
Gemini 4 Carbon: Google's strongest model, its benchmarks and its price
Gemini 4 Carbon from Google: available since November 10, 2026, $2 and $10 per million tokens until Nov 30, plus cybersecurity benchmarks. Data as of Nov 15.
gemini-4-carbon) is a Google DeepMind model that Google presents as its strongest model. It announced it on October 28, 2026 with initial access for cybersecurity defenders and opened it to paying developers on November 10, 2026. Google presents it as its model with state-of-the-art cybersecurity and long-horizon capabilities.The data on this page is as of November 15, 2026. Carbon’s prices change on December 1, so check them again.
What is Gemini 4 Carbon and when did it come out?
In its October 28 announcement, Google describes it as “our strongest model with state-of-the-art cybersecurity and long horizon capabilities”, and calls it the next frontier model. It arrives two weeks after Gemini 4 Argon and is built for sustained agentic workflows and long, multi-step engineering tasks.
The launch happened in two steps. On October 28 it became available to a group of cyber defenders; for developers, businesses and consumers the announcement said “coming soon”. On November 10, 2026 Google published that Carbon was available through the API in Google AI Studio and in Gemini for business. The November 10 post also mentions a gradual rollout to Google AI Ultra subscribers (in Antigravity, the Gemini app and AI Mode in Search).
What is it for?
- Cybersecurity: finding vulnerabilities in real software, assessing whether they are exploitable, and proposing patches. This is where Google pushes hardest.
- Long-running agents: engineering tasks that take many steps. To understand how they chain together, see reasoning models.
If you want Google’s assistant for your day-to-day work, see Gemini for designers; this page covers the model.
What is the quick spec sheet?
| Item | Value (as of Nov 15, 2026) |
|---|---|
| API ID | gemini-4-carbon |
| Lab | Google (DeepMind) |
| Announcement | October 28, 2026 |
| API availability | November 10, 2026 (paid tier) |
| Context | 1 million tokens (per the AI Studio models page) |
| Max output | “64K / 1M” on the models page, with no explanation of what each value means |
| Standard price | $2 input / $10 output through Nov 30, 2026; $4 / $20 from Dec 1, 2026 |
| Open or closed | Closed (API access) |
What benchmarks did it publish and how do you read them?
These scores are reported by Google in its October 28, 2026 announcement. The announcement does not say who ran each evaluation; some benchmarks and indexes come from third parties and their origin is noted. Each lab’s benchmarks are not comparable with each other, so do not use them to rank across labs. To learn how to read them, see how to read AI benchmarks.
| Benchmark | Carbon score | Who measured it |
|---|---|---|
| Artificial Analysis Intelligence Index v5.0 | 59 (rank 1, per Google) | Artificial Analysis index (linked by Google); reported by Google |
| SEC-bench Pro | 88.2% | Reported by Google |
| ExploitGym | 54.8% | Reported by Google |
| CWE-bench (Collinear) | 71.0% | Collinear’s benchmark; reported by Google |
| WebDev Arena | 1768 Elo (rank 4) | Arena (linked by Google); reported by Google |
| Gray Swan Indirect Prompt Injection | 0.0% attack success (Argon: 0.7%) | Reported by Google |
How to read them, using what Google says in the announcement:
- SEC-bench Pro measures finding vulnerabilities: “autonomously uncovering complex flaws buried in real-world software”.
- ExploitGym is, per Google, a popular exploitation benchmark. The announcement adds that defenders increasingly rely on models to check which vulnerabilities are exploitable.
- CWE-bench measures whether the model fixes the root cause: it turns validated findings directly into working code fixes.
- Gray Swan IPI measures, per Google, prompt injection resistance. It is the attack success rate, so lower is better.
- WebDev Arena: Google ties it to better output and artifact quality in web development. It reports 1768 Elo, +89 points (the announcement does not say relative to what).
- The Artificial Analysis Intelligence Index is an external index (Google links to Artificial Analysis). The announcement does not explain what it includes, only that Carbon shows gains in coding and agentic tasks.
Google also notes that SEC-bench Pro, ExploitGym and CWE-bench were evaluated without the model’s guardrails. So if you are building a product, test the model with the protections you will actually use. Internal and partner figures cited in the announcement (for example, from the Chrome team) cannot be verified from the page; treat them as Google’s claims.
How much does it cost and how much context does it have?
According to Google AI Studio’s pricing page, as of November 15, 2026, per million tokens (output includes thinking tokens):
| Tier | Input through Nov 30 | Output through Nov 30 | Input from Dec 1 | Output from Dec 1 |
|---|---|---|---|---|
| Standard | 2.00 | 10.00 | 4.00 | 20.00 |
| Batch or Flex | 1.00 | 5.00 | 2.00 | 10.00 |
| Priority | 3.60 | 18.00 | 7.20 | 36.00 |
Points to keep in mind:
- The launch price ends on November 30, 2026. From December 1 the standard price doubles.
- There are no prompt-length thresholds: each tier has a single price.
- Context caching is implicit (explicit caching is not supported) and costs $0.10 per million through November 30 and $0.20 from December 1 on the standard tier.
To understand the bill, read cost per token, tokens and context window.
When should you pick it and when not?
Pick it when:
- Your task is cybersecurity or long, complex engineering.
- You want Google’s top-capability model to compare against a cheaper one from the same lab, such as Gemini 4 Hydrogen.
Skip it when:
- The task is repetitive and high-volume. A small model costs far less.
- You need open weights: see open vs. closed models.
- You are on a free account: Carbon requires the paid tier.
- You need the cybersecurity access of the Fairwind Program: Google grants it gradually to trusted defenders and partners.
Example
Illustrative example: a product team maintains a web app and wants to review a login module before a release. They give Carbon the module’s code and ask: “List possible vulnerabilities, say which can be exploited, and propose a root-cause fix.” The model returns five findings. The team does not apply anything blindly: a security person confirms which are real and which are false positives before touching the code.
What to try tomorrow: pick a long task you currently split across several parts for a model, give it whole to Carbon and to a smaller model, and compare how often each loses the thread.
Common mistakes
- Reading “rank 1” on an index as “the best for my case”: a general index does not measure your task.
- Comparing its scores with another lab’s as if they were the same test.
- Missing the price-change date: on December 1 the standard price goes from $2 and $10 to $4 and $20.
- Accepting a security answer without human review: a finding can be a false positive.
- Forgetting that Google ran the cybersecurity benchmarks without guardrails: test the model with the configuration you will use in production.