Same model, same tasks: GLM-5.2 jumped from 23% to 52% by switching the harness

The pitch used to be simple. Pick the best model, wire it into your product, ship. Whoever had the smartest model won. That story is falling apart. Analysts, YC, and Chinese labs are now saying the same thing from different corners: the model is the easy part to copy, and the durable value sits in the layers around it. Let me catch you up.

The deep cut

• The model is the specialist, not the platform. Cursor became one of Anthropic's biggest customers, then watched Claude Code compete with it.
• Build a stack, not a subscription. Databricks Omnigent sits above Claude Code and Codex so harnesses swap without a rewrite.
• Copying the frontier is the strategy now, not the crime. Garry Tan wants U.S. labs distilling frontier models the same way Chinese labs do.

When your best feature becomes their product update

The threat is no longer the startup down the street. It's the platform you build on. TechCrunch is running a whole session at Disrupt called "What Happens When OpenAI Ships Your Roadmap", and the framing is blunt: your competitive edge can become someone else's release notes in a few months.

Cursor is the clearest example. It grew into one of Anthropic's largest customers while Anthropic's Claude Code moved into the same job at the application layer. Cursor's answer was to invest in its own stack, its Composer models, its own agent runtime, and a router. When the vendor ships the same work as a product, the partnership flips, and you want to own something they can't easily copy.

The four layers under the hood

The fix is architectural, not a matter of picking a better model. The control-plane argument lays out a stack: models, a router that triages each request, a harness that runs the model, and a meta-harness above the harnesses.

The router decides which model sees each request based on cost, quality, and risk. It is not caching or failover. Known, bounded work goes to a cheap validated model. Hard, ambiguous work escalates to the frontier. That's how you stop spending frontier dollars on work a smaller model handles fine.

Databricks' open-source Omnigent is the public example of the top layer. It gives one interface across Claude Code, Codex, Cursor, and Pi, so you can swap or compose harnesses while your tools and policies stay put. The point is to keep policy, spend, and evals above any single vendor.

The harness matters more than you think

If you assumed the model does the heavy lifting and the wrapper is decoration, the benchmarks say otherwise. One SWE-bench Pro sweep ran 10 harnesses against the same tasks with fixed models. GLM-5.2 moved from 23.2% to 52.4% just by changing the harness. The rank correlation between two model leaderboards was -0.05, meaning the wrapper reshuffled almost the whole field.

That cuts both ways. Standardizing on one coding agent recreates the lock-in you left the model layer to escape, now with tools, transcripts, and habits baked in. Swapping Claude Code for Codex is a runtime change, not a name change. Report capability at the model-harness level, and keep the harness swappable.

Copying the frontier is now the open strategy

Open models are the pressure that makes all this real. Moonshot AI is targeting $2 billion in annual revenue off its open-weight K3, with as many as 300 billion tokens a day flowing through it on OpenRouter. Anthropic alleges Moonshot ran a distillation campaign, routing nearly 300,000 requests to Claude Opus and collecting over 23 million responses for training.

Here's the twist. Y Combinator's Garry Tan doesn't want regulators to stop this. He told CNBC he'd "do nothing" and floated an "American distillation regime," where U.S. open-weight labs copy frontier models the same way, through the front door. His fear is one monolithic company owning all the intelligence. Nathan Lambert's reading list backs the trend: the open-closed gap has shrunk to roughly four to six months, and Western firms like Thomson Reuters have already moved off Claude onto open models to cut costs.

Three questions for your team

• If our top model provider shipped our core feature tomorrow, what do we still own that they can't copy?
• Can we swap our model or coding agent without a rewrite, and do we have evals that prove which work is safe to send to a cheaper one?
• Do we measure cost per successful outcome, or just cost per token, and what would routing save us this quarter?