OpenAI will move GPT-5.6 out of closed preview and open Sol, Terra, and Luna to the public on Thursday, July 9, the company said in a post on X on July 8. Preview access is also expanding globally starting immediately, pushing the three-tier family past the geographic limits that had kept it largely confined to closed testing. The bigger claim buried in the announcement is pricing: OpenAI says Terra, the mid-tier model, matches GPT-5.5’s performance while costing half as much to run.
The three-tier split itself was not new. OpenAI previewed Sol, Terra, and Luna, along with a reasoning-effort slider, weeks earlier in a closed rollout pending a US security review. What changes on Thursday is access and price, not the underlying architecture. OpenAI did not say in the announcement what resolved the review or whether it cleared in full, only that global preview access is expanding now.
Sol is positioned as the flagship, the model OpenAI expects developers to reach for when raw capability matters more than cost. Terra is pitched as the everyday workhorse: OpenAI says it holds competitive performance against GPT-5.5 at roughly half the price, an explicit bid to pull production traffic off the older model before it ages out of favor. Luna sits at the bottom of the lineup, described by OpenAI as the fastest and cheapest of the three while still delivering, in the company’s own words, strong capability.
None of those comparisons come with independent benchmark results. OpenAI’s performance claim for Terra rests on its own testing, and the announcement does not name a third-party evaluation or publish underlying scores. That is worth flagging, not dismissing: OpenAI has a direct incentive to make Terra look like a clean upgrade path off GPT-5.5, since every customer who migrates instead of staying put improves OpenAI’s own margins on inference.
The pricing structure mirrors what Anthropic and Google have already normalized across their own lineups: a flagship model for capability shoppers, a mid-tier model tuned for production cost, and a cheap, fast option for high-volume, low-complexity calls. OpenAI landing on the same three-rung ladder a year into that pattern suggests tiered pricing has become the default shape for a frontier model family, not an experiment specific to GPT-5.6.
Teams currently running production workloads on GPT-5.5 should benchmark Terra against their own evaluation set before Thursday’s cutover, rather than trusting OpenAI’s stated performance parity at face value. If the 2x price claim survives independent testing, GPT-5.5 pricing becomes difficult to justify for any workload that migrates without a quality loss.
Per OpenAI’s announcement on X, July 8, 2026.
