GPT Image 2.5 is live — OpenAI's newest image model, targeted edits that leave the rest of the frame alone
All resources

· reAPI Team

Gemini 4 Neon vs Carbon, Hydrogen, and Argon: Which to Use

Compare Gemini 4 Neon vs Carbon, Hydrogen, and Argon across pricing before and after Dec 1, 2026, four thinking levels, 64K vs 32K REST hops, and routing.

Gemini 4 Neon vs Carbon, Hydrogen, and Argon: Which to Use

Comparing Gemini 4 Neon vs Carbon, Hydrogen, and Argon comes down to three concrete numbers that change how you spend tokens in production: per-million rates, default thinking depth, and per-request REST output limits. Google launched Gemini 4 Neon (gemini-4-neon) on November 18, 2026, six days after Hydrogen (gemini-4-hydrogen), completing a four-model Gemini 4 lineup whose element names reveal nothing about cost or routing order[1][2].

Until November 30, 2026, Google's promotional pricing creates a brief inversion: Gemini 4 Carbon and Argon bill $2.00 per million input tokens and $10.00 per million output tokens, which is lower per token than Neon's $3.00 / $15.00 standard rate[3]. Starting December 1, 2026, Carbon and Argon double to $4.00 / $20.00, placing Neon cleanly in the middle between Hydrogen ($1.50 / $7.50) and Carbon ($4.00 / $20.00)[3]. This guide maps every verified specification across all four models so you can pick the right default without guessing.

TL;DR

  • Pick Gemini 4 Neon (gemini-4-neon) for high-frequency complex software engineering, long tool-heavy agent sessions, quantitative knowledge work, and defensive code hardening where Hydrogen lacks depth and Carbon is heavier than necessary[1][4].
  • Pick Gemini 4 Carbon (gemini-4-carbon) for your hardest long-horizon architecture, science, mathematics, and security problems; its standard Google paid rate rises from $2.00/$10.00 to $4.00/$20.00 per 1M tokens on December 1, 2026[3].
  • Pick Gemini 4 Hydrogen (gemini-4-hydrogen) for fast everyday coding, interactive UI iteration, and high-volume worker or verifier sub-agents at $1.50/$7.50 per 1M tokens, plus a Google AI Studio free tier[3][5].
  • Pick Gemini 4 Argon (gemini-4-argon) if you already run an established long-horizon workflow tuned to Argon's lower per-turn thinking spend relative to Carbon; it shares Carbon's $2.00/$10.00-to-$4.00/$20.00 schedule and 32K JSON REST hop limit[3][6].
  • Zero schema migration across all four models: All four share the same 1,048,576-token unified context window, implicit-only caching with a 95% discount on cache hits, and Google's native GenerateContent request format, though thinking defaults (high vs medium) and non-streaming REST caps (65,536 vs 32,768) differ[4][6][7].

Where Gemini 4 Neon fits in the four-model lineup

Google adapted Gemini 4 Neon directly from Gemini 4 Carbon rather than scaling up Hydrogen[8]. According to Google's model card and migration guide, Carbon stays the heaviest option for deep, long-horizon reasoning, Hydrogen handles rapid everyday execution, Argon serves as a more token-efficient frontier sibling to Carbon, and Neon bridges Hydrogen and Carbon for complex tasks you run repeatedly every day[1][3][8].

When teams evaluate Gemini 4 Neon vs Carbon, the practical difference shows up in multi-step tool sessions and total completion overhead. Google designed Neon to finish repository-level debugging and multi-file edits in fewer steps while cutting redundant file reads and handling context compaction cleanly across long tool chains[4]. That makes Neon a natural primary model for coding agents, spreadsheet and PDF synthesis, and vulnerability patching, whereas Carbon stays reserved for problems where you willingly pay for deeper single-turn deliberation[1][8].

