Gemini 4 Neon vs Carbon

Pick a Gemini default for coding agents and knowledge work. Compare access, benchmarks and the December price change, then try a matched task on each model.

Published

Tools
  • GGemini 4 Neon
  • GGemini 4 Carbon
Gemini 4 Neon vs Carbon

The choice in Gemini 4 Neon vs Carbon turns on how you check the answer. Pick Neon for coding and knowledge work with a clear acceptance test, and keep Carbon for problems that take several rounds of investigation. In early evaluations quoted in Google’s Neon launch post, Glean found Neon matched or exceeded Carbon’s overall quality at 3× the speed and 60% of the cost.

Gemini 4 Neon vs Carbon: which should you pick?

Pick Gemini 4 Neon when a developer is waiting for a patch they can review, or an analyst needs a memo they can check against the source documents. These jobs can require considerable judgment. What makes Neon a good default is that you can tell when the work is done. For a change to an invoice export, you can inspect the callers, run a regression test and check the resulting file. On a Google AI Pro subscription, Neon is also the model of this pair you can select in the Gemini app.

Keep Gemini 4 Carbon when the expensive part is figuring out what went wrong. Suppose the export passes its tests but sometimes double-counts an invoice after a retry. A patch that removes the duplicate row might hide a second payment attempt. Give Carbon the event history, the relevant code and the requirement that must hold across retries. Its job is to distinguish the competing explanations and show which one fits the evidence. Google describes this sustained verification as Carbon’s strength. Google’s guide to choosing a model

Verified November 20, 2026 against Google’s model specifications, launch posts and standard US Gemini API prices. A token is a unit of text or other model input or output. The context window is the budget shared by the request, reasoning and response.

Decision pointGemini 4 NeonGemini 4 Carbon
Stable API model IDgemini-4-neongemini-4-carbon
Unified context window1,048,576 tokens1,048,576 tokens
API thinking levelLow, medium, high (default), xhighLow, medium (default), high
Input per million tokens$3.00$2.00 through Nov 30, then $4.00
Cached input per million tokens$0.15$0.10 through Nov 30, then $0.20
Output per million tokens, including thinking$15.00$10.00 through Nov 30, then $20.00
Gemini app, chats and tasksGoogle AI Pro or UltraGoogle AI Ultra
Google AI StudioPaid API project, or try with Pro or UltraPaid API project, or try with Ultra
Gemini APIPaid tierPaid tier
Google AntigravityPro or UltraUltra
Gemini for businessAPI or model dropdownAPI or model dropdown

For an app user, the subscription decides which options appear in the picker. For a developer, a paid API project is a separate route to both models. Carbon’s October announcement described limited early access. Google opened it to paying developers and enterprises on November 10. Carbon availability announcement

Which model costs less for the same work?

Carbon costs less per token through November 30. Neon costs less per token starting December 1. A finished job can reverse the November result if Neon uses fewer billed tokens to deliver an acceptable answer. Google reports a 32% lower cost per Artificial Analysis Intelligence Index task for Neon at its default high thinking level versus Carbon. That figure describes the evaluation workload. Glean’s separate 3× speed finding describes its own early evaluations.

Consider a batch of accepted jobs using one million uncached input tokens, one million cached input tokens and 100,000 billed output tokens, including reasoning. At standard rates, Neon costs $3.00 + $0.15 + $1.50 = $4.65. Carbon costs $3.10 in November and $6.20 from December 1. This is a fixed-volume calculation before tool fees. Both models use implicit caching, so count a cache hit only when the usage record shows one. Gemini Developer API pricing

The same token workload costs $3.10 on Carbon in November, $4.65 on Neon and $6.20 on Carbon from December.
Holding token use fixed isolates the effect of Carbon’s scheduled price change.

Through November, every Neon rate is 1.5 times its Carbon counterpart. At the same mix of input, cached input and output, Neon breaks even if it finishes on two-thirds as many tokens. In practice, the mix can change as well. If a model saves cheap cached reads but writes a much longer answer, price the actual buckets separately before changing the default.

