our strongest reasoning model". Google's speed labels for the two are "Moderate" for Neon and "Deliberate" for Carbon.

Running both models on one prompt

Both models are on AIHubMix's OpenAI-compatible endpoint, so comparing them is a loop over two model IDs. AIHubMix's parameter schema lists reasoning_effort values from low to xhigh for Neon and from low to high for Carbon, so high is the top value both accept.

import os
from openai import OpenAI

client = OpenAI(
    base_url="https://aihubmix.com/v1",
    api_key=os.environ["AIHUBMIX_API_KEY"],
)

for model in ["gemini-4-neon", "gemini-4-carbon"]:
    response = client.chat.completions.create(
        model=model,
        reasoning_effort="high",  # Neon also accepts "xhigh"; Carbon tops out at "high"
        max_completion_tokens=32768,  # Carbon's non-streaming limit; Neon allows 65,536
        timeout=900,  # hard prompts can take several minutes
        messages=[{"role": "user", "content": "Find the race condition in this retry queue and propose a fix: ..."}],
    )
    print(model, response.usage.completion_tokens)
    print(response.choices[0].message.content[:500])
Enter fullscreen mode Exit fullscreen mode

The cap sits at 32,768 because the call doesn't stream, and that's Carbon's non-streaming limit; stream the response to go up to 65,536 on Carbon. AIHubMix's OpenAI-compatible endpoint can't continue past 65,536. Its schema lists Google's continuation token on the native Gemini endpoint, a path this post didn't test. The loop prints completion_tokens per model because the two spend very different numbers of tokens on the same job.

No live test was run for this post, so the sample shows the request shape, not a verified response. The compare page puts the two model listings side by side.

Score, time per task and first token

The figures in this post come from Artificial Analysis (AA): its live leaderboard and model pages, and its coding-agent index, all re-checked 2026-11-19 at 13:27 UTC. They are third-party measurements of Google's API and change as AA re-runs models.

Setting AA Index Min/task First answer (s)
Neon Xhigh 61.4 21.5 56.5
Carbon High 60.0 36.6 72.9
Neon High (default) 59.6 14.8 40.6
Carbon Medium (default) 58.3 29.3 32.2
Neon Medium 54.9 8.3 20.1
Carbon Low 54.2 29.8 8.7
Neon Low 44.1 3.7 13.9

"First answer" is AA's time to first answer token. "Min/task" is AA's average time per Intelligence Index task.

Output speed explains most of the time gap: AA measures Neon at about 100 tokens per second and Carbon at about 55. Neon's default scores 1.3 points above Carbon's default, and Neon Xhigh, at 61.4, is the highest Gemini 4 setting on the index. Google's launch post calls that "#3 overall". It holds when each model counts once, since only Claude Fable 5.2 and GPT-6.2 Astra score higher; AA ranks every setting separately, and there Neon Xhigh is sixth.

The first answer token runs on a different clock. Neon High cuts Carbon High's 72.9-second wait to 40.6, but Carbon's default Medium starts in 32.2 seconds, sooner than Neon High. For chat-style use, compare Carbon Medium with Neon Medium: 20.1 seconds to the first answer and 8.3 minutes per task, at 3.4 points lower on the index.

The index average hides where the two differ. On AA's individual evaluations, Neon High beats Carbon High on FrontierCode (43.3 vs 40.6), Terminal-Bench-Science (65.7 vs 64.8), AutomationBench (79.2 vs 77.8) and the AA-Briefcase knowledge-work rating (1632 vs 1611). Carbon High leads on CritPt, AA's research-level physics evaluation (69.0 against 59.0, and no Neon setting scores above 62), on Terminal-Bench 4.0 (65.2 against 61.6), on the non-hallucination rate in AA's Omniscience test (88.6% against 86.1%, and 95.8% for Carbon Medium) and narrowly on HLE-Diamond (69.1 vs 68.4).

AA's coding-agent leaderboard runs each model inside a coding agent; Google's models run in its Antigravity CLI. There, Neon Xhigh is first at 58.6, ahead of Carbon High at 55.5. Its lead over second place, Claude Code with Fable 5.2 Xhigh, is less than 0.2 points: a tie at the top. Google's model card doesn't settle the comparison either, because its table sets Neon against Claude Opus 5.5 and GPT-6.2 Sol, not against Carbon.

Cost per task before and after December 1

Model and period Input Output Cached AIHubMix today
Neon $3.00 $15.00 $0.15 same
Carbon, to Nov 30 $2.00 $10.00 $0.10 same
Carbon, from Dec 1 $4.00 $20.00 $0.20 not published

Google's Gemini API pricing, USD per million tokens, standard tier. Output includes thinking tokens. Neon has one rate across the full 1M window and no introductory period.

AIHubMix lists both models at Google's current rates. It hasn't published a Carbon price for December, so check the Carbon model page then.

Per token is the wrong unit for these two, because they spend very different numbers of tokens on the same job. AA's cost per Intelligence Index task is a better measure.

