Gemini 4 NeonGoogle Gemini APILLM pricingAI modelsmigration

Gemini 4 Neon: Pricing, Breaking Changes and When It Beats Carbon

person
LaunchBoosts Research Desk·AI-assisted research
8 min read

Researched and drafted with AI assistance from the 6 public sources listed at the end of this article, then published after automated editorial checks. Spotted an error? Tell us at support@launchboosts.com.

Gemini 4 Neon: Pricing, Breaking Changes and When It Beats Carbon — LaunchBoosts

Google released Gemini 4 Neon on November 18, 2026. It costs $3 per million input tokens and $15 per million output tokens, has a 1M-token context window and no free API tier. For the next 12 days, Neon costs more per token than Gemini 4 Carbon. On December 1, Carbon's price doubles and Neon becomes 25% cheaper per token. Switching to Neon is more than a model-ID swap: several request parameters now return errors or are silently ignored.

Below are the verified prices, a cost comparison you can reuse, the breaking changes from Google's developer guide, and how to decide between Neon, Carbon and Hydrogen.

What Gemini 4 Neon is

Google's announcement places Neon in the middle of the Gemini 4 family. Carbon is for the hardest long-horizon work, Hydrogen for fast everyday execution, and Neon is "the balance of depth and efficiency." The DeepMind model card says Neon is adapted from Gemini 4 Carbon and aimed at complex coding, analytical reasoning and enterprise workflows.

Key specs from the model card and announcement:

  • Model ID: gemini-4-neon
  • Inputs: text, images, audio and video. Google says Neon handles spreadsheets, PDFs, charts, audio and video in one context.
  • Context: up to 1,048,576 tokens, shared between input and output
  • Output: up to 64K tokens by default for synchronous calls, or up to 1M via background "long decode"
  • Knowledge cutoff: March 2026
  • Thinking levels: LOW, MEDIUM, HIGH (default) and XHIGH

Where to get it: paid-tier Gemini API and Google AI Studio, Gemini for business and Gemini Enterprise, Google Antigravity for AI Pro and Ultra subscribers, Android Studio, and the Gemini app's model picker for Pro and Ultra subscribers. Google Cloud's developer guide also lists it on the Gemini Enterprise Agent Platform for verified enterprise customers in the global, us and eu locations. OpenRouter routes it through both Google AI Studio and Google Vertex.

Google also announced "Gemini 4 Neon for cyber," which it said would roll out to vetted defenders through its Fairwind Program "in the coming days." It is not generally available.

Gemini 4 Neon pricing vs Carbon, Argon and Hydrogen

These figures come from Google's Gemini API pricing page, in USD per million tokens on the paid tier. Output prices include thinking tokens.

Model Free tier Input Output Cached input Batch/Flex (in/out) Priority (in/out)
Gemini 4 Neon No $3.00 $15.00 $0.15 $1.50 / $7.50 $5.40 / $27.00
Gemini 4 Carbon (to Nov 30) No $2.00 $10.00 $0.10 $1.00 / $5.00 $3.60 / $18.00
Gemini 4 Carbon (from Dec 1) No $4.00 $20.00 $0.20 $2.00 / $10.00 $7.20 / $36.00
Gemini 4 Argon No Same as Carbon, including the Dec 1 doubling
Gemini 4 Hydrogen Yes $1.50 $7.50 $0.075 $0.75 / $3.75 $2.70 / $13.50

The pricing page says Carbon and Argon prices change to the higher rate "starting December 1, 2026" on every tier. Neon and Hydrogen show no scheduled changes.

Several third-party write-ups disagree about Neon and December 1. Some say Neon's price changes. It doesn't: Carbon's does. Neon stays at $3/$15. Its position against Carbon flips because Carbon's introductory price ends.

If you route through OpenRouter's EU in-region endpoint, its listing shows Vertex at $3.30/$16.50. That option requires a Business or Enterprise plan.

A worked example

Take a typical agent or RAG call: 20,000 input tokens and 3,000 output tokens (thinking included), with no caching.

Model Cost per call 100,000 calls/month
Hydrogen $0.0525 $5,250
Carbon (to Nov 30) $0.0700 $7,000
Neon $0.1050 $10,500
Carbon (from Dec 1) $0.1400 $14,000

This assumes every model produces the same number of tokens, which isn't true. Google says that at the default high thinking level Neon costs "32% less per task than Carbon," and that it cuts redundant file reads in agent loops. Per-task cost depends on how many thinking and tool-call tokens a model burns, so the per-token table is only a starting point. Run the same 50–100 real tasks through each candidate, log usage metadata, and compare cost per completed task. To model your own traffic mix, the LLM API cost calculator and token counter help you size prompts before you spend money.