What the benchmarks say about coding and knowledge work

Neon wins the frequent document and automation jobs on the closest published comparisons. The Vals Index v2.2 gives Neon 70.9% and Carbon 70.4%. On Chartography, which tests reading dense charts, the scores are 73.5% and 72.5%. Zapier’s AutomationBench puts Neon at 56.3% and Carbon at 53.9%. For someone compiling a weekly variance memo from a workbook and a PDF, these results make Neon the sensible first trial. The acceptance test is whether the memo cites the right cells and explains the discrepancy. Google DeepMind’s Neon model card

Google AI Studio’s Neon model specifications and supported capabilities.
Gemini 4 Neon

Carbon retains an edge on particular problems. On Agents’ Last Exam v1.1, Carbon scores 77.4% against Neon’s 75.8%. On GraphWalks, which asks the model to trace relationships across a long input, Carbon scores 96.2% against 95.2% on the same 256K-to-1M breadth-first search measure. That is a reason to try Carbon on a dependency investigation where missing a single link changes the conclusion. Google DeepMind’s Carbon model card

Google AI Studio’s Carbon model specifications and unified token limit.
Gemini 4 Carbon

For coding agents, the surrounding software matters. Google reports 12.8% on SWE-sweep for Neon in mini-SWE and 39.0% in Google Antigravity. The harness is the software that lets the model read files, take actions and check its work. Even a seemingly exact tie can mislead here. Both cards print 65.2% on Terminal-Bench, but Neon ran version 4.1 and Carbon ran 4.0. Their Artificial Analysis Intelligence Index launch scores also come from different versions, 61 on v5.0.1 for Neon at xhigh and 59 on v5.0 for Carbon.

When to give Neon parallel Hydrogen workers

In Antigravity, Neon can coordinate Gemini 4 Hydrogen subagents, workers that take bounded assignments in parallel. On the invoice export, give one worker the callers to trace and another the tests to inspect. Have Neon reconcile their findings before anyone edits the shared interface. If the next investigation depends on the result of the previous one, parallel workers buy little. Keep a single investigator on the case and escalate to Carbon when the proposed explanation keeps failing.

Neon assigns caller tracing and test inspection to separate Hydrogen workers and reconciles their findings.
Split work where the evidence can be gathered independently. Keep responsibility for the final change with one model.

The Neon guide covers its thinking settings, and the Carbon guide covers its reasoning profile. If the decision changes which tools can see your repository, the Antigravity and Claude Code comparison covers that separate choice.

Secure code and prompt injection

Neon has the stronger published result for generating secure code. Google says Neon beats Carbon across the easy, medium and hard tiers of AutoBaxBench, reaching 76% on hard. Neon scores 90.5% on SEC-bench Pro for finding existing vulnerabilities and 70.7% on Collinear’s CWE-bench v2 for patching. Google’s Neon launch post

Carbon more than doubled the secure-code generation success rate of Gemini 4 Argon on Google’s internal evaluation. It scored 88.2% on SEC-bench Pro. Its 71.0% patching score comes from CWE-bench v1, so the two launch scores measure different versions of that test. Google’s Carbon launch evaluations

Prompt injection is a tie on the named test. An indirect prompt injection is an instruction planted in material the agent reads, such as a document or repository file. Google reports 0.0% attack success for both models on Gray Swan’s indirect prompt injection benchmark. Neon reached that result at its highest thinking level with real-time safeguards. Google also reports that continued testing found room for improvement on Carbon. Keep an agent preparing an invoice memo on read-only access to the underlying records. Neither model needs permission to change a payment to explain it.

Run a matched-task test in an afternoon

Use closed work with a known outcome. One bounded coding ticket, one investigation with a misleading initial symptom and one document-based decision give you a small but informative sample. This is a proposed team exercise. All measured results above come from the named sources.

  1. Write down what earns a pass

    For the code change, require the regression test to pass and adjacent behavior to hold. For the investigation, require an explanation that accounts for the misleading symptom. For the memo, list the source cells or passages the reviewer should be able to trace. Hide the accepted patch and memo from both models.

  2. Replay the same starting state

    Give each model a clean copy of the same commit or document packet, identical instructions, tools, permissions and time limit. Pin both to high for the matched run. Record the setting, since leaving both on default compares different effort levels. Keep subagents off for this pass so the model is the change you are measuring.

  3. Blind the review and count the retries

    Remove model names from the outputs. Ask the ticket owner to mark pass or fail and record minutes spent checking or correcting each result. Log elapsed time to acceptance, uncached and cached input, billed output and tool charges. Divide all spend, including failed attempts, by the number of accepted results. If nothing passes, record that outcome directly.

