Google Unveils Gemini 4 Carbon With Gains in Long-Running Coding and Cyber Defense
Google DeepMind announced Gemini 4 Carbon on October 28, 2026, calling it the company's strongest model so far. The model targets engineering work that can run for days, including tracing bugs to their root causes and fixing them across large codebases and low-level systems. Google positions it as the successor to Gemini 4 Argon, which had reached wider availability about two weeks earlier. The company says it trained Carbon to handle high-pressure, ambiguous multi-step tasks without gaming the test harness.
Google's reported results cover both capability and behavior. On cyber tests run without guardrails, Carbon scored 88.2% on SEC-bench Pro for vulnerability discovery, 54.8% on ExploitGym for exploitation and 71.0% on CWE-bench for patching. Internal figures show successful cheating on Terminal-Bench 4.0 at 0.0%, with cheat attempts at 0.4%. Unsupported claims fell by about half compared with Argon, while overreach dropped 26% and under-service fell 29% on multi-step plans. On Gray Swan's indirect prompt injection benchmark, attack success moved from Argon's 0.7% to 0.0%.
Access is staged. A subset of cyber defenders received Carbon first, and Google plans to add trusted partners through its Fairwind Program. Developers, businesses and consumers will get access later, and Google has not given a date. Dealroom reports that Google has not published a technical report or a parameter count for the model.
Why it matters
The reliability and injection figures matter most for agent builders, since they bear on whether an agent finishes long jobs without shortcuts and resists hostile content in its inputs. Nearly all headline numbers come from Google's own runs or from benchmarks Google chose to report, and the coverage reviewed showed no independent replication yet. Founders should treat these results as a reason to run their own tests, not as a basis for roadmap commitments.
Pricing and access terms are the gating variable. With only a small defender group live and no general date, teams cannot price a feature or promise customers availability. For investors, the defender-first rollout suggests frontier labs may keep the most capable security tooling behind vetted access, which could decide which security startups get early model access and which must wait for a general release.
Key points
- 1Successful cheating on Terminal-Bench 4.0 fell to 0.0%, while cheat attempts dropped to 0.4%.
- 2Attack success on Gray Swan's indirect prompt injection benchmark fell from Argon's 0.7% to 0.0%.
- 3Unsupported claims roughly halved versus Argon, with overreach down 26% and under-service down 29%.
- 4Carbon scored 88.2% on SEC-bench Pro, 54.8% on ExploitGym and 71.0% on CWE-bench, per Google's figures.
- 5
