Unconfirmed OpenAI Model Codenames 'Terra', 'Sol' and 'Astra' Surface on OpenRouter

What surfaced

Three model identifiers—gpt-terra-latest, gpt-sol-latest, and gpt-astra-latest—appeared on OpenRouter under a tilde-prefixed ~openai namespace, a convention OpenRouter uses for listings that are not verified as officially sanctioned by the named lab. All three pages are live at openrouter.ai but carry minimal information: no confirmed parameter counts, no disclosed context window sizes, no published pricing per million tokens, and no benchmark scores.

The naming pattern—celestial and astronomical terms (Terra, Sol, Astra)—mirrors a style OpenAI and other labs have used for internal codenames during testing phases, which has fueled speculation that these could be early or leaked references to in-development checkpoints. That speculation is unconfirmed. OpenAI has not issued a statement, blog post, or changelog entry acknowledging any of the three names as an official product.

What we know versus what we don't

Confirmed: three listings exist on OpenRouter's platform using an ~openai tilde prefix, which the platform typically reserves for community-indexed or unverified model routes rather than lab-confirmed releases.

Not confirmed: whether these correspond to real, distinct trained checkpoints; whether they are test deployments, internal aliases, third-party proxies, or artifacts of routing infrastructure; training data cutoff dates; context window limits; input/output pricing; or any benchmark performance (MMLU, HumanEval, or otherwise).

No official OpenAI channel—API documentation, developer changelog, or press communication—references GPT Terra, GPT Sol, or GPT Astra as of this writing. Readers should treat all three as unverified until OpenAI or a comparably authoritative source confirms their existence and specifications.

Why this keeps happening

OpenRouter aggregates models from multiple providers and occasionally surfaces routes before a lab has made a public announcement, either because a partner is testing an integration or because a naming convention gets indexed automatically. This is not the first time codenamed or tilde-prefixed listings have generated attention disproportionate to their confirmed substance. The pattern typically resolves in one of three ways: the lab confirms the release with full specifications, the listing is removed as a testing artifact, or the name quietly disappears without explanation.

What this means

For builders evaluating whether to integrate GPT Terra, Sol, or Astra into production systems: don't, yet. There is no confirmed context window, no pricing, and no independent benchmark data to evaluate cost-performance tradeoffs. Treat these as rumor-stage signals rather than production-ready endpoints.

The broader pattern is worth watching. Labs increasingly test infrastructure through third-party routers before public launch, which means aggregators like OpenRouter can become an early—if unreliable—signal of upcoming releases. Until OpenAI confirms specifics, the responsible read is that three names exist on a listing page and nothing more. Anyone claiming performance numbers, pricing, or capabilities for these models is speculating, not reporting.