Most reviews evaluating AI fashions are primarily based on benchmarks of efficiency, however a current analysis report from Sonar takes a special method: grouping completely different fashions by their coding personalities and searching on the downsides of every in terms of code high quality.
The researchers studied 5 completely different LLMs utilizing the SonarQube Enterprise static evaluation engine on over 4,000 Java assignments. The LLMs reviewed had been Claude Sonnet 4, OpenCoder-8B, Llama 3.2 90B, GPT-4o, and Claude Sonnet 3.7.
They discovered that the fashions had completely different traits, resembling Claude Sonnet 4 being very verbose in its outputs, producing over 3x as many strains of code as OpenCoder-8B for a similar downside.
Primarily based on these traits, the researchers divided the 5 fashions into coding archetypes. Claude Sonnet 4 was the “senior architect,” writing refined, advanced code, however introducing high-severity bugs. “Due to the extent of technical issue tried, there have been extra of those points,” mentioned Donald Fischer, a VP at Sonar.
OpenCoder-8B was the “fast prototyper” because of it being the quickest and most concise whereas additionally doubtlessly creating technical debt, making it ultimate for proof-of-concepts. It created the very best challenge 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 must be a top-tier mannequin, nevertheless it solely had a go price of 61.47%. Moreover, 70.73% of the vulnerabilities it created had been “BLOCKER” severity, probably the most extreme kind of bug, which prevents testing from persevering with.
GPT-4o was an “environment friendly generalist,” a jack-of-all-trades that could be a frequent alternative for general-purpose coding help. Its code wasn’t as verbose because the senior architect or as concise because the fast prototyper, however someplace within the center. It additionally prevented producing extreme bugs for probably the most half, however 48.15% of its bugs had been control-flow errors.
“This paints an image of a coder who appropriately grasps the primary goal however usually fumbles
the small print required to make the code strong. The code is more likely to operate for the meant state of affairs however will likely be stricken by 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, in addition they shared related strengths and weaknesses. The frequent strengths had been that they shortly produced syntactically appropriate code, had stable algorithmic and information construction fundamentals, and effectively translated code to completely different languages. The frequent weaknesses had 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 turn out to be inclined to refined points within the code they generate, and so there’s this correlation between functionality and danger introduction, which I believe is amazingly human,” mentioned Fischer.
One other fascinating 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 had been 93% extra more likely to be “BLOCKER” severity.
“Should you suppose the newer mannequin is superior, give it some thought another time as a result of newer isn’t truly superior; it’s injecting increasingly points,” mentioned Prasenjit Sarkar, options advertising 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 growing reasoning has a diminishing return on useful efficiency. Bumping up from minimal to low leads to the mannequin’s go price rising from 75% to 80%, however medium and excessive solely had a go price of 81.96% and 81.68%, respectively.
By way of safety, excessive and low reasoning modes remove frequent assaults like path-traversal and injection, however exchange them with harder-to-detect flaws, like insufficient I/O error-handling. The low reasoning mode had the very best share of that challenge at 51%, adopted by excessive (44%), medium (36%), and minimal (30%).
“We’ve seen the path-traversal and injection turn out to be zero %,” mentioned Sarkar. “We will see that they’re attempting to resolve one sector, and what’s occurring 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 downside that has skyrocketed. Should you take a look at 4o, it has gone to 15-20% extra within the newer mannequin.”
There was the same sample with bugs, with control-flow errors lowering past minimal reasoning, however superior bugs like concurrency / threading growing alongside the reasoning issue.
“The trade-offs are the important thing factor right here,” mentioned Fischer. “It’s not as simple as to say, which is one of the best mannequin? The way in which this has been seen within the horse race between completely different fashions is which of them full probably the most variety of options on the SWE-bench benchmark. As we’ve demonstrated, the fashions that may do extra, that push the boundaries, in addition they introduce extra safety vulnerabilities, they introduce extra maintainability points.”