Most experiences evaluating AI fashions are primarily based on benchmarks of efficiency, however a current analysis report from Sonar takes a distinct strategy: grouping totally different fashions by their coding personalities and looking out on the downsides of every in relation to code high quality.
The researchers studied 5 totally different LLMs utilizing the SonarQube Enterprise static evaluation engine on over 4,000 Java assignments. The LLMs reviewed have been Claude Sonnet 4, OpenCoder-8B, Llama 3.2 90B, GPT-4o, and Claude Sonnet 3.7.
They discovered that the fashions had totally different traits, equivalent to Claude Sonnet 4 being very verbose in its outputs, producing over 3x as many strains of code as OpenCoder-8B for a similar drawback.
Primarily based on these traits, the researchers divided the 5 fashions into coding archetypes. Claude Sonnet 4 was the “senior architect,” writing subtle, complicated code, however introducing high-severity bugs. “Due to the extent of technical issue tried, there have been extra of those points,” stated Donald Fischer, a VP at Sonar.
OpenCoder-8B was the “speedy prototyper” because of it being the quickest and most concise whereas additionally probably creating technical debt, making it ultimate for proof-of-concepts. It created the very best concern density of all of the fashions, with 32.45 points per thousand strains of code.
Llama 3.2 90B was the “unfulfilled promise,” as its scale and backing implies it ought to be a top-tier mannequin, but it surely solely had a go fee of 61.47%. Moreover, 70.73% of the vulnerabilities it created have been “BLOCKER” severity, essentially the most extreme sort of bug, which prevents testing from persevering with.
GPT-4o was an “environment friendly generalist,” a jack-of-all-trades that could be a widespread selection for general-purpose coding help. Its code wasn’t as verbose because the senior architect or as concise because the speedy prototyper, however someplace within the center. It additionally averted producing extreme bugs for essentially the most half, however 48.15% of its bugs have been control-flow errors.
“This paints an image of a coder who accurately grasps the primary goal however usually fumbles
the small print required to make the code strong. The code is more likely to perform for the supposed situation however shall be suffering from persistent issues that compromise high quality and reliability over time,” the report states.
Lastly, Claude 3.7 Sonnet was a “balanced predecessor.” The researchers discovered that it was a succesful developer that produced well-documented code, however nonetheless launched numerous extreme vulnerabilities.
Although the fashions did have these distinct personalities, additionally they shared comparable strengths and weaknesses. The widespread strengths have been that they shortly produced syntactically appropriate code, had strong algorithmic and information construction fundamentals, and effectively translated code to totally different languages. The widespread weaknesses have been that all of them produced a excessive share of high-severity vulnerabilities, launched extreme bugs like useful resource leaks or API contract violations, and had an inherent bias in direction of messy code.
“Like people, they grow to be prone to delicate points within the code they generate, and so there’s this correlation between functionality and threat introduction, which I feel is amazingly human,” stated Fischer.
One other attention-grabbing discovering of the report is that newer fashions could also be extra technically succesful, however are additionally extra more likely to generate dangerous code. For instance, Claude Sonnet 4 has a 6.3% enchancment over Claude 3.7 Sonnet on benchmark go charges, however the points it generated have been 93% extra more likely to be “BLOCKER” severity.
“For those who assume the newer mannequin is superior, give it some thought yet another time as a result of newer shouldn’t be really superior; it’s injecting an increasing number of points,” stated Prasenjit Sarkar, options advertising and marketing supervisor at Sonar.
How reasoning modes affect GPT-5
The researchers adopted up their report this week with new information on GPT-5 and the way the 4 accessible reasoning modes—minimal, low, medium, and excessive—affect efficiency, safety, and code high quality.
They discovered that rising reasoning has a diminishing return on purposeful efficiency. Bumping up from minimal to low ends in the mannequin’s go fee rising from 75% to 80%, however medium and excessive solely had a go fee of 81.96% and 81.68%, respectively.
When it comes to safety, excessive and low reasoning modes get rid of widespread assaults like path-traversal and injection, however substitute them with harder-to-detect flaws, like insufficient I/O error-handling. The low reasoning mode had the very best share of that concern at 51%, adopted by excessive (44%), medium (36%), and minimal (30%).
“Now we have seen the path-traversal and injection grow to be zero %,” stated Sarkar. “We will see that they’re attempting to resolve one sector, and what’s taking place is that whereas they’re attempting to resolve code high quality, they’re someplace doing this trade-off. Insufficient I/O error-handling is one other drawback that has skyrocketed. For those who take a look at 4o, it has gone to 15-20% extra within the newer mannequin.”
There was an identical sample with bugs, with control-flow errors reducing past minimal reasoning, however superior bugs like concurrency / threading rising alongside the reasoning issue.
“The trade-offs are the important thing factor right here,” stated Fischer. “It’s not as simple as to say, which is the very best mannequin? The way in which this has been seen within the horse race between totally different fashions is which of them full essentially the most variety of options on the SWE-bench benchmark. As we’ve demonstrated, the fashions that may do extra, that push the boundaries, additionally they introduce extra safety vulnerabilities, they introduce extra maintainability points.”
