Two model tiers
gstack 1.91.45.0 reviews plans on a frontier model and builds on a smart one
Fable 5.1GPT-6 AstraOpus 5.5GPT-6.1 SolIndependent plan reviews in gstack now default to a frontier model, Claude Fable 5.1 or GPT-6 Astra. For the build, it recommends Opus 5.5 or GPT-6.1 Sol, without switching the session you are in.
Your choice still wins
A model you choose still beats the new default
A model you name still wins. Set the plan review tier to smart to spend less, or to host to keep the model your CLI picks. On Bedrock, Vertex or a custom endpoint, name a compatible model.
gstack-models
gstack-models shows which model each job will use, and which setting chose it
A new command, gstack-models, lists the model each job will use and the setting that picked it, with commands to pin or reset it. It runs offline, so it can't confirm your account has access.
Fixes
Codex reviews keep to one deadline, and /ship reads QA instructions first
Codex model checks remember up to sixteen choices, even when runs overlap. A plan review's readiness check and the review now share one deadline. And /review and /ship read the QA instructions before a probe.
Weekly freshness check
gstack checks every week whether its model defaults are still current
docs/model-policy.mdModel policy guideEach week, gstack checks Anthropic and OpenAI's official model recommendations and retirement notices. A newer model becomes a candidate, and a person reviews any change to the defaults. The model policy guide covers the settings.
