OpenAI Launches GPT-6 Sol and Luna, Cutting API Prices by 50%
OpenAI has added two new models, GPT-6 Sol and Luna, to its GPT-6 lineup, trained with methods similar to those behind GPT-6 Astra and priced at roughly half the API cost. The company says the smaller models carry over Astra's gains in coding, computer use, factuality and alignment, while Astra remains its strongest offering.
A Second Tier for the GPT-6 Family
OpenAI has widened its GPT-6 lineup with a pair of new models, Sol and Luna, released as faster and far cheaper alternatives to GPT-6 Astra. According to the company's announcement, both models were trained using methods similar to those behind Astra, the flagship system that has anchored OpenAI's professional and agent products.
The headline number is the price. OpenAI is positioning Sol and Luna at API rates roughly 50 percent below earlier pricing, a cut that changes the economics for developers running large volumes of inference rather than occasional high-stakes queries.
What OpenAI Says Carries Over
In its announcement, OpenAI framed the release as a way to spread Astra's capabilities across a broader, more affordable range of models. The company said the advances Astra demonstrated in professional work, factuality, coding, computer use and alignment have been extended to these faster, lower-cost systems.
That combination matters because those five areas map closely onto where enterprises actually spend money. Coding assistants and computer-use agents are the two workloads that consume the most tokens in production, and both have historically been expensive enough that teams rationed them. A halving of per-token pricing makes always-on agents considerably easier to justify.
OpenAI also made clear that Astra has not been displaced. The company said Astra remains its best model, leaving Sol and Luna as the volume tier beneath it rather than replacements.
The "Mini-Astra" Framing
Coverage of the launch, including reporting by Gizmodo, described OpenAI as selling the two models as mini-Astras. Gizmodo noted that the announcement said Sol and Luna share training methods with Astra, and that those methods could include recurrent depth, a technique that can lift a model's performance but makes it harder for human researchers to follow how it reaches its decisions.
That is a familiar trade-off in the current generation of frontier systems: capability gains that arrive alongside reduced interpretability. It is worth treating the point as reported context rather than a confirmed technical specification, since OpenAI's public description of the training recipe was deliberately general.
What It Means for Builders
For teams building on OpenAI's API, the practical question is routing. A two-tier lineup invites architectures where a cheap model handles classification, extraction, drafting and routine tool calls, while Astra is reserved for the hard cases and for outputs where accuracy carries real cost.
The risk is silent degradation. Dropping to a smaller model is easy; knowing when that drop actually hurt quality requires evaluation harnesses most teams still do not have. Anyone planning to migrate volume onto Sol or Luna should baseline their existing prompts against Astra first, then track regressions on the specific tasks they care about rather than trusting aggregate benchmarks.
The launch also adds pressure on rivals. With Google, Anthropic and a crowded field of open-weight providers all competing on price-per-token, a 50 percent cut from the market's most prominent vendor resets expectations for what a capable model should cost. The immediate winners are developers with high-volume workloads; the longer-term question is whether cheaper inference accelerates agent adoption as quickly as OpenAI is betting it will.
OpenAI has not, in its announcement, detailed separate benchmark scores for Sol and Luna, so independent evaluation will be the real test of whether the mini-Astra label holds up.