SpecificationGemini 4 HydrogenGemini 4 NeonGemini 4 ArgonGemini 4 Carbon
Stable model IDgemini-4-hydrogengemini-4-neongemini-4-argongemini-4-carbon
Release dateNov 12, 2026Nov 18, 2026Oct 2026Nov 10, 2026
Documented roleFast everyday coding & worker agentsHigh-frequency complex coding, agents & scienceEfficient long-horizon frontier reasoningHardest architecture, math, science & security
Inputs / outputText, image, video, audio, PDF / textText, image, video, audio, PDF / textText, image, video, audio, PDF / textText, image, video, audio, PDF / text
Unified context window1,048,576 tokens1,048,576 tokens1,048,576 tokens1,048,576 tokens
Default thinking levelhighhighmediummedium
Supported thinking levelslow, medium, high, xhighlow, medium, high, xhighlow, medium, highlow, medium, high
Non-streaming REST hop cap65,536 tokens65,536 tokens32,768 tokens32,768 tokens
Streaming SSE hop cap65,536 tokens65,536 tokens65,536 tokens65,536 tokens
Context cachingImplicit only (min 4,096 tokens)Implicit only (min 4,096 tokens)Implicit onlyImplicit only
Google API free tierAvailable (with limits)Paid tier onlyPaid tier onlyPaid tier only
Knowledge cutoffMarch 2026March 2026Documented per model cardDocumented per model card

Source: Google AI Studio model specifications, long-decode guide, and DeepMind model cards, checked November 20, 2026[2][5][6][8].

All four models reject image generation, audio generation, and the Live API[2][5]. Every token the model processes counts against the same 1,048,576-token unified pool, meaning prompt input, thinking tokens, tool calls, and final answer text share one ceiling rather than stacking separately[6].

Gemini 4 Neon pricing before and after December 1, 2026

Pricing is the single biggest surprise when comparing Gemini 4 Neon vs Carbon in late November 2026. All four Gemini 4 models use flat per-token pricing across the entire 1M context window on Google's paid tier, with no long-context surcharge threshold and a 95% discount on implicit cache hits[3]. Every output rate includes thinking tokens[3].

ModelGoogle Standard paid tier (through Nov 30, 2026)Google Standard paid tier (from Dec 1, 2026)Cached input (from Dec 1, 2026)Live reAPI group rates
Gemini 4 Hydrogen$1.50 in / $7.50 out$1.50 in / $7.50 out$0.075 / 1MHydrogen pricing
Gemini 4 Neon$3.00 in / $15.00 out$3.00 in / $15.00 out$0.15 / 1MNeon pricing
Gemini 4 Argon$2.00 in / $10.00 out$4.00 in / $20.00 out$0.20 / 1MArgon pricing
Gemini 4 Carbon$2.00 in / $10.00 out$4.00 in / $20.00 out$0.20 / 1MCarbon pricing

Source: Google official Standard paid-tier rates checked November 20, 2026[3]. Grounding with Google Search or Google Maps carries separate per-query charges after shared monthly allowances ($14 per 1,000 queries on Google's paid tier)[3].

Notice two practical consequences of that schedule:

  1. Before December 1, 2026, Carbon and Argon carry a lower per-token sticker price ($2.00/$10.00) than Neon ($3.00/$15.00), so your real-world cost comparison between Gemini 4 Neon vs Carbon depends on how many thinking tokens each model consumes to finish the same job[3]. Because thinking counts as output, a model that finishes a refactor in half the output tokens can cost less per task even at a higher per-token rate.
  2. From December 1, 2026 onward, Neon ($3.00/$15.00) sits 25% below Carbon and Argon ($4.00/$20.00) and double Hydrogen ($1.50/$7.50)[3]. On reAPI, each model exposes live gateway group multipliers on its model page: for example, gemini-4-neon published in group gemini-0.4 at a 0.55x ratio on November 20, 2026, corresponding to $1.65 input, $8.25 output, and $0.0825 cached input per 1M tokens — check Gemini 4 Neon on reAPI for live group availability and current rates.

Gemini 4 Neon thinking levels and the 32K vs 64K REST hop split

Two API defaults separate the newer November pair (Neon and Hydrogen) from the earlier pair (Carbon and Argon): thinking configuration and single-request REST output limits.

First, thinking cannot be turned off on any Gemini 4 model[7]. Carbon and Argon expose three native levels (low, medium, high) and default to medium. Neon and Hydrogen expose four native levels (low, medium, high, xhigh) and default to high[4][7]. Sending native thinkingLevel: "minimal" returns an upstream error on all four models, while the OpenAI-compatible /api/v1/chat/completions endpoint maps reasoning_effort: "minimal" to low and preserves xhigh on gemini-4-neon and gemini-4-hydrogen only[9].

Second, direct non-streaming REST calls (:generateContent) default to and cap at 65,536 tokens per hop on Neon and Hydrogen, compared to 32,768 tokens per hop on Carbon and Argon[6]. Streaming calls (:streamGenerateContent?alt=sse) allow 65,536 tokens per hop across all four models[6]. For tasks requiring more than one hop, Google's latest GenAI SDKs (google-genai>=2.30.1 for Python and @google/genai>=2.29.1 for JavaScript) enable automatic_continuation by default, whereas direct REST callers and reAPI execute strictly one request per call and return an opaque continuationToken when finishReason is CONTINUATION[4][6].

