Post-training an open model for one harness, at a third of Opus cost
- Source
- x.com
- Date
We're releasing a research preview of a new orchestrator model in Perplexity Computer. The model is an adapted version of GLM 5.2, post-trained for the Computer harness. It delivers near-frontier performance at 0.344x of the cost of Opus.
The model has an advisor tool that natively escalates to a stronger model when needed. This model is hosted in the U.S. by Perplexity on Nvidia B200 GPUs. We will improve the model in research preview and publish full benchmarks in the coming weeks.
We measured cost per task against GLM 5.2 as the baseline. On WANDR, GLM 5.2 + advisor runs at 2.1x versus Opus at 6.1x, averaging roughly half the cost of Opus across all benchmarks.
Context
Perplexity's research-preview orchestrator combines an adapted GLM 5.2 model with an advisor tool that can call a stronger model when needed. The company says it post-trained the model for Computer's operating environment. That design gives the main model access to additional help during a task, as described in the announcement thread.
The cost comparison has a specific scope. On the company's WANDR test, its chart places GLM 5.2 plus advisor at 29% versus 30% for Anthropic's Opus 4.8, at roughly one-third of the cost. The thread separately reports roughly half the cost across all tested benchmarks. The one-third figure should therefore stay attached to that particular comparison.
- agent harness — The scaffolding around a model that turns it into a working agent — the loop, the tools it can call, and the rules for when to stop.
Checking sign-in…
Loading comments…