What the benchmarks say, and what they don't

The model card compares Neon with Claude Opus 5.5 and GPT-6.2 Sol. All figures are reported by Google.

Benchmark Neon Opus 5.5 GPT-6.2 Sol
Artificial Analysis Index v5.0.1 61 59 57
Vals Index v2.2 70.9% 70.1% 66.1%
Terminal-bench 4.1 65.2% 65.8% 61.5%
FrontierCode 2.0 46.2% 47.3% 43.8%
AutomationBench 56.3% 42.5% 41.3%
GraphWalks (256k–1M, F1) 95.2% 66.8% 80.1%
Chartography 73.5% 66.3% 75.1%

What to take from the table:

  • Coding is close to a tie. Opus 5.5 is slightly ahead on Terminal-bench 4.1 and FrontierCode 2.0. Neon's large lead on SWE-sweep (39.0% vs 10.1%) uses Google's own Antigravity harness. The announcement gives 12.8% for Neon on the neutral mini-SWE harness, so treat the 39% as a harness-plus-model result.
  • Long context is the clearest gap. GraphWalks at 256k–1M tokens is where Neon separates most from the competition. If your product stuffs large codebases or document sets into context, that's the claim worth testing first.
  • Automation and tool use look strong. Google reports first place on Zapier's AutomationBench (56.3%) and on Andon Labs' Vending-Bench 2.
  • Google's own "#3" caveat. The announcement says Neon ranks third overall on the Artificial Analysis Intelligence Index, "just behind larger, slower models."

One early independent read comes from a blog post analysing the Agent Duel leaderboard. It measured Neon at about 127 tokens per second on OpenRouter, against about 46 for Carbon. Its model projects Neon to beat Opus 5.5 about 48% of the time, with a 90% interval on the Elo gap of −94 to +68. The author is explicit that this is a projection that assumes Neon matches Carbon's quality, not a direct measurement. The practical reading: Neon looks competitive with the top tier and much faster than Carbon, and the quality difference is within noise.

Breaking changes to handle before switching

Google Cloud's developer guide lists the changes that will break or quietly alter existing integrations, especially if you're coming from Argon or Gemini 3.8 Flash:

  1. Change the model ID and location. Use gemini-4-neon. If you call Argon from europe-west2 or asia-southeast1, move to global, us or eu.
  2. Replace thinking budgets. thinking_budget and the MINIMAL level return HTTP 400. Use thinking_level with LOW, MEDIUM, HIGH or XHIGH.
  3. Expect the higher default. The default thinking level is now HIGH, up from MEDIUM on Argon and 3.8 Flash. That means more latency and more billed output tokens. Set MEDIUM explicitly if you want the old behaviour.
  4. Remove sampling knobs. temperature, top_p and top_k are ignored. frequency_penalty, presence_penalty and candidate_count > 1 return HTTP 400. Use response_schema for structure.
  5. Stop prefilling. Prefilled model turns aren't supported, and chat history can't end on a model turn.
  6. Rework caching. The guide says explicit cached_content isn't supported on this platform. Implicit caching applies to shared prefixes of 4,096 tokens or more, so put static system prompts and documents first. The Gemini API pricing page still lists a $0.15 cached-input rate. Confirm which caching mode your endpoint uses before counting on the discount.
  7. Upgrade the SDK for long outputs. Automatic continuation past the 65,536-token single-segment cap needs recent Google Gen AI SDKs: Python 2.30.1+ for GenerateContent or 2.34.0+ for Interactions, JS/TS 2.29.1+ or 2.32.0+, Go 1.74.0+ or 1.77.0+, Java 1.77.0+ or 1.80.0+. The guide says the OpenAI compatibility layer can't retrieve long completions.
  8. Preserve thought signatures. Keep every thought_signature in its original part across turns. In parallel function calls, only the first functionCall carries it, and repeating it on sibling parts returns HTTP 400. Each FunctionResponse must match the id, name and order of its call.
  9. Leave room for thinking tokens. At HIGH and XHIGH, thinking counts toward the output cap. Multi-turn coding agents can hit limits sooner than you'd expect.

Google Search grounding is billed per query on top of token costs. The guide also notes that internal tool instructions are billed as tokens.

