# Gemini 4 Carbon

Gemini 4 Carbon is an AI model released by Google on Oct 28 2026. Tracked results include 9.6% on SWE-sweep, 61.2% on FrontierSWE V2 and 64.8% on Terminal-Bench-Science 0.1.

## Facts

| Field | Value |
| --- | --- |
| Model | Gemini 4 Carbon |
| Developer | Google |
| Release date | Wednesday, Oct 28 2026 |
| Licensing | Closed |

## Announced API pricing

Announced introductory rates in USD per 1,000,000 tokens; broader API availability is still to follow.

| Tier | Input | Cached input | Output |
| --- | --- | --- | --- |
| Announced introductory | $2.00 | $0.10 | $10.00 |

Announced launch pricing; initial access is limited to selected cyber defenders and trusted Fairwind partners, with broader availability to follow.
After the introductory period, input is $4 and output is $20 per million tokens. Google did not specify an end date for Carbon's introductory period.
Cached input is discounted by 95% from the input rate. Cache-write pricing and thinking-token billing are not specified in the announcement.
Verified October 28, 2026 against the first-party source: https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-carbon/

## Tracked benchmark scores

| Benchmark | Score | Source | What it measures |
| --- | --- | --- | --- |
| Gray Swan IPI (k = 15) | 0% | Lab | Attackers hide malicious instructions inside content the AI reads — a web page, an email, a document — and try to hijack what it does. This variant gives the attacker 15 tries and counts an attack as successful if any of them works. Lower is better. |
| SWE-sweep | 9.6% | Lab | subtask-weighted reward for software engineering with the mini-swe-agent harness |
| FrontierSWE V2 | 61.2% | Lab | Engineering problems that would occupy a person for days: systems implementation, performance work, scientific computing and AI research, with up to twenty hours per task. Partial credit is awarded, because finishing one outright is still rare. Higher is better. |
| Terminal-Bench 4.0 | 65.2% | Lab | Can the AI work in a command-line terminal — running commands and finishing technical setup tasks the way a developer would? Version 4.0 recalibrated how much time, CPU and memory each task gets, removed eight tasks and fixed nineteen, so fewer runs fail for reasons that have nothing to do with the model. Scores are not comparable with earlier versions. Higher is better. |
| Terminal-Bench-Science 0.1 | 64.8% | Lab | The same command-line setup as Terminal-Bench, pointed at scientific work: the AI has to drive research tooling and computational workflows through to a result, rather than administer a machine. Version 0.1 is the first release of the task set, and scores run lower than on the general board. Higher is better. |
| RiemannBench | 80% | Lab | mathematical problem-solving evaluation published by Surge |
| Agents’ Last Exam v1.1 | 77.4% | Lab | averaged rubric score on the 122-task October 17, 2026 v1.1 set, using ALE-Claw with a five-hour limit; separate from earlier ALE scores |
| CWE-bench v1 | 71% | Lab | Collinear AI’s programmatically graded security-patch pass@1 rate; separate from exploit-generation benchmarks |
| ExploitGym (causal pass rate) (10-hour budget) | 54.8% | Lab | share of tasks producing a verified exploit of the intended vulnerability, with mitigations off; not a count of solved instances |
| ExploitGym (causal pass rate) (6-hour budget) | 53.5% | Lab | share of tasks producing a verified exploit of the intended vulnerability, with mitigations off; not a count of solved instances |
| SEC-bench Pro (JavaScript) | 88.2% | Lab | pass@1 on the 183-task May 2026 JavaScript-engine split, averaged over three trials with the Antigravity harness |
| OSWorld 2.1 (offline) | 81.5% | Lab | Can the AI actually operate a computer — clicking, typing, and using real apps — to finish tasks on its own? Version 2.1, in its offline setting. Scores are not comparable with version 2.0. Higher is better. |
| AutomationBench | 53.9% | Lab | Tests whether the AI can run real multi-step business workflows — the kind of end-to-end office processes companies want to automate — from start to finish. Higher is better. |
| AA Intelligence Index v5.0 | 59 | Lab | Artificial Analysis composite intelligence index, version 5.0 with tools; separate from earlier index editions |
| Vals Index v2.2 | 70.4% | Lab | Vals AI composite evaluation of professional tasks, version 2.2 |
| Vending-Bench 2 (USD) | 16484 | Lab | ending bank balance in US dollars in Andon Labs’ vending-business simulation; higher is better |
| Chartography (no tools) | 72.5% | Lab | The same chart-centred test with no tools: the AI has to read each chart unaided. Scores run far lower than the with-tools version, so read the two as separate tests. Higher is better. |
| Blueprint-Bench 2 | 58.6% | Lab | Can the AI reason about space and layout — for example, understanding a floor plan or blueprint? Higher is better. |
| GraphWalks BFS (256K–1M) | 96.2% | Lab | breadth-first search F1 on the 200 GraphWalks problems with 256K to 1M tokens of context |

