AI Agent StoreSEC-bench Pro: 88.2%
Finding software vulnerabilities
Google tested 183 JavaScript-engine tasks from the May 2026 set, running each task three times and averaging the single-attempt success rate.
Read the evaluation details ↗Google · Announced · restricted access
Gemini 4 Carbon
Watch Carbon for demanding software-engineering and cybersecurity work. Access starts with selected defenders and trusted partners; most developers cannot switch to it yet.
Announced 2026-10-28 · First reviewed 2026-10-29 · Checked 2026-11-03. Announced October 28; this is not a general-availability release.
We still need comparable independent tests. You can compare the confirmed features and limits below, but we cannot yet give this version an Agent Fit or Practical Value rating. Test it on your own tasks before switching; results for another version or thinking setting may not apply.
What Google’s published tests show
Google published these results using its evaluation setup. They are not measurements from our matched independent ranking, and the tests differ from a real deployment. No effort-by-effort comparison or Agent Fit score is assigned.
ExploitGym: 54.8%
Checking whether a vulnerability can be exploited
Google ran this test without internet access or deployment safeguards. Tests allowed six or ten hours per task; Google does not specify which limit produced this headline figure.
Read the evaluation details ↗What is announced so far
Token prices measure the small pieces of text the model reads or produces. Capacity is the amount the service accepts in a request; check the limits below for any fallback to another model. A dash means the rate is unverified.
We could not confirm a model ID for general API access. Google describes improvements in sustained coding work and multimodal understanding. Public input formats, tool support and API identifiers are not yet confirmed.
Initial access is through Google’s Fairwind program. Wider developer, business and consumer availability is planned, without a public launch date.
Public context and output limits are not confirmed. A long-document benchmark does not establish the capacity your account will receive.
Google announces $2 input and $10 output per million tokens, with cached input 95% cheaper ($0.10). After the introductory period, input/output become $4/$20. Carbon’s promotion end date and other billing rules have not been announced. Allow for retries and any tool fees. Models that think before answering may also charge for that work.
Provider release and specifications ↗
Check availability and plans ↗
More setup and licensing details ↗
Test it before switching
Give each model the same examples, tools and definition of success. Record correct outcomes, response time, retries and total cost. A cheaper token price does not guarantee a cheaper completed task.
Compare caching and long-context costs ↗