That hop difference matters directly when evaluating Gemini 4 Neon vs Carbon over REST. If your service sends non-streaming JSON requests with high thinking, Carbon hits its 32,768-token hop ceiling twice as early as Neon, forcing an extra continuation round trip (and a second request billed against the repeated context) sooner on long code generations.

How to switch between Gemini 4 Neon, Carbon, Hydrogen, and Argon in code

Switching between any two Gemini 4 models requires changing only the model ID in the URL and reviewing thinkingLevel and maxOutputTokens[4]. Here is a single-shot native request targeting reAPI's gemini-4-neon endpoint with x-goog-api-key:

curl https://reapi.ai/api/v1beta/models/gemini-4-neon:generateContent \
  -H "x-goog-api-key: YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "contents": [{
      "role": "user",
      "parts": [{"text": "Audit this retry queue for race conditions and propose a minimal patch."}]
    }],
    "generationConfig": {
      "maxOutputTokens": 65536,
      "thinkingConfig": {"thinkingLevel": "high"}
    }
  }'

When migrating an existing app from Gemini 3 or testing Gemini 4 Neon vs Carbon side by side, watch four traps documented in Google's migration checklist[4]:

  1. Drop low maxOutputTokens caps: Legacy values like 4096 or 8192 frequently truncate answers because thinking tokens consume the same budget as visible text. Leave maxOutputTokens omitted (65,536 on Neon and Hydrogen) or set an explicit higher ceiling, and lower thinkingLevel when you want shorter runs.
  2. Remove cachedContent: None of the Gemini 4 models support explicit context caching resources. Implicit caching runs automatically on prompts of at least 4,096 tokens[4].
  3. Remove temperature, topP, topK, and candidateCount: In-range custom sampling values are ignored upstream, while out-of-range values return HTTP 400[4].
  4. Preserve thoughtSignature and continuationToken verbatim: On multi-turn function calling, send back the model's exact parts array rather than rebuilding assistant turns from text alone. On CONTINUATION, resend the original contents unchanged with the returned token[6].

See the Gemini 4 Neon API docs for complete Python, Node.js, Go, streaming SSE, and group-selection examples.

Multi-model routing patterns that cut token waste

You rarely need one model for every step of a pipeline. Google's documentation highlights pairing a stronger planner or orchestrator with Hydrogen workers, and the four-model lineup maps cleanly onto three routing topologies[4][5]:

  • Default production coding and agents (Neon at high, dropping to medium for interactive edits): Start complex refactors, test debugging, security patching, and multi-document analysis on Gemini 4 Neon. Escalate only failed or high-stakes tasks to Neon xhigh or Gemini 4 Carbon.
  • Orchestrator and worker swarms (Neon or Carbon planning with Hydrogen execution): In a Gemini 4 Neon vs Hydrogen split, let Neon or Carbon break a specification into bounded file tasks, then dispatch independent module edits, unit-test drafts, and schema extractions to Gemini 4 Hydrogen at half Neon's list token rate.
  • High-volume extraction and classification: Skip Gemini 4 reasoning overhead altogether and keep a lightweight flash tier for simple routing or tagging, reserving Gemini 4 models for tasks where mandatory thinking improves accuracy[4].

Before committing to a default in production, run 10 to 20 representative prompts from your own workload across Hydrogen (high), Neon (medium and high), and Carbon (medium and high). Measure pass rate, wall-clock latency, and total billed tokens per completed task rather than raw per-million list rates alone.

Frequently asked questions

How does Gemini 4 Argon compare to Neon, Carbon, and Hydrogen?

In a Gemini 4 Neon vs Argon comparison, Google positions Gemini 4 Argon (gemini-4-argon) as an efficient frontier reasoning model that spends fewer thinking tokens per turn than Carbon while sharing Carbon's $2.00/$10.00-to-$4.00/$20.00 pricing schedule, three thinking levels (low/medium/high, default medium), and 32K non-streaming REST hop cap[3][6]. Compared to Argon, Neon adds the xhigh thinking tier, doubles the non-streaming REST hop cap to 65,536 tokens, and sits 25% lower per token once Carbon and Argon move to $4.00/$20.00 on December 1, 2026[3][4].

