Gemini 4 Carbon API Pricing 2026: $2/$10 Intro

Gemini 4 Carbon announced Oct 28 2026: $2/$10 per 1M intro pricing, $4/$20 after. SEC-bench Pro 88.2%, AA Index 59 (#1), but not yet callable.

The short version

Google announced Gemini 4 Carbon on October 28, 2026. It is the strongest model Google has shipped and the second in the Gemini 4 frontier line, arriving exactly two weeks after Gemini 4 Argon.

For API buyers the announcement contains three concrete numbers and one hard constraint:

• Introductory price: $2 per million input tokens and $10 per million output tokens, with cached input at 95% off the input rate.
• Standard price after the introductory period: $4 per million input tokens and $20 per million output tokens.
• Availability: not open. Carbon is rolling out to a subset of cyber defenders and trusted partners through the Fairwind Program. There is no developer signup, no model ID, and no API endpoint you can call yet.

That last point is the whole story for most developers. Carbon is a benchmark headline today and a pricing decision you will have to make later. The model you can actually call today is Gemini 4 Argon, which launched to developers on October 13 at the same $2/$10 introductory price.

Both models sit at the same introductory price. When Carbon opens up, the real question is not which one is cheaper — it is whether you should be holding a long-horizon agent workload on Argon now, or waiting for a model built to run that workload for days at a time.

Pricing: identical introductory rates, identical cliff

Carbon launches at the same introductory price as Argon, and both carry the same footnote: after the introductory period expires, the price doubles. This is a deliberate pricing structure for a frontier line, and it is worth understanding before you model your costs.

Two things jump out of that table.

First, Carbon introduces no new price point. At $2/$10 it is the same introductory tier as Argon and the same as GPT-6.1 Sol. The $4/$20 steady-state rate is exactly the Claude Opus 5.5 rate. Google is positioning Carbon as a peer of the existing frontier tier, not as a premium above it.

Second, the 95% cached-input discount is the number that actually changes unit economics. A cached input token costs $0.10 per million at the introductory rate and $0.20 at steady state, against $2 and $4 for uncached input respectively. For a long-horizon agent workload that re-reads the same repository, the same specification document, or the same conversation history on every turn, the difference between a cached and uncached request is a factor of 20. Any agent architecture built on Gemini 4 needs to use context caching as a default, not an optimization.

What the introductory period means for a production budget

The footnotes on both announcements read the same way: after the introductory period expires, the price of $4 per 1M input tokens and $20 per 1M output tokens will apply. Google does not publish a date. That means you cannot compute a reliable steady-state cost from the announcement alone, and any vendor quote you get for Carbon later may be the $4/$20 number rather than the $2/$10 headline everyone will repeat.

The practical approach for budgeting a Gemini 4 workload today is to price against $4/$20 and treat $2/$10 as a temporary subsidy. If a workload only makes sense at the introductory rate, it is priced against a promotion, not a product.

Benchmarks: what Carbon actually changed

Carbon's headline claim is position, not raw power. It ranks number one on version 5.0 of the Artificial Analysis Intelligence Index at 59, and it leads three cybersecurity benchmarks. But the more interesting number for an API buyer is how much of the gain is in the categories you would actually bill for.

Three findings are worth pulling out of that table.

The security scores are domain-specific, not general capability. Three of the six benchmarks are cybersecurity. A 54.8% on ExploitGym tells you Carbon is good at reasoning about exploitability, which does not directly translate into better JSON generation or better retrieval quality. If your workload is not security work, these numbers are less load-bearing than they look in a launch post.

The coding gain is real but it is an increment over Argon, not a step over everything. The +89 elo jump lands Carbon at number four on WebDev Arena, not number one. Google is reporting a strong improvement over its own predecessor rather than a clean sweep of the field.

The alignment numbers are the ones Google chose to lead with. Zero percent attack success rate on Gray Swan's Indirect Prompt Injection benchmark, a 26% reduction in overreach and a 29% reduction in under-service, and roughly half the unsupported claims of Argon. For anyone deploying an agent that touches untrusted content, that is arguably the most commercially relevant data in the announcement, and it is a departure from how frontier labs usually lead.

