Claude, Gemini, and GPT-5 can handle every SDLC task. Almost none of them should.
The New Stack Jeff Michael
New Stack argues big AI models like GPT-5 shouldn't run your whole dev pipeline. Smaller specialized models can do most SDLC jobs cheaper and more reliably.
There's a pattern forming in enterprise AI that looks a lot like the early cloud rush, and anyone who lived through that first decade of AWS migrations will recognize the shape of it immediately. Teams raced to adopt, the bills showed up later, and suddenly governance and cost control became the whole story. The New Stack's argument is that software development is now hitting the same wall, just with LLMs instead of EC2 instances.
The core claim is refreshingly specific: throwing Claude, Gemini, or GPT-5 at every task in the software delivery lifecycle is a waste, not a flex. Most of what happens between planning and deployment — writing unit tests, running build validation, checking compliance boxes — doesn't need a model that can reason across physics and poetry. It needs something narrow, fast, and boringly consistent. The piece leans on a Star Wars comparison that actually lands: frontier models are C-3PO, impressive generalists that cost a fortune to keep running; most SDLC work needs an R2-D2, a specialist built for one job and nothing else.
What's interesting is where this leaves the giant models. Not obsolete, just repositioned. Under this framework, a frontier model shows up at the start of a workflow to turn business goals into requirements and a plan, then again at the end as a kind of AI judge, checking whether the output actually satisfies what was asked. Everything in between — the actual grinding work of generating tests, reviewing code, verifying policy — gets handed to smaller, cheaper, fine-tunable models that behave predictably because they were never asked to be geniuses in the first place.
The governance angle is where this stops being theoretical. Once you've got five or six different models touching a single pipeline, you can't bolt compliance on at the end anymore, the way security reviews used to happen after code was already written. You need to know in real time which model touched what, how much it cost, and whether it followed policy. That's the gap the piece says AI control planes and orchestration layers are starting to fill, and it's also the gap fueling a fresh wave of consulting and FinOps tooling built specifically around AI token spend.
None of this is really about model capability anymore. It's about architecture, and about resisting the very natural urge to solve every problem with the biggest hammer available just because it happens to be the one everyone's talking about.
My take
This is basically the industry rediscovering a lesson microservices taught a decade ago: bigger and more general isn't automatically better, it's just more expensive and harder to audit. The uncomfortable part nobody wants to say out loud is that frontier-model vendors have every incentive to keep selling the C-3PO fantasy, because a fleet of cheap specialized models running on someone else's open-weight stack is a much worse business for them than one giant subscription doing everything badly.
Read more about this at: The New Stack