Is Gemini 4 Argon still good after Neon's launch?

Yes, if your existing prompts are already evaluated on gemini-4-argon and you rely on its lower per-turn reasoning behavior compared to Carbon, it remains a stable 1M-context model[3]. For new workloads starting in late November 2026, most teams should benchmark Neon and Hydrogen first because both offer 64K non-streaming REST hops, four thinking levels including xhigh, and lower post-December-1 token rates[3][4].

Who has access to Gemini 4 Argon and Gemini 4 Neon?

On Google's own platform, both Gemini 4 Argon and Gemini 4 Neon are restricted to the Gemini API paid tier with no Google free tier, while Neon also rolls out to Google AI Pro and Ultra subscribers in Google AI Studio, Google Antigravity, Gemini for business, and the Gemini app[1][3]. Developers can also call all four models pay-as-you-go through reAPI using one API key and credit balance without a separate Google billing account; see each model page for live group availability.

How to use Gemini 4 Carbon alongside Neon or Hydrogen?

Because Carbon, Neon, Hydrogen, and Argon share the exact same native GenerateContent request schema, you can route between them by changing only the URL model segment (/api/v1beta/models/gemini-4-carbon:generateContent vs /api/v1beta/models/gemini-4-neon:generateContent) and checking thinkingLevel (medium default on Carbon vs high default on Neon and Hydrogen)[4][7]. See the Gemini 4 Carbon API docs for full native request details.

Does Gemini 4 Neon have a free tier?

No. Google's official pricing table lists Gemini 4 Neon as "Not available" on the Free tier[3]. Among the four Gemini 4 models, only Gemini 4 Hydrogen offers a limited Google AI Studio free tier[3]. On reAPI, all four models bill pay-as-you-go from your credit balance with no monthly subscription.

Should I use Gemini 4 Neon xhigh instead of Carbon?

Treat xhigh as an escalation setting rather than a default. Google advises starting Neon at high and reserving xhigh for hard multi-step engineering, scientific, or analytical tasks where high falls short, because xhigh increases both latency and thinking-token spend[4]. Compare Neon xhigh directly against Carbon medium and high on your failing cases and keep whichever delivers higher pass rates per dollar.

Can any Gemini 4 model generate images or audio?

No. All four models (Neon, Carbon, Hydrogen, and Argon) accept text, image, video, audio, and PDF inputs, but output text only[2][5]. Use dedicated image, video, or audio generation models when your application needs generated media.

Picking your default Gemini 4 model

If you are deciding between Gemini 4 Neon vs Carbon, Hydrogen, and Argon today, use a straightforward three-step rule: start everyday high-frequency coding, tool agents, and document analysis on Gemini 4 Neon at high; offload high-volume bounded sub-tasks to Gemini 4 Hydrogen; and escalate only the hardest long-horizon problems to Gemini 4 Carbon. Watch the December 1, 2026 rate change if you currently run Carbon or Argon at promotional pricing[3], and test your actual prompt suite across thinking levels before locking production routing.

References

  1. Google Blog, Introducing Gemini 4 Neon (Nov 18, 2026): https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-neon/
  2. Google AI Studio, Gemini 4 Neon Model Specifications (retrieved Nov 20, 2026): https://aistudio.google.com/docs/models/gemini-4-neon
  3. Google AI Studio, Gemini API Pricing — Carbon, Argon, Neon, and Hydrogen (retrieved Nov 20, 2026): https://aistudio.google.com/docs/pricing#gemini-4-neon
  4. Google AI Studio, What's new in Gemini 4 Neon (retrieved Nov 20, 2026): https://aistudio.google.com/docs/whats-new-gemini-4-neon?generate-content=true
  5. Google AI Studio, Gemini 4 Hydrogen Model Specifications (retrieved Nov 14, 2026): https://aistudio.google.com/docs/models/gemini-4-hydrogen
  6. Google AI Studio, Long decode guide (retrieved Nov 20, 2026): https://aistudio.google.com/docs/long-decode?generate-content=true
  7. Google AI Studio, Gemini thinking guide (retrieved Nov 20, 2026): https://aistudio.google.com/docs/thinking?generate-content=true
  8. Google DeepMind, Gemini 4 Neon Model Card (Nov 18, 2026): https://deepmind.google/models/model-cards/gemini-4-neon/
  9. Google AI Studio, OpenAI compatibility — Thinking (retrieved Nov 20, 2026): https://aistudio.google.com/docs/openai

More Posts

All resources