CONFIRMED

What was a loose string in an Nvidia commit became an official launch in under 24 hours. On September 22 OpenAI announced GPT-6 Sol and GPT-6 Luna, the two models the MaxAssistant signal desk had been tracking since the start of the month. The company describes both as trained with methods similar to GPT-6 Astra, bringing part of its state-of-the-art performance in professional work, factuality, coding and computer use into faster, more affordable models.

What changes for users

OpenAI cut API pricing for Sol and Luna by 50% versus the promotional GPT-5.6 versions, arguing that gains in caching and inference let it pass the savings directly to users. Sol targets complex coding and longer agentic workflows, Luna targets high-volume, narrow-focus tasks. Both are already available in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise and Edu users; Free and Go accounts get Luna in the desktop app. In regular ChatGPT, outside Work and Codex, the models are not yet available, and Enterprise admins must enable access.

The detail nobody predicted

The rumor circulating for weeks bet on three tiers, Sol, Terra and Luna, mirroring the GPT-5.6 structure. OpenAI dropped the middle tier: there is no GPT-6 Terra. The company kept just three sizes overall, Astra at the top, Sol as large, and Luna as small, simplifying the family instead of replicating the previous structure.

The numbers OpenAI released

On the OSWorld 2.0 offline computer-use benchmark, the company says GPT-6 Sol at xhigh effort scores close to Claude Opus 5 at medium effort, 60.5% versus 60.3%, at roughly 80% lower cost per task. GPT-6 Luna at max mode reportedly beat GPT-5.6 Sol at medium mode at one tenth of the cost. OpenAI also cited internal coding-agent usage growth: median daily token spend valued at API prices rose to $600 for the median researcher and $7,000 for the 90th percentile, an argument used to justify why cutting cost per task matters more now than an isolated list-price cut.

What the rumor got right and wrong

The signal desk correctly caught the central artifact, the "GPT-6 Sol medium" string seen in an Nvidia commit on September 21, and correctly read the silent description change on GPT-5.6 Sol in OpenAI's catalog around the same time as a sign something new would take the top spot. It missed on pricing: the circulating figures, $2.50 to $4 input per million tokens, didn't match the 50% cut over GPT-5.6 promotional pricing OpenAI actually applied, and a GPT-6 Terra never materialized.

MaxAssistant's read

This is the pattern worth watching going forward: a rumor born from a real technical artifact, gaining strength when the company itself changes documentation language without announcing anything, turning into a launch within days. For anyone building agents and products on OpenAI's API, the 50% cut on Sol and Luna is the most relevant practical takeaway from this round, more than the isolated benchmarks, because it resets cost per task for the tier most production applications actually use.

Update note

Originally published September 7 as a RUMOR, raised to STRONG SIGNAL on September 22 after the Nvidia commit and OpenAI's catalog change, and updated the same day to CONFIRMED after the official GPT-6 Sol and GPT-6 Luna announcement.

Sources

OpenAI, GPT-6 Sol and GPT-6 Luna announcement: https://openai.com/index/gpt-6-astra/ | 9to5Mac, OpenAI upgrading ChatGPT and Codex with two more GPT-6 models: https://9to5mac.com/2026/09/22/openai-upgrading-chatgpt-and-codex-with-two-more-gpt-6-models/ | OrcaRouter, GPT-6 Luna and GPT-6 Sol Leak: What's Actually Verified: https://www.orcarouter.ai/blog/gpt-6-luna-and-gpt-6-sol-leak