GPT-6 Sol and Luna: What OpenAI's Cheaper Model Tiers Mean for Automation

What OpenAI released on September 22, 2026

OpenAI released updated versions of GPT-6 Sol and GPT-6 Luna on September 22, 2026, roughly three weeks after GPT-6 Astra shipped on September 3-4, 2026. Sol and Luna are OpenAI's smaller, cheaper tiers, API pricing for both dropped 50% compared to their GPT-5.6 predecessors, and OpenAI says they were trained with methods similar to Astra's.

Astra sits at the top of the current GPT-6 lineup as the flagship model. Sol is the mid tier, built for coding and agentic work. Luna is the smallest tier, meant for high-volume, low-complexity tasks. That is the full picture of what OpenAI confirmed this week: a three-model lineup, not four. Some outlets have floated a future fourth "Terra" tier, but OpenAI has not confirmed one, so treat that as speculation, not fact.

Why OpenAI shipped two cheaper models instead of one bigger one

OpenAI is chasing the same problem every model lab is chasing right now: token costs at scale. Sol and Luna let OpenAI sell Astra-level training methods at a fraction of Astra's price, aimed at the high-volume, repetitive work that doesn't need a flagship model. Anthropic released Claude Opus 5.5 the same day, cut about 40% for typical workloads, according to Gizmodo (September 22, 2026).

That timing was not a coincidence. Both labs are racing to make everyday, high-frequency AI usage affordable, because that's where the bulk of API spend for most companies actually lives, not in the flashy one-off flagship queries.

What Sol and Luna are actually for, according to OpenAI

According to OpenAI's own announcement, cited by TechCrunch (September 22, 2026), Sol is designed for complex tasks like coding, while Luna is built for "high-volume tasks with a clear goal, like summarizing documents, extracting information, or answering quick questions." That's a direct split between agentic, multi-step reasoning work and repetitive, single-shot classification work.

OpenAI also claims a specific factuality improvement for Sol. Per the company's internal evaluation, based on de-identified real-world conversations where users flagged mistakes, GPT-6 Sol makes about half as many mistakes as its predecessor, reaching what OpenAI calls "Astra-level reliability" at a lower cost (TechCrunch, September 22, 2026). That's OpenAI's own internal benchmark, not an independent audit, so treat it as a vendor claim until a third party replicates it.

OpenAI also published a computer-use comparison: on OSWorld 2.0 offline, GPT-6 Sol at "xhigh" effort scored 60.5% versus Claude Opus 5 at medium effort's 60.3%, at roughly 80% lower cost per task, and GPT-6 Luna at max effort reportedly beat GPT-5.6 Sol at medium effort at about one-tenth the cost (9to5Mac, September 22, 2026, citing OpenAI's announcement). Again, this is OpenAI's own benchmark, run on OpenAI's chosen settings, not a neutral third-party test.

Availability: who gets access and when

Sol and Luna are rolling out to ChatGPT Work and Codex for Plus, Pro, Business, Enterprise, and Edu accounts, plus the API, starting September 22, 2026 (9to5Mac, TechCrunch). Free and Go users get access to Luna through the desktop app. Enterprise workspace admins need to enable the new models manually before their teams see them.

What a tiered model lineup means for how you should route automation work

A three-tier lineup (flagship, mid, small) is a routing decision waiting to happen, not just a pricing update. If you're running automations that call a model dozens or hundreds of times a day, using the same flagship model for every step is the most common way teams overspend on tokens without getting better output.

The practical split, based on OpenAI's own stated positioning for these tiers, looks like this:

• High-volume classification, tagging, routing, and data extraction: a small/cheap tier (Luna-class) is designed for exactly this. If a task has a clear, narrow goal and doesn't require multi-step reasoning, it's a candidate for the cheapest tier that still hits your accuracy bar.
• Complex, multi-step agent work, coding, and tool-calling chains: a mid tier (Sol-class) is where OpenAI is positioning its factuality gains, which makes it a reasonable default for agent workflows that need real reliability without flagship pricing.
• Judgment calls, ambiguous edge cases, and anything customer-facing where an error is costly: keep the flagship tier (Astra-class) in the loop, either as the default or as an escalation path when the cheaper tier's confidence is low.

We have not independently benchmarked Sol or Luna against Astra or against Claude's tiers for our own automation workloads. Treat vendor-published benchmarks as a starting point for testing, not a substitute for testing on your own data and your own success criteria. The right tier for a given task depends on your accuracy requirements, your volume, and what a mistake actually costs you, and that only shows up when you run your own workload through it.

The one thing to do next

If you're running an automation pipeline that calls the same model for every step regardless of task complexity, audit which steps are high-volume and low-complexity versus which ones need real reasoning. That single change, routing cheap-tier models to the repetitive steps and reserving flagship models for the judgment calls, is usually the fastest way to cut token spend without touching output quality on the steps that matter.

If you want help mapping your current automation stack to the right model tier for each step, talk to AY Automate about an AI automation build.