JOURNAL / 2026.10.29
Gemini 4 Carbon improves software defense, but access remains restricted
Google announces a model for sustained engineering and cybersecurity. External evaluations support gains over Argon; their conditions and the limited rollout leave a gap between measured capability and a tool any team can use.
Google announced Gemini 4 Carbon on October 28. It describes internal runs lasting several days to solve engineering problems and is beginning to offer the model to a group of cybersecurity defenders, with gradual expansion through Fairwind. General access for developers, businesses and consumers comes later. This is a capability development worth examining, even though it does not yet support planning an ordinary migration.
There is evidence beyond the announcement. Artificial Analysis places Carbon first on its Intelligence Index v5, at 59 points against Argon’s 50. On its coding agent index, Carbon with Antigravity trails Opus 5.5 with Claude Code. The improvement within Gemini is substantial; the result compares systems with tools and configurations, and does not establish superiority in every repository.
The measuring instrument has changed too. Version 5 of the index, published the previous day, gives 60% of its weight to evaluations with private tasks or held-out answers. It combines coding, science, professional work and general capabilities. Comparing Carbon and Argon within that version makes sense; stringing together headlines from earlier rankings can confuse a change in the model with a change in the exam.
The most useful test ends with a patch
Collinear AI’s CWE-bench v1 presents 120 private audit-and-patch tasks. The agent receives a repository without being told its vulnerabilities. Verifiers check that the attack stops working and existing tests still pass. On the single-attempt programmatic measure, Carbon reaches 71%, against Argon’s 68%; both use Antigravity. This is a concrete difference within a bounded test, not a predicted success rate for a commercial audit.
That design seems closer to a real need than producing a convincing explanation of a flaw. A patch that blocks the sample attack but breaks a legitimate function does not fulfill the assignment. Still, passing the available verifiers does not demonstrate the absence of other vulnerabilities. The value lies in delivering something another person can inspect and test.
Google’s evaluation methodology for Carbon requires separating other launch figures. SEC-bench Pro is calculated on 183 JavaScript engine tasks, using Antigravity and criteria for reproducing the flaw. ExploitGym uses an offline environment, disabled mitigations and different time budgets. These measurements are vendor runs; CWE-bench comes from Collinear. Their percentages cannot be multiplied to estimate how many flaws a single agent would find, exploit and repair from start to finish.
I would not equate persistence with productivity either. Working for days could allow investigation of a difficult dependency or prolonged pursuit of a mistaken hypothesis. I would want to evaluate the cost of reaching an accepted change, how many proposals the maintainer rejects and whether the agent preserves evidence that contradicts its solution. A longer process pays off only when it improves the reviewable outcome.
The agent’s own safety remains a separate question. Google reports less misconduct in its internal tests; that does not establish that a prolonged run will respect every project boundary. The Gemini comparison page also shows uneven gains across tasks: a general ranking summarizes different capabilities and leaves out the particular conditions of each integration.
My reading is that Carbon strengthens a valuable direction: assigning software investigations with a verifiable deliverable and keeping a review process able to reject it. For a team without access, the immediate decision is to prepare its own cases and acceptance criteria. When the rollout expands, there will be a more useful question than “Is it number one?”: does it find and fix relevant problems our previous process missed, at a manageable review cost?
Primary sources consulted on October 29, 2026. We have not run Carbon or reproduced the evaluations. The proposed evaluation approach reflects our judgment, not observed results.