OpenAI launches GPT-6 Sol and Luna at half the previous price

What happened

OpenAI's newest model family is led by GPT-6 Astra, its most capable and most expensive model. Sol and Luna follow it as lower-cost options. According to MacRumors, both were built with methods similar to Astra's and improve on their predecessors in professional work, factual accuracy, coding, computer use and alignment, which means behaving as their designers intend. Sol is meant for everyday work and reasons more deeply. Luna is cheaper and tuned for quick answers and large volumes, such as summarising documents or pulling data out of text.

The prices, as reported by MacRumors, are 2 dollars per million input tokens and 10 dollars per million output tokens for Sol. Luna costs 0.10 dollars for input and 0.50 dollars for output, down from 0.20 and 1.20 dollars for the previous Luna. Reading input from the cache, where a repeated part of a prompt is stored and reused, gets a 90 percent discount. OpenAI's forum post described the API prices as 50 percent lower than the earlier promotional prices, and an OpenAI staff member called the cut permanent.

OpenAI says Sol makes about half as many mistakes as GPT-5.6 Sol on its internal tests, and improves on it in coding and computer use. MacRumors reported that both new models use Astra's style of simpler language, less jargon and slightly shorter answers. Both are available in the API. In ChatGPT and the Codex coding tool they rolled out to paying Plus, Pro, Business, Enterprise and Education users, and free users can try Luna in the desktop app.

The engineering behind it

Language models are priced per token, a small piece of text that is often part of a word. Input tokens are what you send, and output tokens are what the model writes back. Output usually costs more, because the model must generate it one token at a time. This pricing makes cost a design question. A system that sends long instructions with every request, or asks for long answers, can cost many times more than one that is designed carefully.

A simple calculation shows how cheap high-volume work has become. Suppose a project classifies 10,000 short messages. If each request uses 500 input tokens and returns 20 output tokens, the total is 5 million input tokens and 200,000 output tokens. At Luna's listed prices, that is about 0.50 dollars for input and 0.10 dollars for output. The same job on Sol would cost about 12 dollars. These are example figures based on the published prices, not measured costs.

Price per token is not the whole story. MacRumors pointed out that Sol is half the price per token of Anthropic's Claude Opus 5.5, released the same day, but that Opus 5.5 uses fewer tokens per task, so the two cannot be compared directly. A model that reasons longer may use many more tokens to reach an answer. The useful measure is the total cost to finish a task correctly, which can only be found by testing on real work.

Users on OpenAI's developer forum shared their own early tests. One ran three small coding tasks on each model. Both new models passed all three, and the Luna runs cost well under one cent each. Another user reported that Sol at its highest effort setting made little progress on some problems while costing more. These are informal reports from individual users, but they show the kind of testing developers do before choosing a model.

What it means in Nepal

The sources do not mention Nepal. What they show applies to any student or developer with an API account: the price of using a capable language model has fallen sharply, and the cheapest tier now costs very little for simple tasks. A project that sorts customer messages, extracts fields from forms or summarises reports can be tested on thousands of examples for a few dollars, as the example calculation above shows.

This makes cost estimation a practical skill. Before running a large job, a careful developer counts the tokens in a typical request, multiplies by the number of requests, and compares models on a small sample first. They also decide whether a cheap, fast model is good enough or whether the task needs a stronger one. Many real systems use both: a cheap model for most requests and a stronger one for difficult cases.

Prices and models change often. OpenAI's forum already listed newer releases within weeks of this launch. Anyone planning a project should check the current price page before starting, keep the code flexible enough to switch models, and record the costs they actually see. That record is often more reliable than any published benchmark.