Availability: why you cannot call Carbon yet

This is the part that gets lost in benchmark-driven coverage. Carbon is not in the Gemini API. The announcement describes a deliberately staged rollout:

• Access begins with a subset of cyber defenders this week.
• Additional trusted partners follow through the Fairwind Program.
• Google says it is actively engaged in the U.S. government's voluntary process for pre-release model testing.
• Developers, businesses, and consumers come later, with no date given.

Argon took a similar path and reached developers on October 13, roughly two weeks after its own announcement. If Carbon follows the same cadence, a developer model ID is plausible in November 2026. That is a reasonable expectation, not a commitment, and the announcement contains nothing that would let you plan around a specific date.

The contrast with Argon is instructive because Argon demonstrates what a Google frontier model looks like on the API today. Argon runs an industry-leading 1 million token output limit, up from the previous 64K, and it is reachable on the paid tier of the Gemini API in Google AI Studio and Android Studio.

How to call Argon while you wait

Argon exposes a standard Gemini API model with a 1M context window. The OpenAI-compatible surface and the native Google SDK both work; the differences are in parameter naming, not in capability.

curl -X POST "https://generativelanguage.googleapis.com/v1beta/models/gemini-4-argon:generateContent"   -H "x-goog-api-key: $GEMINI_API_KEY"   -H "Content-Type: application/json"   -d '{
    "contents": [{ "role": "user", "parts": [{ "text": "Summarise this diff and flag the two riskiest hunks." }] }],
    "generationConfig": { "temperature": 0.2, "maxOutputTokens": 65536 }
  }'

import os
from google import genai

client = genai.Client(api_key=os.environ["GEMINI_API_KEY"])

resp = client.models.generate_content(
    model="gemini-4-argon",
    contents="Summarise this diff and flag the two riskiest hunks.",
    config={
        "temperature": 0.2,
        "max_output_tokens": 65536,
    },
)

print(resp.text)

# Every Gemini 4 request is a pricing decision. Cache the context you re-send.
usage = resp.usage_metadata
cached = getattr(usage, "cached_content_token_count", 0) or 0
print("input tokens:", usage.prompt_token_count, "| cached:", cached)

Two details in that second example matter more than they look. The first is that a 1M output-token ceiling changes what a single agent turn can cost. A model that can emit 65536 tokens of reasoning is a model that can burn $0.66 per call at the $10/M introductory output rate if it actually uses the full budget. Long-horizon agents need output ceilings set deliberately, not left at defaults.

The second is the cached-token count. At the introductory rate a cached input token costs $0.10 per million against $2.00 for an uncached one. A 500K-token context re-sent on every turn costs $1.00 uncached and $0.05 cached — a 95% saving on the largest line item in most agent bills.

Argon versus Carbon: which one you should wait for

Both models share a price, a safety stack, and a rollout shape. They differ in what they were built to do, and Google is unusually explicit about the difference.

The design targets point in different directions. Argon is a knowledge-work and migration model — it powers Google’s own C/C++ to Rust migrations across core libraries, including the Fuchsia Zircon kernel at 800K+ lines, and it is the leading model on the Vals Index for economic impact across finance, coding, legal, and tax work.

Carbon is a persistence model. Google’s description is that internal teams have seen Carbon run for days, persisting through dead ends to root-cause and patch subtle bugs across large codebases and low-level systems. It also reports more than 2x uplift in secure-code generation success rate over Argon, and the Chrome team reporting Carbon performed more than 10x better than another lab’s large frontier model at generating high-quality patches.

The practical read: if your workload is a well-scoped migration or research task, Argon is sufficient today and you can start now. If your workload is an agent that must survive a multi-day loop without losing the thread, Carbon is the model designed for it — and waiting is the correct call rather than trying to force Argon into a job it was not shaped for.

