Follow the difficult bug.
Google describes sustained, multi-step engineering across large codebases and low-level systems.
A task to try
Trace this bug through the codebase. Explain the cause, propose a focused fix and identify the checks needed.
Gemini 4 Carbon.Stay with the hard problem.
Google DeepMind’s frontier model
Offloop is the AI workspace for small teams. Keep the brief, files and review together as you explore a new generation of coding and reasoning models.
- SEC-bench Pro · vulnerability discovery
- 88.2%
- ExploitGym · exploit validation
- 54.8%
- CWE-bench · vulnerability patching
- 71.0%
From finding a flaw to fixing it.
Google’s published evaluations cover vulnerability discovery, exploit validation and patching. These charts show model evaluations without guardrails, not results produced in Offloop.
Explore Google’s evaluationsSEC-bench Pro
Pass@1 · higher is better.
Hover or tap a bar to inspect its value. Use arrow keys while the chart is focused.
View all values
| Model / setup | Score |
|---|---|
| Gemini 4 Carbon | 88.2% |
| GPT-6 Astra | 85.4% |
| Gemini 4 Argon | 83.4% |
| GPT-6.1 Terra | 83.3% |
| GPT-6.1 Sol | 78.8% |
| MiMo-V2.6-Pro | 66.3% |
| DeepSeek-V4.1-Flash | 62.8% |
| DeepSeek-V4-Pro | 56.4% |
Find the vulnerability.
Google reports 88.2% on SEC-bench Pro.
Evaluation published by Google DeepMindExploitGym
Causal pass rate. Time caps differ between entries; compare the setup as well as the score.
Hover or tap a bar to inspect its value. Use arrow keys while the chart is focused.
View all values
| Model / setup | Score |
|---|---|
| Gemini 4 Carbon · 10h | 54.8% |
| Gemini 4 Carbon · 6h | 53.5% |
| GPT-6 Astra · no time cap | 42.4% |
| Gemini 4 Argon · 10h | 42.0% |
| Gemini 4 Argon · 6h | 39.5% |
| GPT-6.1 Sol · no time cap | 35.1% |
| Claude Opus 5.5 · 6h | 34.5% |
| Claude Mythos 5.1 · 6h | 30.4% |
| MiMo-V2.6-Pro · unknown time cap | 17.8% |
| GLM-5.3 · 6h | 15.0% |
Understand the risk.
Google reports 54.8% on ExploitGym.
Evaluation published by Google DeepMindCWE-bench
Pass@1, with ties broken by Pass@4. Published agent environments differ between entries.
Hover or tap a bar to inspect its value. Use arrow keys while the chart is focused.
View all values
| Model / setup | Score |
|---|---|
| Gemini 4 Carbon · Antigravity | 71.0% |
| Grok 4.7 · opencode | 68.0% |
| Gemini 4 Argon · Antigravity | 68.0% |
| GPT-6 Astra · Codex | 68.0% |
| Claude Opus 5.5 · Claude Code | 67.0% |
| Claude Fable 5.1 · Claude Code | 58.0% |
| Grok 4.6 · opencode | 57.0% |
| DeepSeek-V4.1-Flash · opencode | 55.0% |
| Muse Spark 1.3 · opencode | 55.0% |
| Hy4 Preview · opencode | 53.0% |
| GPT-6 Sol · Codex | 52.0% |
Fix the root cause.
Google reports 71.0% on CWE-bench.
Evaluation published by Google DeepMindFrom investigation to a working fix.
Google describes Carbon as a model for sustained engineering and cybersecurity defense. Start with a scoped task and a clear review process.
Strengthen software defenses.
Carbon targets vulnerability discovery, risk validation and root-cause patching.
A task to try
Review this authorized codebase for security weaknesses. Prioritize findings and propose patches for review.
Build with clearer evidence.
Google reports improvements in secure code generation and fewer unsupported claims in internal evaluations.
A task to try
Review this proposed change against the brief. Separate verified facts from assumptions and unresolved questions.
How Google measures the model.
Google’s October 28, 2026 evaluation overview. Results depend on the benchmark and evaluation setup; cybersecurity scores were evaluated without guardrails. These are not Offloop measurements.
Choose a benchmark
AA Index v5.0
Higher is better. Results reflect the published evaluation setup.
Hover or tap a bar to inspect its value. Use arrow keys while the chart is focused.
View all values
| Model / setup | Score |
|---|---|
| Gemini 4 Carbon | 59 |
| Gemini 4 Argon | 50 |
| GPT-6 Astra | 57 |
| Claude Opus 5.5 | 58 |
| Claude Fable 5.1 | 49 |
Give your team a place to work.
A model announcement is a starting point. Offloop gives people and AI Agents a shared workspace for the work around it.
Bring the brief.
Put the question, source files and acceptance criteria in a Channel. Give colleagues a clear starting point.
Review together.
Discuss findings with your team. Keep assumptions, feedback and decisions close to the task.
Keep the handoff clear.
Share the result and its next steps in the same workspace, so the next person can pick up the work.
A few things to know.
- What is Gemini 4 Carbon?
- Google DeepMind’s successor to Gemini 4 Argon, focused on long-horizon engineering and cybersecurity defense.
- What do these scores mean?
- They are Google-reported results on specific evaluations, run without guardrails. They do not predict performance on every codebase or deployment.
- Can I use Carbon in Offloop?
- This page does not announce an Offloop integration. Check the models available in your workspace.
- Where can I check pricing and access?
- Use Google’s official product and developer pages for current availability and pricing. Google API prices are separate from Offloop plans.
- How should our team review the work?
- Set acceptance criteria before starting. Review code in its intended environment and open the sources behind research claims. Keep decisions with the task in a shared Channel.
- How do we measure the value?
- Compare completed tasks with your usual process, including briefing, review and rework. Evaluate whether colleagues can understand and use the result.
Start with a shared goal.
Bring your people, AI Agents and next project into one workspace.
Open Offloop