GPT-6 Sol and Luna Halve OpenAI's API Prices

OpenAI released GPT-6 Sol and GPT-6 Luna on 22 September 2026 at roughly half the API price of the GPT-5.6 models they replace. The discount is the headline, but the more useful detail is buried in the coverage: OpenAI confirmed to VentureBeat that these rates are permanent, not promotional. For anyone who has tried to budget an AI feature twelve months out, that single word is worth more than the percentage.

The rates

Sol is 50% cheaper in both directions. Luna is 50% cheaper on input and about 58% cheaper on output. In the API the models are gpt-6-sol and gpt-6-luna, and OpenAI says both were trained with methods similar to those behind GPT-6 Astra, extending Astra's gains in professional work, factuality, coding, computer use and alignment into faster and cheaper models.

Why permanence is the story

When we covered the GPT-6 Astra launch earlier this month, the sharpest thing on the pricing page was a footnote: GPT-5.6 Sol's attractive rate was promotional and scheduled to end on 21 November 2026. A promotional rate is not a planning input. You cannot sign a twelve month client contract, or size a free tier, or decide whether a feature is viable at scale, on a number that expires in eight weeks and reverts to double.

That footnote is gone. The new rates are the rates. If you deferred a decision in early September because the economics might revert in November, the decision is unblocked, and the number to plan against is $2 and $10 rather than $4 and $20.

What GPT-6 Sol and Luna are each for

OpenAI positions Sol for complex work that developers and knowledge workers repeat: building features, reviewing code, debugging and analysing data. Luna is the high-volume option for tightly defined jobs such as summarisation, extraction and answering straightforward questions.

Luna at $0.10 per million input tokens changes which workloads are worth automating at all. Document extraction over a large archive, classification passes over a support queue, or enriching a CRM record by record are jobs where the per-item cost used to decide whether the project happened. At ten cents per million input tokens, the cost of the model stops being the constraint and engineering time becomes the constraint again. The familiar tiering pattern behind this split is the same one we described in why models ship in mini, flash and lite versions.

On the numbers OpenAI published and VentureBeat reported, Sol scores 33.2% on AutomationBench 1.0.6 at $0.27 per task, and on DeepSWE 1.1 Sol reaches 68.8% with Luna close behind at 66.6%. A cheap model landing within a couple of points of the expensive one on a coding suite is the kind of result that should send you to test the cheap one on your own tasks rather than trusting either score.

Availability

Both models are in the API from launch. In ChatGPT Work and Codex the rollout covers Plus, Pro, Business, Enterprise and Edu users from 22 September, and Free and Go users reach Luna through the desktop app.

The number that still moves

A 50% rate cut is not a 50% bill cut unless your token counts stay the same, and the two labs that cut prices on the same day both made claims about doing more work in fewer steps. Price per token and cost per completed task are different measurements, and the second one is the one on your invoice. We cover the gap in what token efficiency means in practice, and the arithmetic for sizing a workload before you run it in how to estimate tokens for an AI task.

Anthropic shipped Claude Opus 5.5 the same morning, which CNBC framed as both labs responding to pressure from open-weight competitors. Two permanent price cuts in a day is unusual enough to be worth a calendar reminder to re-run your own cost model, and our guide to keeping up with AI news covers how to do that without checking a feed all day.

Update, 29 September 2026: OpenAI released GPT-6.1 Sol at DevDay at $2 and $10 per million tokens. The GPT-6.1 Sol launch summary has the pricing and availability.