How the pricing stacks up against the frontier tier

Carbon lands in a crowded price band. Measured at current per-million-token rates for the frontier tier, with cached input treated as a separate line:

Carbon at steady state would be priced exactly alongside Claude Opus 5.5 and exactly double GPT-6.1 Terra. If you are running a frontier-tier workload today, the cost argument for switching to Carbon when it lands is weak on price alone — the argument has to be about what it completes that the cheaper model cannot.

That is where the honest comparison is actually decided. On the Artificial Analysis Intelligence Index, Carbon takes number one at 59. On the specific benchmarks that measure the thing Carbon was built for, it sets the state of the art. Whether that translates into fewer failed agent runs for your specific workload is something only a benchmark against your own prompts can answer.

If you want to test Carbon against your own traffic the moment it opens, without writing provider-specific integration code, a routing aggregator is the fastest path. FreeModel exposes an OpenAI-compatible endpoint that fronts multiple upstream providers, so you can point the same client at a different model ID and compare cost and quality without touching your integration layer.

Limitations, stated plainly

You cannot buy Carbon today. The Fairwind Program cohort is the only access path, there is no developer model ID, and no published date for general availability. Any planning that assumes a Carbon endpoint in the next sprint is planning against an assumption.

The introductory price has no published expiry. Both announcements carry the same footnote with no date. Model budgets against $4/$20 and treat the headline as temporary.

Context and output limits are not published for Carbon. Argon’s 1M output-token limit was a headline feature. Carbon’s numbers are not in the announcement, which means the per-call cost ceiling for a long-horizon agent is genuinely unknown.

Three of the six headline benchmarks are cybersecurity. If your workload is not security work, the leaderboard position is less meaningful than the announcement implies. DeepSWE v1.1 at 77.9% for Argon is a better proxy for long-horizon software engineering than a 54.8% exploitation score.

The rollout pattern is deliberate, and deliberate rollouts can stall. Google has committed to pre-release testing with the U.S. government and to iterating on guardrails before broad release. That is the right posture for a model with these capabilities and also a reason the timeline is open-ended. Argon took two weeks; there is no guarantee Carbon takes two weeks.

Where each model fits

Start on Argon now if you are doing large-scale code migration, enterprise research across legal or financial documents, or any multi-step task where the current model gets to the right answer without needing to survive an extended autonomous loop. Argon has the 1M output ceiling, the 1M context, the benchmark record on DeepSWE v1.1, and it is live in the API right now.

Wait for Carbon if your workload is a persistent agent that must run for hours or days, survive dead ends without the operator intervening, or operate on untrusted input where prompt-injection resistance is the binding constraint. Carbon’s zero-percent injection attack success rate and its days-long persistence are the two properties that would change your architecture.

Do not switch on price. Carbon’s introductory rate equals Argon’s introductory rate, and both double afterward. If cost is your deciding factor, GPT-6.1 Terra at $1/$5 is the cheaper frontier option today.

Verdict

Gemini 4 Carbon is a genuinely strong model announcement with a genuinely unshipped product behind it. The benchmarks are real and the safety posture is the most substantive part of the release, but as of October 28 there is nothing to integrate, no model ID to call, and no date for general availability.

For API buyers the practical response is to treat Carbon as a planning input, not a build target. Build against Argon, which is live and cheap at the same introductory rate, with the output ceiling set deliberately and context caching enabled from the first request. Then price the workspace against $4/$20 so the end of the introductory period is a budget adjustment rather than a surprise.

When Carbon opens, the question worth asking is not whether it is a good model — on the evidence so far it is — but whether the specific workload you are running today fails on capability rather than on price. If it fails on price, the $1/$5 tier is already winning. If it fails on capability, Carbon is the model that was built for it.

Sources

• Google — Introducing Gemini 4 Carbon (October 28, 2026)
• Google — Introducing Gemini 4 Argon
• Google — Gemini 4 Argon is now available for developers (October 13, 2026)
• OpenRouter models API — per-token pricing and context lengths for the comparison table

Related Articles

• Gemini 3.8 Live API Pricing 2026 — the audio-to-audio tier for voice agents
• GPT-6 Astra API Pricing 2026 — the $10/$50 frontier tier for comparison
• OpenAI Moderation API 2026 — moderation built into the same request