The $3.36 question: do you still need to rent your coding model?

For two years the answer was easy. You wanted serious AI coding help, you paid Anthropic or OpenAI, and you ate the bill. That math just changed. Open-weight models are now good enough for real work, and some of them run on a GPU sitting under your desk. Let me catch you up on what shifted and what it means for your roadmap.

The bill that stopped making sense

Claire Vo ran GLM 5.2, an open-weight model from Z.AI, through four real jobs in her actual production codebase. An architecture audit, a UI redesign, and a 45-minute autonomous bug hunt that pulled from Sentry and Vercel logs. The total came to $3.36 for roughly 6 million tokens. She shipped a bug-fix dashboard from it and got a landing page that matched her design system on the first try.

The headline is the price, but the real story is "open-weight." You can plug GLM 5.2 into Cursor or Claude Code through OpenRouter and keep your existing workflow. Same tools, different engine, a fraction of the cost. That is the part to sit with. The switching cost is low.

Good enough is the new bar

The other shift is where these models run. KDnuggets walked through seven coding models you can run locally in 2026 on a single GPU like an RTX 3090. Not demos. Real coding agents, repo edits, debugging, terminal work.

The trick is efficiency. Models like Nemotron Cascade 2 carry 30B parameters but only fire about 3B at a time, so they reason well without the full cost. Qwen3.6 27B is the pick for an all-round local setup. There is even a 9B option for smaller machines. The author's advice is plain: start with one, don't waste time hopping between them.

The work changed, so the model has to

Here is the part that hits your team directly. Coding is not just writing functions anymore. Your engineers read docs, check errors from screenshots, study architecture diagrams, and dig through messy project files. The models on that list reflect it. Gemma 4 31B and EXAONE 4.5 are multimodal, so they handle a screenshot of a broken UI, not just the code behind it.

That matters because the value of a coding assistant now lives in the full loop. Pull logs, read the error, propose a fix, ship it. Vo's 45-minute autonomous run was exactly that loop. When the tool can do the whole thing cheaply and locally, the question is no longer "is it smart enough." It's "why are we still renting it."

Where it still trips

Don't rip out your contracts on Monday. GLM 5.2 stumbled in places, and the local models trade some raw reasoning for the privilege of running on your hardware. The 9B model won't beat a frontier model on hard problems, and even Vo, who is replacing Opus in her own workflow, didn't say it wins everywhere.

So treat this as a portfolio move, not a swap. Use a cheap open-weight model for the high-volume, lower-stakes work: audits, exploration, first drafts, bug triage. Keep the expensive frontier model for the gnarly stuff where being wrong costs real time. The point is you now have a cheaper tier that didn't exist a year ago.

The deep cut

The thing easy to miss: the lock-in was never the model, it was the assumption that you couldn't move. Open-weight means you own the weights and can run them where you want, on OpenRouter, on your own box, or on a rented GPU. That changes your leverage at renewal time. You can run the same task on three engines this week and compare cost and output yourself.

So the move is to run that test before your next vendor conversation. Pick one real task from your backlog, run it through your current paid model and through GLM 5.2 or a local Qwen build, and put the cost and quality side by side. Walk into the renewal knowing what you'd actually lose by switching. Most teams have never measured that.

Three questions for your team

1. What share of our AI coding spend goes to work a cheaper open-weight model could handle, and have we ever tested that split?
2. If our main vendor doubled prices tomorrow, how many days would it take us to move to an open-weight model in the same tools?
3. Which of our coding tasks actually need a frontier model, and which are we sending there out of habit?