Setting $/task to Nov 30 $/task from Dec 1
Neon Medium 1.90 1.90
Neon High (default) 3.74 3.74
Neon Xhigh 6.44 6.44
Carbon Medium (default) 3.29 6.59
Carbon High 3.88 7.76

AA's cost per task in USD, same snapshot. AA prices Carbon at its December rate ($4/$20). Every Carbon rate is exactly half that until November 30, so the two Carbon figures in the first column are AA's numbers halved: computed estimates, not AA measurements.

Until November 30, Neon High and Carbon High cost about the same per task, $3.74 against about $3.88, and Carbon Medium, at about $3.29, is cheaper than Neon High. Before December, Neon saves time, not money. From December 1, Neon High costs 52% less per task than Carbon High and 43% less than Carbon Medium, at a similar or higher score.

Google's launch post claims "a 32% lower cost per task than Carbon" at Neon's default level. Today's AA numbers give 43–52% after December 1, and anywhere from 3.5% cheaper to 14% dearer before it. Treat Google's figure as Google's own estimate.

One agent turn, priced by hand

Take an agent turn with a 200,000-token prompt. 150,000 of those tokens are a stable prefix that hits the implicit cache, and the model writes 20,000 tokens of thinking and answer.

  • Neon: (50,000 × $3 + 150,000 × $0.15 + 20,000 × $15) per million = $0.47.
  • Carbon until November 30: $0.32.
  • Carbon from December 1: $0.63.

That holds the token counts equal, which flatters Carbon against Neon High but not against Neon Xhigh. AA's per-task figures imply that at equal prices, Neon High would cost about 36% less than Carbon High on the same tasks, and Neon Xhigh about 11% more. AA's model pages say Neon Xhigh generated 283 million tokens across the Intelligence Index and Carbon High 205 million, against a 65.1 million median.

Your own number depends on two settings: the thinking level you send, and whether your stable prefix clears the 4,096-token minimum for implicit caching.

Failure modes after the switch

Thinking you didn't ask for. A Carbon integration that never set a thinking level runs at high on Neon. That matches Carbon High on the index and costs twice as much per task as Neon Medium. Set the level on purpose for each call type.

A 400 when a request goes back to Carbon. If one code path serves both models, a non-streaming request with max_output_tokens above 32,768 is fine on Neon up to 65,536, and Google documents a 400 for it on Carbon. Keep non-streaming Carbon calls at 32,768, as the sample does, or stream.

A slower first token at the default. A chat surface that moves from Carbon Medium to Neon's default high waits longer for the first answer token: 40.6 seconds against 32.2. Send medium for interactive turns. It starts answering 12 seconds sooner than Carbon Medium, at 3.4 points lower on the index.

A smaller saving in agent loops. On AA's coding-agent leaderboard, Neon Xhigh averaged 30.5 minutes per task against 33.6 for Carbon High, about 9% less. Agent runs also spend time executing tools, which output speed doesn't change. AA puts those runs at $6.86 and $9.75 per task, pricing Carbon at its December rate; at today's rate Carbon High would be about $4.88, cheaper than Neon Xhigh. Measure your own tool-heavy loops before planning around the index numbers.

A 65,536-token ceiling per request. Through AIHubMix's OpenAI-compatible endpoint the limit is 65,536 output tokens per request, and 32,768 for non-streaming Carbon calls. On Google's own API, Neon can go further through long decode in the Interactions API or GenerateContent.

More made-up answers than Carbon. On AA's non-hallucination rate, Neon High scores 86.1% against 88.6% for Carbon High and 95.8% for Carbon Medium. If a confident wrong answer costs more than a slow right one, that gap matters more than the index.

Where Carbon is still the better model

Keep Carbon for research-grade reasoning and security audits. CritPt is the clearest gap, 69.0 for Carbon against 59.0 for Neon High, and Google's guidance points the same way: use Carbon for "system architecture, multi-file debugging, dense multimodal analysis, formal proofs, and security audits, or when a faster model gets stuck." You can also run both, with Neon for the bulk of complex work and Carbon for the cases Neon gets wrong.

Carbon Medium also wins this month's bill. Until November 30 it costs less per task than Neon High, so if your evaluations pass on Carbon Medium and nobody is waiting on the answers, staying until the price change costs nothing. Carbon Low returns its first answer token in 8.7 seconds, the fastest of any setting here, and scores 54.2, but it still takes 29.8 minutes per task, so it helps only where the first token is the metric.

If cost per task drives the decision, run your own evaluation set on Neon High and on Carbon in the last week of November, and move on the day the prices change. If timeouts drive it, switch now and start at Neon's default high: it matches Carbon on the index in half of Carbon Medium's time per index task.

What this comparison doesn't verify

No live test was run for this post; the code shows the request shape. Every measurement is AA's snapshot of Google's API from 2026-11-19 at 13:27 UTC, and AA re-runs models. The pre-December Carbon costs per task are AA's figures halved, computed estimates rather than measurements. "#3 overall" and "a 32% lower cost per task" are Google's claims. AIHubMix hasn't published its December price for Carbon, and the continuation token its schema lists on the native Gemini endpoint is untested here.

Keep reading: the Gemini 4 Neon series

Sources

Top comments (0)