  4. Make a decision for each kind of work

    Move the bounded jobs to Neon if it clears the same acceptance bar and reduces time or spend. Keep Carbon on investigations where it finds a cause Neon misses. When both fail, inspect what evidence was available before paying for another run. Test parallel Hydrogen workers separately after choosing the coordinator.

Switch from Carbon to Neon without losing your baseline

Change the API request’s model from gemini-4-carbon to gemini-4-neon in one workload on a paid project. Gemini 4 uses the same request schema across the pair. Pin thinking_level explicitly if you want to preserve your comparison. An omitted setting changes from medium on Carbon to high on Neon. For a client that manages its own generateContent history, pass prior model turns and their thought signatures back unchanged. Google’s Neon migration checklist

Keep the old model ID and effort setting together so you can restore both if accepted work gets slower or review time rises. Leave a long-running Carbon investigation on its current configuration through the next checkpoint. Teams coming from Gemini 4 Argon can use the same matched-task exercise. For a shortlist that extends beyond Google, compare the best AI models for coding.

How good is Gemini 4 Carbon?

Carbon is the stronger choice here for a hard investigation with several plausible causes. It leads Neon on Agents’ Last Exam v1.1 and the long-input GraphWalks measure. Use a case with a known root cause to see whether that advantage matters in your codebase.

Which Gemini model is most accurate?

It depends on what counts as a correct answer. Neon leads on Vals Index v2.2 and Chartography. Carbon leads on Agents’ Last Exam v1.1. A chart-based weekly brief and a multi-stage investigation can reasonably have different defaults.

How to use Gemini 4 Carbon?

Select Carbon in the Gemini app with Google AI Ultra, or set the model to gemini-4-carbon in a paid Gemini API project. Google also offers Carbon in AI Studio, Antigravity and Gemini for business. Check the access row for the surface you plan to use.

Get the AI Tools Map for Business Owners with the newsletter.

Published
Category
AI
Related Articles
ChatGPT Dots: What a Dot Is, Who Gets One and How to Start

ChatGPT Dots: What a Dot Is, Who Gets One and How to Start

Find out who can create a ChatGPT dot, where it can work between conversations, and how to set approvals, pause a project or delete its memories.Nov 20, 2026AI
GPT-6.2 vs Claude Fable 5.2

GPT-6.2 vs Claude Fable 5.2

Choose Astra, Fable or Sol for coding agents, long documents and daily work. See where their API rates diverge and which model your team can access.Nov 20, 2026AI
GPT 6.2 vs GPT 6.1

GPT 6.2 vs GPT 6.1

Which Sol workloads should move, when Terra is still the better buy, and what to change in your API requests before switching to GPT-6.2.Nov 20, 2026AI
How to Use GPT-6.2

How to Use GPT-6.2

Find GPT-6.2 by plan, enable Astra and Sol for your workspace, choose a model for your job, and make your first API call.Nov 20, 2026AI
Gemini Computer Use Mac

Gemini Computer Use Mac

Set up Gemini Computer use on your Mac, see which accounts qualify, and try prompts for shipping quotes, resale listings, presentations and creative work.Nov 19, 2026AI
Genie 4 Explained: Google's World Model and Who Can Use It

Genie 4 Explained: Google's World Model and Who Can Use It

Who can try Genie 4 in Project Genie, how live editing and audio work, and when creators and developers should use it for a prototype.Nov 19, 2026AI
Gemini 4 Neon: What It Does, What It Costs and How to Start

Gemini 4 Neon: What It Does, What It Costs and How to Start

Where Gemini 4 Neon fits beside Carbon, Hydrogen and Argon, who can use it, and how to compare it with your current model on a real coding task.Nov 19, 2026AI
Always-On Agents: Anthropic's Five Primitives, Shown Working

Always-On Agents: Anthropic's Five Primitives, Shown Working

An always-on agent mostly sleeps. Anthropic's five primitives wake it, carry its memory, guard its actions and deliver its work. Costs and limits included.Nov 19, 2026AI
Newsletter

One letter, every Sunday.Working systems, not hot takes.

Weekly. No spam. Unsubscribe anytime.