OpenAI ships GPT-6 Sol and Luna at half the price of GPT-5.6

OpenAI has released GPT-6 Sol and GPT-6 Luna, the two cheaper models in its GPT-6 family, and cut their API prices in half compared with the GPT-5.6 promotional pricing [1]. GPT-6 Sol now costs $2 per million input tokens and $10 per million output tokens; GPT-6 Luna costs $0.10 and $0.50 [1]. For anyone who runs agents, automations or an app on OpenAI’s API, the bill for the same volume of tokens drops by half [1].

What was announced

OpenAI introduced GPT-6 Astra earlier this month and calls it its most capable model [1][2]. Sol and Luna sit below it. TechCrunch describes Sol as the tier for complex tasks such as coding, while OpenAI positions Luna for “high-volume tasks with a clear goal, like summarizing documents, extracting information, or answering quick questions,” according to TechCrunch [2]. OpenAI says it trained both with methods similar to those used for Astra, and that improvements in caching and inference let it serve them at lower cost [1].

The price change is the part that is concrete and checkable. OpenAI’s own table shows GPT-6 Sol at $2 input and $10 output per million tokens, against $4 and $20 for GPT-5.6 Sol, and GPT-6 Luna at $0.10 and $0.50, against $0.20 and $1.20 for GPT-5.6 Luna [1]. OpenAI labels both reductions 50% cheaper [1]. Simon Willison adds two useful details: cached input on GPT-6 Sol is $0.20 per million and on Luna $0.01 per million, and GPT-5.6 has a 25% price increase scheduled for November, so the 50% cut is measured against promotional pricing that was due to end anyway [3].

OpenAI also changed prompt caching, which matters most for agents and long conversations that resend the same context. The company says caching for GPT-6 now delivers higher hit rates by default, with a 90% discount on cached input-token reads [1]. It lists three developer changes: a Prompt Caching Dashboard and a diagnostics tool that explain missed caching; the ability to change reasoning effort or enable and disable tools without breaking the cache; and explicit breakpoints that let developers choose where a cached prefix ends [1]. OpenAI says GitHub reports these improvements have reduced the share of prompt tokens needing fresh processing by more than 50% across billions of requests over several months [1].

On quality, the claims are OpenAI’s and run on its own terms. It says GPT-6 Sol makes about half as many mistakes as its predecessor on an internal factuality test built from real conversations in which users flagged errors, and that Luna at higher effort matches GPT-5.6 Sol on that test at about a hundredth of the cost [1]. On AutomationBench, a test of business workflows across 47 tools, OpenAI reports GPT-6 Sol at xhigh effort scoring 33.2% at $0.27 per task, ahead of Claude Opus 5 at max effort at 9% of its cost per task [1]. On DeepSWE it reports Sol at 68.8%, within 1.1 percentage points of Claude Fable 5’s 69.9%, at about 80% lower cost per task, and Luna at 66.6% [1]. On OSWorld 2.0 offline, a computer-use test, it reports Sol at 60.5% against 60.3% for Claude Opus 5 at medium effort [1]. OpenAI notes that competitor scores come from public reports and that it used Fable 5 scores where Fable 5.1 scores were unavailable [1]. These comparisons are against Anthropic’s previous models, and TechCrunch points out that Anthropic released Claude Opus 5.5 about 90 minutes before OpenAI’s announcement [2].

What changed, and for whom

Availability is split. OpenAI says both models are in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise and Edu users, rolling out gradually through the day of launch [1]. Free and Go users can use GPT-6 Luna in the desktop app, and the models are not yet available in Chat [1]. In the API they are named gpt-6-sol and gpt-6-luna [1]. TechCrunch reports the same rollout, including Luna for Free and Go users [2].

The competitive picture is where the change bites. Simon Willison’s price table puts GPT-6 Sol at $2 input and $10 output, the same input price as Grok 4.7 and GPT-5.6 Terra, and below Claude Opus 5.5 at $4 and $20 [3]. He notes that with GPT-5.6 Terra priced the same as GPT-6 Sol, the remaining reasons to use Terra “just evaporated” [3]. At the top end, GPT-6 Astra and Claude Fable 5.1 both sit at $10 input and $50 output per million in his table [3]. In practice, most day-to-day agent and automation work now has at least three credible vendors at or below $4 per million input tokens, which makes routing work between models a cost decision you can revisit monthly rather than a one-time choice.

OpenAI also says Sol and Luna carry over the communication style introduced with Astra, with slightly shorter answers, less jargon and fewer low-value details [1]. Shorter output means fewer output tokens per task, so the effective saving on some workloads could be larger than the list-price cut; OpenAI does not publish a figure for that, so treat it as something to measure rather than assume. On safety, OpenAI says both models improve on their GPT-5.6 counterparts in its alignment evaluations, including lower rates of misleading claims about their coding work, and points to the system card for full results [1].

Who it is for — and not

This release matters most to people with a real API bill: developers running agents that resend long context, small teams with automations calling OpenAI in the background, and anyone who built on GPT-5.6 Sol or Luna and has not looked at costs since. For them, a 50% cut on list price plus a 90% discount on cached reads is worth an afternoon of re-costing [1]. The caching changes also remove a common source of waste: switching reasoning effort or toggling tools mid-conversation used to break the cache, and OpenAI says it no longer does [1].

It matters less if you use ChatGPT through a flat subscription, where the per-token price is not what you pay. For those users the relevant change is which model the Work tab and Codex now run, and whether Luna in the Free and Go desktop app is enough for simple tasks [1][2]. And the benchmark tables should be read for what they are: vendor-run comparisons, often against competitor models that have since been replaced [1][2]. The prices and the model names are the facts to build on; the benchmark margins are not.

One more caution. Willison’s point about the November price rise cuts both ways [3]. If you stay on GPT-5.6 Sol or Luna, your costs go up by 25% in November; if you move, you should test the new models on your own work first, because a model that is cheaper per token but needs more retries is not cheaper per task.

We will update this page if OpenAI changes GPT-6 Sol or Luna pricing, brings the models to Chat, or publishes more detail on the GPT-5.6 price change in November.

∴ Mid-tier OpenAI tokens now cost half as much; re-price any workload you last costed on GPT-5.6.