OpenAI Launches GPT-6 Sol and Luna as Faster, Lower-Cost AI Models

Picture a local council’s back-office team, grinding through thousands of planning documents a week. Summarise this. Classify that. Route it to the right department. It’s the kind of repetitive, high-volume work that takes human hours and costs real money. On 23 September 2026, OpenAI announced two new models that are squarely aimed at exactly that sort of task — and at cutting the cost of doing it.

GPT-6 Sol and GPT-6 Luna arrived with a headline number: API prices cut by 50% compared with the promotional pricing of their predecessors, GPT-5.6 Sol and GPT-5.6 Luna. OpenAI said the reduction was made possible by improvements in caching and inference efficiency. In plain terms, the models do the same work with less computational effort — and OpenAI is passing some of that saving on.

Two Models, Two Jobs

The distinction between Sol and Luna is deliberate. GPT-6 Sol is built for the harder stuff — demanding reasoning, complex coding, and what OpenAI calls agentic workflows, where the model takes a sequence of actions rather than just answering a single question. GPT-6 Luna is built for speed and scale: summarisation, data extraction, classification, routing. High volume, lower complexity, lower cost.

Think of it as the difference between hiring a specialist consultant and running a well-trained production line. Both are useful. Neither replaces the other.

Pricing on the API reflects that split. GPT-6 Sol costs around £1.57 per million input tokens and around £7.87 per million output tokens. GPT-6 Luna, by contrast, comes in at around £0.08 per million input tokens and around £0.39 per million output tokens — figures that make large-scale automated processing far more commercially accessible than it was even a few months ago. (OpenAI publishes prices in US dollars; GBP equivalents are approximate based on current exchange rates.)

Both models support a context window of around 1.05 million tokens — enough to process a very long document, or a large batch of shorter ones, in a single pass. Sol and Luna both cap their maximum output at 128,000 tokens.

Where You Can Use Them

The models are available now through the OpenAI API as `gpt-6-sol` and `gpt-6-luna`. Enterprise, Business, Pro, Plus and Edu subscribers can access both through ChatGPT Work and through Codex, OpenAI’s coding-focused product. Free and Go users get access to GPT-6 Luna through the desktop app, though OpenAI noted the models were not initially available in the standard ChatGPT product.

AWS also confirmed general availability of both models on Amazon Bedrock on the same day, meaning organisations already running workloads on Amazon’s cloud infrastructure can plug them in without switching platforms.

Where GPT-6 Astra Still Sits

Sol and Luna are not replacements for GPT-6 Astra. OpenAI’s own model guidance is clear on that. Astra remains the most capable model in the GPT-6 family, intended for the most complex end-to-end tasks. OpenAI has also disclosed that Astra reached the Critical level of cybersecurity capability under its internal Preparedness Framework — a designation that carries its own effect on how the model is handled and deployed.

Sol and Luna sit below Astra in capability but above it in speed and cost-efficiency. The choice between them is a trade-off, not a hierarchy of quality alone.

A Note on What the Announcement Doesn’t Say

OpenAI’s announcement sets out pricing and positioning clearly. What it doesn’t provide — and what independent observers have pointed out — is verified, third-party evidence of real-world accuracy, reliability, or bias performance across specific use cases. The 50% price reduction and the efficiency claims come from OpenAI’s own product documentation. Reuters reported the same headline figures, but the underlying performance data remains primarily OpenAI’s own.

That matters especially for public-sector and regulated-industry deployments, where procurement decisions can’t rest on a company’s self-description alone. Any organisation handling personal data, health information, or commercially sensitive material would need to work through data protection obligations under the Data Protection Act 2018 and UK GDPR, as well as any sector-specific requirements, before putting either model to work.

Sam Altman, OpenAI’s chief executive, has spoken previously about the company’s goal of making powerful AI accessible at lower cost. But the real test for Sol and Luna will come in how organisations actually deploy them — and whether the efficiency gains hold up under the pressure of real workloads.

What This Means for Kent Residents

For Kent businesses, councils, schools, and NHS organisations, the arrival of GPT-6 Sol and Luna at much lower API prices means that automated document processing, customer-service tools, and administrative workflows become more affordable to run at scale — though any public-sector deployment by bodies such as Kent County Council, Medway Council, or NHS Kent and Medway ICB would need to satisfy data protection, procurement, and cybersecurity requirements before going live. Kent residents may encounter these models indirectly, through services and systems that organisations choose to build on top of them. The UK’s regulator-led approach to AI governance means there is no single sign-off process, so the burden of responsible deployment falls on individual organisations and their legal teams.