You have 1 article left to read this month before you need to register a free LeadDev.com account.
Estimated reading time: 3 minutes
The latest generation of AI models have proven to be a major step forward for a wide range of coding tasks – but at what cost?
OpenAI’s GPT‑5.2 and Anthropic’s Claude Opus 4.5 models landed within weeks of each other at the end of 2025. Together, they have reset expectations for what AI coding assistants and agents can achieve.
The use of AI within developer workflows has split developers since ChatGPT kicked off the generative AI revolution three years ago. While many coders have been wowed by their potential, there has also been a steady decline into the trough of disillusionment. But the late-year release of these two models has many believing that an inflection point has been reached.
Gergely Orosz called the “model releases in November and December […] the tipping point where AI got really good at generating code,” in his Pragmatic Engineer newsletter.
“If you can give a coding agent a good enough specification and a good enough definition of what you want done, they can get there now,” says Simon Willison, co-creator of Django.
Willison is less interested in the one-shot ability of the models to try and produce what he wants. Instead, he’s impressed by the way the models can now run “very long sessions” of calling other tools for hours without degrading performance, making coding agents meaningfully more dependable in real projects.
There are still issues with hallucination, Willison admits – but in a way that is now fixable. If an error is introduced into code, Willison says it is trivial to ask the model to fix its own error.
Your inbox, upgraded.
Receive weekly engineering insights to level up your leadership approach.
What now?
Peter Steinberger, the creator of PSPDFkit, wrote that “the step from GPT 5/5.1 to 5.2 was massive.”
“Whereas in ~May I was amazed that some prompts produced code that worked out of the box, this is now my expectation. I can ship code now at a speed that seems unreal.”
For Willison, the last month or so has seen a step change in how AI models work. “What’s interesting about these two new models, which came out in November and December, is that part of the reason there’s so much buzz right now is that a lot of people spent the Christmas holidays playing with them,” he says. But the change isn’t just in user perception: he thinks something is fundamentally different, which he attributes to the application of reinforcement learning to code.
It’s believed that companies like OpenAI haven’t made new major training runs for their models in months, he points out, but instead have focused on fine-tuning the outputs of their models.
Coders benefit from the fact that reinforcement learning has taken place on code – a decision taken by model makers because it’s easy to figure out whether the model got the right answer or not. “It either compiles or it doesn’t,” he says.
As well as improvements among the models’ performance, users are also playing their part. “People have learned how to use the models more effectively,” says Liz Fong-Jones, technical fellow at Honeycomb. “There is better tooling and guardrails around the models.”
More like this
Jevons paradox in practice
The big question is what that means for the AI models’ biggest adherents. Willison wonders how the Jevons paradox – where the increasing availability and decreasing cost of a resource can increase the rate at which that resource is used – will play out for software developers.
The dystopian version would see AI bring the cost of software development down to zero. If so, what are the implications for professional developers? Orosz, for one, is optimistic that it will separate engineers from coders. “Being a solid software engineer and not just a “coder”, is going to be more sought-after than before,” he wrote.
The more optimistic version would see developers elevated to a higher plain of performance.
“I can produce 10 times the amount of software that I used to be able to because I don’t have to type it all into my computer myself,” Willison explains. “That makes me more valuable, and it makes the cost of software so much cheaper that companies that would have never have commissioned software can now.”
That said, not everyone wants to be more efficient. Fong-Jones has learned to love it. Her wife, less so. “She enjoys writing lines of code, and she looks at me like I have two heads, and she’s like, ‘Why would you take away the fun part?’”
London • June 2 & 3, 2026
Rands, Nicole Forsgren & Matej Pfajfar confirmed
Willison is in no doubt that the tide is rising – and he worries whether those in the industry who are currently avoiding using AI altogether might get washed over. As Nordhealth CTO James Stanier advises new engineering managers: “Use of AI is now required.”
