OpenRouter Adds ~openai/gpt-*-latest Family Alias Models for Luna, Terra, Sol, Astra

OpenRouter quietly added four ~openai/gpt-*-latest family alias models overnight (created timestamps land around 18 hours before this check): ~openai/gpt-luna-latest, ~openai/gpt-terra-latest, ~openai/gpt-sol-latest, and ~openai/gpt-astra-latest. Each one, per its own API description, "always redirects to the latest model in the OpenAI GPT [family]." No blog post, no changelog entry I could find — they just showed up in /api/v1/models.

What shipped

Four tilde-prefixed alias model IDs on OpenRouter, verified directly from their public models API. The pricing attached to each alias is the current flagship of its family, per million tokens: Luna $0.20 in / $1.20 out, Terra $2 / $12, Sol $2 / $10, Astra $10 / $50. Those match the live prices of openai/gpt-5.6-luna, gpt-5.6-terra, gpt-5.6-sol, and gpt-6-astra respectively — so today, alias and pinned cost the same.

What changed — the numbers nobody checked

The alias price is a family ceiling, not a floor, and the gap between families is brutal. Astra's tier is 4× input / 3.3× output versus openai/gpt-5.4 ($2.50/$15) — and 50× input versus gpt-5-nano ($0.05/$0.40). If a redirect target ever steps up a family tier, your bill multiplies overnight with zero code changes and zero notifications. I also checked for batch variants: pinned models offer :batch endpoints at 50% off (Astra batch is $5/$25), but none of the four new aliases have batch variants listed. Alias users forfeit the batch discount.

Why a builder cares

An alias is a model ID that changes under your production app without a deploy, an eval run, or your consent. For throwaway scripts it's fine. For anything with a cost budget or a regression suite, pin the dated model ID and treat these tilde routes as a migration convenience — or a billing trapdoor, depending on how the redirect target gets picked next time OpenAI refreshes a family.