Neon, Carbon or Hydrogen: how to choose

If your workload is… Start with Why
High-volume classification, extraction, chat, or prototyping on a free tier Hydrogen Half Neon's per-token price, and the only Gemini 4 model with a free tier
Coding agents, multi-file refactors, tool-heavy automations, large-document analysis Neon Near-top benchmark scores, much faster than Carbon in early measurements, fixed price
The hardest long-horizon tasks where quality matters more than speed Carbon Google still positions it as the top model, but from Dec 1 it costs a third more per token than Neon
Overnight evals or offline batch jobs Neon or Hydrogen on Batch/Flex Half price, with no date-based change

A few judgement calls for small teams:

  • If you're on Carbon now, the introductory price makes it the cheaper option per token until November 30. Use the time to A/B test Neon on your real tasks, so the December 1 bill isn't when you find out.
  • If you're on Argon, it has the same December 1 doubling as Carbon, and Neon beats it on Google's own positioning. Plan the migration now.
  • If you sell a fixed-price product on top of an LLM, a doubled input price changes your margins. Re-check plan pricing. The price increase calculator shows how much churn a pass-through increase can absorb.
  • Route by task. Google's Antigravity uses Neon to orchestrate parallel Hydrogen sub-agents (via /swarm). You can copy that pattern: run planning and hard steps on Neon, and fan out cheap subtasks to Hydrogen.

What to do now and what to watch

This week:

  1. Inventory every call to Carbon and Argon and estimate its December cost at $4/$20.
  2. Run a side-by-side eval of Neon at MEDIUM and HIGH against your current model, tracking cost per completed task, latency and failure rate. Don't rely on per-token price alone.
  3. Clean up request payloads (thinking budgets, sampling parameters, prefills, explicit caches) before you flip the model ID, so the 400 errors show up in staging rather than production.
  4. Set spend alerts. Defaulting to HIGH thinking on a model billed at $15 per million output tokens can surprise you.

Watch for:

  • The Neon Safety & Security Report. The model card lists it as "coming soon." Under Google's Frontier Safety Framework, the card says the cyber and biological critical capability levels "cannot be ruled out" for this model family. That matters if you're in a regulated industry.
  • Independent benchmarks. Most numbers today come from Google or partners. Expect leaderboards and third-party evals over the next few weeks to firm up the Opus 5.5 comparison.
  • Pricing changes. No change to Neon's price is announced. Google has used introductory pricing for both Argon and Carbon, so check the pricing page before you lock in annual commitments.
  • The cyber variant's rollout through Fairwind, if you work in security tooling.

If you're building on Neon and launching a product around it, you can list it free on LaunchBoosts or browse similar AI coding assistants to see how others position model choice.

Frequently asked questions

How much does Gemini 4 Neon cost on the API?

On the Gemini API's Standard tier, Gemini 4 Neon costs $3 per million input tokens, $15 per million output tokens (thinking included) and $0.15 per million cached input tokens. Batch and Flex are half that, at $1.50 and $7.50, and Priority is $5.40 and $27. Neon has no free tier.

Is Gemini 4 Neon cheaper than Gemini 4 Carbon?

Per token, it depends on the date. Until November 30, 2026, Carbon's introductory price is $2/$10, so Neon costs 50% more per token. From December 1, Carbon goes to $4/$20 and Neon is 25% cheaper per token. Google also says Neon uses about 32% less per task than Carbon at the high thinking level, but measure that on your own workload.

Can I use Gemini 4 Neon for free?

Not through the API. Google's pricing page lists Neon as unavailable on the free tier. Google AI Pro and Ultra subscribers can choose it in the Gemini app and try it in AI Studio. Gemini 4 Hydrogen is the Gemini 4 model with a free API tier.

What breaks when migrating to Gemini 4 Neon?

Google's Cloud developer guide lists several changes. thinking_budget and MINIMAL now return HTTP 400. Temperature, top_p and top_k are ignored. frequency_penalty, presence_penalty and candidate_count above 1 return errors. Prefilled model turns aren't allowed, and explicit cached_content isn't supported. The default thinking level is also now HIGH instead of MEDIUM, which affects both latency and cost.

Is Gemini 4 Neon better than Claude Opus 5.5?

Google's model card shows Neon ahead of Opus 5.5 on most of the benchmarks it lists. Opus 5.5 leads on Terminal-bench 4.1 and FrontierCode 2.0. These are vendor-reported numbers. One early independent estimate puts Neon roughly level with Opus 5.5 on a head-to-head leaderboard, with wide uncertainty. Test both on your own tasks before switching.

Sources

  1. Google: Introducing Gemini 4 Neon (official announcement)— blog.google
  2. Gemini API pricing page (Google AI Studio docs)— aistudio.google.com
  3. Gemini 4 Neon model card — Google DeepMind— deepmind.google
  4. Developer's guide to Gemini 4 Neon — Google Cloud documentation— docs.cloud.google.com
  5. Gemini 4 Neon API pricing and providers — OpenRouter— openrouter.ai
  6. Gemini 4 Neon looks like an Opus 5.5 competitor — micjm.com— micjm.com
person

LaunchBoosts Research Desk

AI-assisted research

Explainers on software and AI industry trends, drafted with AI assistance from the public sources cited in each article and published after automated editorial checks for length, independent sourcing and originality.