OpenAI's GPT-6 Sol and Luna arrive on Azure and AWS at lower prices

OpenAI on 22 September 2026 introduced GPT-6 Sol and GPT-6 Luna, two models it describes as bringing frontier intelligence to everyday work with different balances of capability and cost. They sit below GPT-6 Astra, the most capable and most expensive model in the family.[1][3]

On the same day, Microsoft and Amazon Web Services said both models were generally available on their AI platforms, Microsoft Foundry and Amazon Bedrock. For many companies, that means the models can be bought through an existing cloud account rather than a separate OpenAI contract.[2][3]

AWS describes Sol as a model for demanding tasks that recur through the week: writing features, debugging, reviewing code, analysing data and completing multistep work across tools. It says OpenAI's own internal factuality test found Sol made about half as many factual mistakes as its predecessor, GPT-5.6 Sol. That figure comes from OpenAI and has not been independently checked.[3]

Luna is the smaller, faster sibling. Microsoft recommends it for extraction, summarisation, routing requests and routine customer conversations, while AWS stresses workloads that run thousands of times a day. Both clouds suggest mixing models in one system, for example Luna to sort incoming requests, Sol to investigate harder cases and Astra only where deeper reasoning changes a decision.[2][3]

Microsoft published full rates for its Global Standard tier. Sol costs US$2 per million input tokens and US$10 per million output tokens for short prompts. Luna costs US$0.10 and US$0.50. Astra costs US$10 and US$50. Rates roughly double for long-context requests, and Microsoft's US and EU data-zone deployments carry a premium over Global.[2]

AWS did not print a rate card in its post, saying only that both models come at significantly lower API pricing than their GPT-5.6 predecessors. Microsoft urged buyers to look at cost per task rather than price per token, because a model that needs fewer retries can end up cheaper even at a higher rate.[3][2]

Both models support prompt caching, which lets repeated instructions, policies or reference documents be reused instead of processed again on every call. On Microsoft's Global tier, cached input for Luna is listed at US$0.01 per million tokens. For a support desk that sends the same policy text with each question, that is where much of the saving is likely to sit.[2][3]

AWS says inference data on Bedrock is not used for model training and customers do not have to opt in to share data with OpenAI. Traffic flagged by its abuse classifiers can be kept for up to 30 days, and zero data retention is available on request through an AWS account team.[3]

We could not read OpenAI's own launch page, which refuses automated readers, so its title, date and summary were confirmed from OpenAI's official news feed. OpenAI's direct API prices for Sol and Luna are therefore not reported here, and they may differ from Microsoft's cloud rates.[1]

Location is the bigger gap for Malaysian buyers. Microsoft says standard deployment covers all 28 of its Global regions plus US and EU data zones, but it does not list the regions or offer an Asian data zone. AWS points readers to its documentation for supported regions without naming any in the post.[2][3]