OpenAI cuts token prices in half for new GPT-6 Sol and Luna models

Sam Altman posted on X on the afternoon of September 22, 2026, as OpenAI published the announcement for GPT-6 Sol and GPT-6 Luna. “They are also half the price per token, and even less per task,” he wrote.

Both models went public that same afternoon, extending the GPT-6 family below GPT-6 Astra, which had launched on September 3. Sol took the complex coding and agentic work. Luna took the high-volume clerical jobs—document summarization, information extraction, short answers. They sat as the lower-cost counterparts beneath the flagship, listed in the API as gpt-6-sol and gpt-6-luna.

“GPT-6 Astra has launched a new era of intelligence; these models enhance its advantages by making that intelligence more efficient and accessible,” OpenAI stated. “While the most demanding and important projects still call for Astra's full depth, work happens at different scales, rhythms, and budgets.”

Kenneth Green, writing in Tech Weekly that day, framed the release as hierarchy expansion. The initial Sol and Luna models had appeared earlier in the year as distinct tiers, Sol for intricate coding and Luna for clerical volume with clear objectives. The GPT-6 versions updated those places and set them under Astra so the family could cover work that did not need the flagship’s full depth. The tiers were in place. How cheaply the stack would run came next.

Sol now costs $2 per million input tokens and $10 per million output. Those rates replace the $4 and $20 that GPT-5.6 Sol charged under promotional pricing. Luna costs $0.10 per million input tokens and $0.50 per million output, cut from $0.20 and $1.20. A token is a chunk of text the model reads or writes; API usage is billed by the million. OpenAI set the 50 percent reduction against the GPT-5.6 promotional figures and attributed it to improvements in caching and inference. Caching stores a prompt prefix so later calls can reuse it without full reprocessing.

GPT-6 Astra had supplied the methods template. The company trained both new models with the same approach and said it was passing the infrastructure savings directly through. Cached input-token reads now carry a 90 percent discount. Developers can set explicit breakpoints where a cached prompt prefix ends, and they can change reasoning effort or switch tools on and off without losing the cached context.

“The GPT‑6 models lead across the cost–intelligence curve, combining exceptional capabilities at every tier with infrastructure that delivers them efficiently at scale,” the announcement stated. OpenAI, the San Francisco public benefit corporation behind the GPT line, tied the new rates to that infrastructure. The comparisons that followed measured cost per task.

“In our internal factuality assessments, which are based on anonymized real-world interactions where users identified mistakes made by our models, GPT-6 Sol commits roughly half the errors as its predecessor, achieving Astra-level reliability at a significantly lower cost,” the announcement said.

On AutomationBench, Sol at extra-high effort scored 33.2 percent at $0.27 per task. Claude Opus 5 at maximum effort scored 26.9 percent at 11.1 times that cost. GPT-6 Astra at low effort scored 30.3 percent at 3.9 times. Agents’ Last Exam put Sol at maximum effort on 56.4 percent, beating Opus 5’s best result at 60 percent lower cost per task. DeepSWE v1.1 gave Sol 68.8 percent against Claude Fable 5’s 69.9 percent at roughly 80 percent less cost per task; Luna scored 66.6 percent, comparable to Opus 5 and Fable 5 at medium effort but 93 percent and 96 percent cheaper. FrontierCode 1.1 listed Sol at 49.3 percent maximum effort for $2.14 per task, Fable 5.1 at 50.3 percent for $12.83, and Opus 5 at 53.4 percent for $4.31 at medium effort.

“While GPT‑6 Astra remains the world's best model for computer use, GPT‑6 Sol and Luna offer more cost-efficient performance than their predecessors. On OSWorld 2.0 offline, GPT‑6 Sol at xhigh effort achieves a similar score to Claude Opus 5 at medium effort—60.5% versus 60.3%—at approximately 80% lower cost per task,” the announcement stated. Artificial Analysis, reported by Matthias Bastian, found intelligence scores staying near GPT-5.6 levels, with gains in some evaluations and regressions in others.

The same afternoon already held a second launch on a ninety-minute lead. Anthropic released Claude Opus 5.5 roughly ninety minutes before OpenAI published. The model cost around 40 percent less to run than Opus 5 and performed close to Fable 5.1 on most work. Neither release was a new frontier model. Both labs took capability that already existed and sold it for less.

Chinese open-weight models from Alibaba, DeepSeek, and Moonshot already did much of the same work for little or nothing. Startups had been moving to those open weights as the bills arrived. The pressure hit on two fronts at once—cheaper tiers and outright cuts on the expensive models—and eroded pricing power on both sides of the closed market. The Sol line traced the same shift in concrete steps. In June, OpenAI had limited GPT-5.6 Sol to twenty government-approved partners. Three months later its successor sat in the full API at half the price.

GPT-6 Sol and GPT-6 Luna became available September 22, 2026 in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise, and Edu users. Free and Go users received Luna in the desktop app. Neither model was yet in ChatGPT Chat. OpenAI said the models would roll out gradually through the day to keep the service stable.

Zac Hall at 9to5Mac cited the ChatGPT desktop app release notes, which described Sol and Luna rolling out at lower token prices than their GPT-5.6 predecessors and using fewer tokens to run. Availability depended on rollout and workspace settings. Enterprise administrators had to enable the new models.

In the OpenAI Developer Community thread that afternoon, VeitB posted, “Please welcome GPT-6 Sol and GPT-6 Luna to the GPT-6 universe.” LarisaHaster replied, “Yay!” Derek Foss wrote that Sol’s claimed drop in errors at these rates made it worth testing against the older models for daily work. Marcus Bell wrote that the lower prices on Sol and Luna should finally let smaller teams run bigger batches without constant cost worries.

Luna appeared in the desktop app for Free and Go users as the day’s rollout continued. The afternoon’s launches left the ceiling of intelligence where it stood. What changed was only the cost of reaching for it, and the windows that would decide who got the cheaper models first were still opening through the day.