Gemini 4 Carbon benchmarks: what Google’s scores show
The headline number is 59 in Artificial Analysis Intelligence Index v5.0, according to Google. But a useful benchmark comparison separates coding, computer use, reasoning and cyber capability. The tables below keep those categories apart and show where the new model still loses.
👁 132 reads
Update 10 November 2026: Gemini 4 Carbon is available to everyone
Since 10 November 2026 Gemini 4 Carbon is generally available. Model ID gemini-4-carbon in the Gemini API and Google AI Studio (paid tier only, no free tier), google/gemini-4-carbon in OpenRouter, and Google AI Ultra in the Gemini app, AI Mode and Google Antigravity (Low, Medium and High thinking levels). It is not part of Google AI Pro.
| API price per 1M tokens | Until 30 Nov 2026 | From 1 Dec 2026 |
|---|---|---|
| Input | 2 USD | 4 USD |
| Output (incl. thinking) | 10 USD | 20 USD |
| Batch / Flex (in / out) | 1 / 5 USD | 2 / 10 USD |
Context window: 1,048,576 tokens shared by input and output, up to 65,536 output tokens by default. Google AI Ultra: $99.99 (5x) or $199.99 (20x) a month.
My first test (10 Nov): 23 practical tasks all correct (0.23 USD, 264 s - Argon took 135 s), 16 of 16 scientific sources real with 0 invented, 8 of 8 planted vulnerabilities found in a code review, a 639-line spreadsheet engine passing 34 of 34 tests. Weak spot: at medium effort it twice used its whole token limit on thinking and returned no answer. Details: Gemini 4 Carbon is available and my Gemini 4 Carbon test vs Argon, Claude Fable 5.2 and GPT-6.1 Sol. The sections below were written on 28 October, before public access; prices and access above are current.
We have not tested Gemini 4 Carbon. Our team has no access to the model. This is an analysis of Google’s announcement and evaluation report, not a hands-on review. Every score below is reported by Google; competing-model figures in its comparison come from vendor reports or leaderboards.
The headline number is 59 in Artificial Analysis Intelligence Index v5.0, according to Google. But a useful benchmark comparison separates coding, computer use, reasoning and cyber capability. The tables below keep those categories apart and show where the new model still loses.
Google-reported benchmarks
Google’s report uses pass@1 and the highest thinking settings through the Gemini API. It runs SWE-sweep itself with mini-swe-agent, and ExploitGym and SEC-bench Pro with the Antigravity harness, without internet and with mitigations disabled. Google also ran Agents’ Last Exam and OSWorld internally. An external benchmark name does not make every submitted score an independent replication.
| Benchmark | Carbon | Argon |
|---|---|---|
| AA Index v5.0 | 59 | 50 |
| Vals Index v2.2 | 70.4% | 69.2% |
Coding and agents
| Benchmark | Carbon | Argon |
|---|---|---|
| SWE-sweep | 9.6% | 4.3% |
| FrontierSWE 2 | 61.2% | 55.0% |
| Terminal-Bench 4.0 (AA) | 65.2% | 57.1% |
| Agents’ Last Exam v1.1 | 77.4% | 69.0% |
| OSWorld-2.1 (offline) | 81.5% | 74.7% |
| AutomationBench (Zapier) | 53.9% | 51.3% |
| Vending-Bench 2 (USD) | 16 484 | 13 718 |
Reasoning and visual tasks
| Benchmark | Carbon | Argon |
|---|---|---|
| Chartography | 72.5% | 71.6% |
| Blueprint-Bench 2 | 58.6% | 54.4% |
| GraphWalks 256k-1M (BFS F1) | 96.2% | 84.2% |
| RiemannBench | 80.0% | 76.0% |
| Terminal-Bench-Science 0.1 (AA) | 64.8% | 52.9% |
Cybersecurity
According to Google, Carbon reaches 88.2% on SEC-bench Pro for vulnerability detection, 54.8% on ExploitGym for exploitation and 71.0% on CWE-bench for patching. These evaluations were performed without guardrails. They describe capability under those conditions, not a safe deployment policy or permission to run an agent against production systems.
| Benchmark | Carbon | Argon |
|---|---|---|
| CWE-bench | 71.0% | 68.0% |
| ExploitGym | 54.8% | 42.0% |
| SEC-bench Pro | 88.2% | 83.4% |
Honesty and agent behaviour
Google reports successful cheating at 0.0% and cheating attempts at 0.4% on Terminal-Bench 4.0. It also reports nearly 80% fewer serious misalignment alerts in live coding, about half as many unsupported claims as Argon, 26% less overreach and 29% less under-service. Gray Swan IPI successful prompt injections fall from Argon’s 0.7% to 0.0%. These are bounded evaluations, not a guarantee that an agent cannot be misled.
Google says Carbon worked for days on internal investigations into subtle bugs in large codebases and low-level systems. It also claims more than twice Argon’s effectiveness at generating secure code. Chrome’s reported gain of over 10 times concerns high-quality patches against an unnamed frontier competitor. Neither claim gives us an independent measure of time saved on a commercial repository.
Access and release date
The announcement is not a public release. Google says selected cyber defenders begin getting access this week, followed by trusted partners through Fairwind. Developers, businesses and consumers are to follow “as soon as possible”, without a calendar date. Paying for a Gemini subscription does not establish access to Carbon.
Fairwind is aimed at governments, national cyber bodies, critical infrastructure operators, major technology platforms and academic laboratories. Google describes more than 650 partners, but does not name eligible countries. The Fairwind page still describes Argon. A US developer or security team should treat the interest form as an application, not an instant API-key request.
What the results do not establish
For an engineering buyer, FrontierSWE 2 and Terminal-Bench answer different questions. Carbon’s stronger terminal score does not erase Astra’s lead on the repository benchmark. Similarly, its OSWorld result does not beat Opus. Choose a row that resembles the task before choosing a winner; do not average percentages from unrelated suites.
There is no English writing test, latency measurement or independent cost-per-completed-task result here. The announced token price cannot tell you how much an agent will spend over several days. A ranking index is also not a percentage of all work completed correctly. Before changing a production workflow, wait for access and compare the same repository, acceptance checks and spending cap. That is a proposed evaluation, not one we have run.
Related reading
Existing Gemini 4 Argon guide | Gemini 4 Carbon
Google · Your sourcesQuantslant higher in your Google - one clickAdd to preferred sources →Sources - accessed 28 October 2026
- Google, 28.10.2026 - 28.10.2026
- Google, Gemini 4 Carbon model evaluation - 28.10.2026
- Fairwind - 28.10.2026
- Sundar Pichai (X) - 28.10.2026
- Google (X) - 28.10.2026
- Gemini 4 Argon safety report - 28.10.2026
- Google One - 28.10.2026
More on this from QUANTSLANT
guides·Higgsfield Supercomputer: Chat Agent, Costs and Setup (2026)Higgsfield Supercomputer is an autonomous chat agent that plans video pipelines, selects generation models, and requests credit approval before running tasks.