Google’s frontier model for complex coding, long-running agent workflows and cybersecurity research.
Last checked October 29, 2026
Carbon looks promising for difficult engineering and security work, but most people cannot use it yet. If you need a model today, choose one you can actually access and test against your own tasks.
Announced October 28, 2026. Google is starting with selected cyber defenders and trusted Fairwind partners; broader access has no published date. Gemini 4 Carbon is not available on Jynno.
Complex code, multi-step investigation and security analysis.
Early access is limited; benchmarks are not a substitute for a hands-on trial.
Carbon is a reasoning model, not a confirmed image or video generator.
Practical starting points based on Google’s reported strengths. These are suggested workflows, not results tested by Jynno.
Ask for a hypothesis, evidence in the code, a minimal fix and tests. Review the patch before running it.
“Trace this failing test to its root cause. Cite the relevant code, propose the smallest fix, and list regression tests.”
Summarize a vulnerability, assess affected components and propose a patch for an authorized system. Have a qualified reviewer validate it.
“Review this authorized system for the reported vulnerability. Explain evidence, affected paths, and a safe remediation.”
Break a multi-step research or engineering task into milestones; require sources, assumptions and checkpoints before decisions.
“Break this project into milestones. List assumptions and sources, and identify decisions that need my review.”
Selected results from Google’s published comparison. Higher is better; benchmarks measure specific tasks and do not establish a universal winner.
| Benchmark | Carbon | Argon | GPT-6 Astra | Claude Opus 5.5 |
|---|---|---|---|---|
| AA Intelligence Index v5.0 | 59 | 50 | 57 | 58 |
| Terminal-Bench 4.0 | 65.2% | 57.1% | 59.1% | 59.6% |
| FrontierSWE 2 | 61.2% | 55.0% | 65.5% | 62.3% |
| CWE-bench | 71.0% | 68.0% | 68.0% | 67.0% |
Carbon leads Argon on all four selected measures, while GPT-6 Astra leads this set on FrontierSWE 2. Pick the test closest to your task; small score differences are not a guarantee in production.
Google DeepMind benchmark tableGoogle announced introductory API pricing of $2 / $10 per million input / output tokens. After the introductory period, the announced price is $4 / $20. Cached input is 95% cheaper than input tokens. Google has not specified the end date of the introductory period.
These are Google’s announced API rates, not Jynno pricing. Availability and billing terms may change.
Illustrative calculation using Google’s announced text token rates; excludes cache, taxes, tools and any provider-specific fees. Actual token usage and availability vary.
Google says it is gradually extending access through its Fairwind program and plans to bring Carbon to developers, businesses and consumers. It has not announced a general release date or a public model ID in its announcement.
Selected defenders and trusted Fairwind partners
Google plans broader developer, business and consumer access
General release date and public model ID
No. This is an information page; Jynno does not currently offer Gemini 4 Carbon.
Google’s published table shows higher scores for Carbon on the selected tests shown above. Your results depend on the task and evaluation setup.
Google describes multimodal understanding, but its announcement does not establish Carbon as an image or video generation model.
This page summarizes Google’s announcement and comparison table. It does not represent an independent hands-on test. Checked October 29, 2026.