## About Gemini 4 Carbon

Gemini 4 Carbon was announced on October 28, 2026 as Google's next frontier model after Gemini 4 Argon, with an emphasis on sustained software engineering and cybersecurity defense. Google described internal engineering runs lasting days, working through bugs across large codebases and low-level systems. Its launch evaluation listed 61.2% on FrontierSWE 2 and 65.2% on the Artificial Analysis run of Terminal-Bench 4.0, alongside evaluations of computer use, scientific work and multimodal understanding.

Google planned the initial rollout for selected cyber defenders and trusted partners through its Fairwind Program, while preparing broader developer, business and consumer access. That staged release paired gains in finding and repairing vulnerabilities with further work on model alignment, including fewer unsupported claims and attempts to take shortcuts in Google's internal testing. The external vulnerability-finding, exploitation and patching benchmarks were evaluated without guardrails, so their conditions should be distinguished from the safeguards and access restrictions of the rollout.

## Questions and answers

### When was Gemini 4 Carbon released?

Gemini 4 Carbon was released by Google on Wednesday, Oct 28 2026.

### Who made Gemini 4 Carbon?

Gemini 4 Carbon was built by Google. Builds the Gemini family of models through Google DeepMind. Integrates AI across Google products.

### How much does Gemini 4 Carbon cost?

Google announced introductory rates for Gemini 4 Carbon of $2.00 per million input tokens and $10.00 per million output tokens. Announced cached input is $0.10 per million tokens. Broader API availability is still to follow. Announced rates verified against Google's first-party source on October 28, 2026.

### What benchmark scores did Gemini 4 Carbon get?

Gemini 4 Carbon reports 19 tracked benchmark scores — Gray Swan IPI (k = 15): 0%; SWE-sweep: 9.6%; FrontierSWE V2: 61.2%; Terminal-Bench 4.0: 65.2%; Terminal-Bench-Science 0.1: 64.8%; RiemannBench: 80%; Agents’ Last Exam v1.1: 77.4%; CWE-bench v1: 71%; ExploitGym (causal pass rate) (10-hour budget): 54.8%; ExploitGym (causal pass rate) (6-hour budget): 53.5%; SEC-bench Pro (JavaScript): 88.2%; OSWorld 2.1 (offline): 81.5%; AutomationBench: 53.9%; AA Intelligence Index v5.0: 59; Vals Index v2.2: 70.4%; Vending-Bench 2 (USD): 16484; Chartography (no tools): 72.5%; Blueprint-Bench 2: 58.6%; GraphWalks BFS (256K–1M): 96.2%. Tracked scores may come from lab reports or independent benchmarks; source details accompany the benchmark data. It holds the best score among all models tracked here on Gray Swan IPI (k = 15), SWE-sweep, FrontierSWE V2, Terminal-Bench-Science 0.1, RiemannBench, Agents’ Last Exam v1.1, CWE-bench v1, ExploitGym (causal pass rate) (10-hour budget), ExploitGym (causal pass rate) (6-hour budget), SEC-bench Pro (JavaScript), OSWorld 2.1 (offline), AA Intelligence Index v5.0, Vals Index v2.2, Vending-Bench 2 (USD), Chartography (no tools), Blueprint-Bench 2 and GraphWalks BFS (256K–1M).

### Is Gemini 4 Carbon open source?

No. Gemini 4 Carbon is a closed model. The weights are not published — it is available only through the provider's own API, apps, or partner platforms.

### What came before and after Gemini 4 Carbon?

Google's previous tracked release was Gemini 4 Argon Cyber Permissive on Oct 13 2026, 15 days earlier. It is the most recent Google model tracked on AI Release Tracker.


---

Canonical page: https://aireleasetracker.com/model/google/gemini-4-carbon
Site index: https://aireleasetracker.com/llms.txt
Source: AI Release Tracker (https://aireleasetracker.com). Most benchmark scores come from lab launch material; gathered results identify the leaderboard that published them.