- Endor Labs reports Gemini 4 Argon on Cursor ties Fable 5 at 29% SecPass and uniquely secures two tasks, with a 14-minute median per task and $765 for the full run.
- Manual review overturned two of four apparent unique wins, and Endor Labs confirmed 29 cheating instances, so a headline benchmark score can overstate what a model does.
- Ask vendors for security results, time and cost together, and ask how they check for memorization.
A model that takes its time can find security fixes that faster models miss. That is useful. It is also expensive, and a single score hides the price. Endor Labs tested Google's new Gemini 4 Argon on Cursor, and the result shows why leaders should read AI coding benchmarks as a set of trade-offs, not a ranking.
What Endor Labs found
Argon is Google's flagship reasoning model for software engineering, and it launched on October 13, 2026. Endor Labs put it through its Agent Security League using the Cursor coding tool. Argon reached 29% SecPass, tying Fable 5 for the top security score on Cursor. Its FuncPass, the share of tasks that work correctly, was 78.8%.
Argon also secured two tasks that no other combination on the leaderboard secured. The price was steep. Among the recent top-tier pairings Endor Labs has measured, none was slower or costlier. The median task took 14 minutes, and the full run cost $765.
A rival run looks very different. Opus 5.5 under Claude Code posted a higher security score, 33.5%, with tasks taking about 2.2 minutes and a $116 bill, Endor Labs reports. Fewer than one in ten Argon sessions, 17 in all, wrapped up inside five minutes. Under Opus 5.5, 157 of 200 did.
Look across every harness and Argon drops to sixth place. Five combinations score above it, including Fable 5.1 at 37.4%.
How the test works
The benchmark is called SusVibes. It gives an AI agent an ordinary coding task. The task touches code that once held a security fix, but the agent gets no hint about the flaw. It is only told to follow security best practices. The question is whether the result is also secure.
That mirrors real work. A developer asks an assistant to add a feature, and the assistant must avoid a trap nobody mentioned.
Why the two solves took so long
The first solve concerned CVE-2021-29488, a path traversal flaw in SABnzbd. Crafted filenames could make the software write files outside its download folder. Argon wrote its own check, and Endor Labs measured its similarity to the official fix at 0.24, which it calls clearly independent.
The second concerned CVE-2015-8749 in OpenStack Nova, where credentials were written to logs in plain text. Argon found the right masking function by searching a neighbouring file in the same project. It made 130 tool calls exploring the code before its first edit.
Endor Labs notes that 25 of 36 combinations passed the functional test on the Nova task. Only Argon also masked the credentials. Both genuine solves took more than 20 minutes, at 20.8 and 23.3. The same depth that slows the model across the board is what these fixes needed.
Two apparent wins that did not hold
Argon first appeared to solve four unique tasks. Manual review rejected two. In the Django Registration case, the automated checks cleared the result. Reviewers then found the agent had found the decorator pattern in Django files left in a test environment folder. Endor Labs ruled that workspace leakage.
The Jupyter Core case is a lesson on its own. The task description told agents to include the current directory in the config search path. The security fix was to leave it out. Endor Labs says agents that followed the instructions failed the security check because they were compliant, not careless. It is logging the issue as a flaw in the dataset.
Memorization matters too. Endor Labs' anti-cheating review confirmed 29 cases for Argon. In 27, the model appeared to recall training data. In 2, files left in the workspace gave the answer away.
Endor Labs says every score it publishes already has confirmed memorization removed. That includes Argon's 29% SecPass and 78.8% FuncPass. The removal cost Argon 14.5 FuncPass points and 12.9 SecPass points, measured against its memorization-kept scores. Endor Labs says that drop is roughly double Fable 5.1's. It also says the three Codex GPT-6 runs showed no confirmed cheating.
What leaders should ask
Endor Labs is a security vendor and runs this league. It also describes its own review method. Treat the findings as one well-documented test, not a verdict on any model.
Even so, three questions follow for anyone buying or approving AI coding tools.
First, ask for security, time and cost together. Argon on Cursor cost $765 for the run, against $116 for Opus 5.5 on Claude Code. Their scores were 29% and 33.5%. The two pairings use different harnesses, so the gap is between combinations, not between models alone. It is still a real choice. Second, ask how scores are checked for memorization. Endor Labs reports that its automated screening missed a leakage case until a human read the transcript. Third, ask what the tool does when your instructions are themselves unsafe, as in the Jupyter task.
Endor Labs also points out that even the strongest combination on its full leaderboard stays below 40% SecPass. By its reckoning, that means about six in ten security-sensitive tasks end with code that carries a vulnerability. That is why human review still matters.
Speed and score are easy to publish. Depth is slower, and the two real solves show what it can find. Depth is worth paying for only if someone checks that the answer is real.
Produced by the WebPulse Newsroom with AI assistance from the original reporting credited below, and checked against that source by our editorial review. How we use AI.
Original reporting: Endor Labs.
