I tried Doubleword’s async tier in Pi for cheaper GLM 5.2 inference, ~25%
I’m not affiliated with Doubleword, I just came across Doubleword.ai while looking at cheaper ways to run coding-agent workloads. I personally think local/open weight models are the way to go in the near future.
Their tiered pricing is interesting if your workflow can tolerate latency:
- Realtime: normal interactive inference, priced similarly to other providers.
- Async: ~1 min time to first token, ~25% cheaper.
- Batch: 24h SLA for batched tasks, ~50% cheaper.
GLM 5.2 pricing for reference:
| Model | Provider | Type | Realtime | Async | Batch (24h) |
|---|---|---|---|---|---|
| GLM 5.2 | zai | Generation | $1.40 in / $4.40 out | $1.05 in / $3.30 out | $0.70 in / $2.20 out |
I think, quite a few tasks like execution after careful research and planning, refactors, background exploration, or eval-style tasks can be offloaded to the async tier, not every call needs realtime latency.
If ~50% of your LLM calls during development can move to the async tier, that is ~12% total savings for the same model.
Batch could save more, although it's less relevant to live coding, but it could be significantly cheaper for things like large eval workloads.
GLM 5.2 also looks strong on coding/agentic benchmarks, but they are vendor-reported, so I’d still test on your own workload, but the price/performance looks worth it.
Added support via a Pi extension to test it out, sharing in case anyone wants to try it out.
Link: https://pi.dev/packages/@snehalyelmati/pi-doubleword-provider