Issue #2 · June 20, 2026 · ~5 min read

Open weights just hit the cheap frontier for coding

Welcome back to Zero to MVP Weekly. Same deal: one signal, one tool, one MVP idea — five minutes, every Saturday. No link dumps. Let's go.

1. The signal of the week

An open model you can download just beat GPT-5.5 on coding — at a fraction of the price.

This week Z.ai's GLM-5.2 became the #1 open-weights model on the Artificial Analysis Intelligence Index, and independent testing had it beating GPT-5.5 on multiple long-horizon coding benchmarks at roughly one-sixth the cost. It's open-weight, the coding plan starts around $12/month, and it carries a 1M-token context window. Days earlier, Moonshot shipped Kimi K2.7 Code — another open, cheap model tuned specifically for agentic software engineering.

Why it matters: for most of the last two years, "use the cheap model for the grunt work" meant accepting worse results. That trade just narrowed to the point where it's barely a trade. For the unglamorous bulk of agentic coding — refactors, migrations, test passes, search-and-replace across a big repo — an open model at a fraction of frontier cost is now a serious default, not a consolation prize.

One caveat to keep you honest: "open weights" still doesn't mean "runs on your laptop." These are large models you'll mostly reach through a hosted plan or API. And benchmark crowns at this tier change hands almost monthly — so don't take the leaderboard's word for it. Take your own. (That's block three.)

2. My take

The cheap-open-model story usually gets framed as "replace your expensive model." That's the wrong frame. I'm not switching — I'm splitting the work.

The big paid models (Opus, GPT) do the thinking: analysis, architecture, planning the hard changes. The open models do the doing: implementation, routine edits, the small repetitive tasks that fill most of the day. GLM-5.2 passed every test I threw at it this week, so I can hand it the bulk of that work and reach for a frontier model only when the problem genuinely needs one — fewer premium tokens spent, without the quality drop I used to accept when I tried to go cheap.

But the part I keep coming back to is control. An open-weight model is one you can download, keep, and — if it ever comes to it — run on your own hardware. Even if "your own hardware" means renting a serious cluster, that option exists. Ask anyone who built on Fable 5 and woke up to find it pulled by government order: with a hosted closed model, access is something someone else can revoke. With open weights, the worst case is "this got more expensive to run," not "this is gone." That's not ideology — it's risk management.

So: route, don't switch. Use the expensive models where they earn it, push everything else to an open model you control, and keep a copy of the weights for the day you're glad you did.

3. Tool worth trying

GLM-5.2 vs Kimi K2.7 Code — and the one I'm keeping.

Two open models, both built for agentic coding, both cheap enough that testing them costs you almost nothing. GLM-5.2 is the long-context generalist — point it at a whole repo and have it plan and execute a multi-file change. Kimi K2.7 Code is the leaner agent specialist, priced at a dollar or two per million tokens, for high-volume runs where reasoning-token cost compounds.

I ran both this week on the same easy, medium, and hard tasks from a real codebase. The honest headline: both held up even on the hard problems — this isn't the usual "great on the easy stuff, falls apart when it matters" open-model story. Either one is good enough for real work.

I preferred GLM-5.2, and it's the one I'm making my default for routine, day-to-day tasks. I'll be straight with you: the gap is narrow and the call is partly subjective — Kimi is a genuinely strong alternative, and if you're optimizing purely for cost-per-token on big agent runs it may be the better pick. The point isn't which one I landed on. It's that two open models are now both good enough that the choice comes down to taste and price, not capability.

Two caveats to keep you honest: both are large enough that you're realistically using a hosted API, not local hardware; and this is my own hands-on testing, not a controlled benchmark — run your own tasks before you commit.

4. MVP idea of the week

Codemod-as-a-service on cheap open models.

Problem: big mechanical migrations — framework upgrades, deprecated-API sweeps, lint-rule rollouts — are slow, boring, and expensive. On closed frontier APIs, the token bill for a large-repo pass is enough to make teams just not do it.
Solution: a service that runs high-volume, methodical code transforms with a cheap open model (GLM-5.2 / Kimi K2.7) in an agent loop, priced per repo or per migration instead of per token.
For whom: agencies and teams sitting on large legacy codebases — the people who feel the pain and have the budget.
Minimal v1: a CLI that takes a codemod spec plus a repo, runs the open-model agent, and opens a PR with the diff and a plain-English summary of what changed. No dashboard, no backend.
Demand check: do three real migrations by hand-holding the tool, free, in exchange for honest feedback. Measure hours saved versus doing it manually. If two of the three ask for the next one, you have something.
Why now: a month ago, the per-token economics of a whole-repo agent pass were the dealbreaker. This week's open-weight models are the first time the math comfortably works — the signal up top is the reason this is newly buildable.

5. Useful links

GLM-5.2 tops the open-weights leaderboard — the independent benchmark behind today's signal; see exactly where it lands.
GLM-5.2 beats GPT-5.5 at ~1/6 the cost — the economics in one headline.
My hands-on: GLM 5.2 vs Kimi K2.7 Code — the full comparison on real coding tasks, if you'd rather watch before you try.
SpaceX is buying Cursor for $60B — the largest dev-tool deal ever, and a reminder not to marry your whole stack to one IDE.
The week the government switched off two Anthropic models — Fable 5 and Mythos 5 got pulled; the flip side of today's "own your model" argument.