GPT-6 Sol and Luna Are Here: OpenAI Launches Cheaper AI Models With Major Upgrades

OpenAI is expanding its GPT-6 lineup with two new AI models designed to make advanced artificial intelligence more affordable: GPT-6 Sol and GPT-6 Luna. Announced in September 2026, the models bring improvements in coding, factual accuracy, computer use, and AI-agent workflows while significantly reducing API costs.

The launch targets developers, businesses, and users who want more capable AI without paying flagship-model prices for every request. GPT-6 Sol is designed for demanding professional tasks and complex coding workflows, while GPT-6 Luna focuses on affordability and efficient performance at scale.

According to OpenAI, both models benefit from training methods similar to those used for GPT-6 Astra, the company’s flagship model. Improvements in inference and prompt caching also help reduce operating costs, making the new models particularly attractive for applications that process large volumes of requests.

What Are GPT-6 Sol and GPT-6 Luna?

GPT-6 Sol and GPT-6 Luna are two models in OpenAI’s GPT-6 family, joining GPT-6 Astra, which remains the company’s top option for the most demanding tasks.

Rather than offering a single model for every workload, OpenAI is building a lineup that balances intelligence, speed, and cost. This approach allows developers to choose a model based on the complexity of their applications and their available budgets.

GPT-6 Sol targets professional work, software engineering, complex reasoning, and AI agents that need to complete multi-step tasks. It is designed for users who need strong results but want more room to experiment without the cost of using the flagship model for everything.

GPT-6 Luna is the more affordable option. It targets workloads where low operating costs matter, including high-volume applications, routine automation, and coding tasks that benefit from capable AI without requiring the most expensive model.

OpenAI says both models improve on their GPT-5.6 counterparts in several areas, including factual reliability, coding performance, computer use, and alignment.

GPT-6 Sol and Luna Pricing: Up to 50% Cheaper

Lower pricing is one of the main selling points of the new models. OpenAI says improvements in caching and inference allow it to offer GPT-6 Sol and Luna at half the promotional API prices of their GPT-5.6 predecessors.

Here is the published pricing per one million tokens:

For comparison, the previous promotional prices were $4 per million input tokens and $20 per million output tokens for GPT-5.6 Sol. GPT-5.6 Luna cost $0.20 per million input tokens and $1.20 per million output tokens.

That means GPT-6 Luna is particularly inexpensive for applications that generate large numbers of responses. GPT-6 Sol offers a stronger option for more demanding workloads while remaining substantially cheaper than a flagship model in many scenarios.

The actual cost of a project depends on the number of tokens processed, reasoning settings, cache usage, and the number of tool calls an agent makes. A cheaper model does not necessarily produce the lowest total cost if it needs more attempts to complete a task successfully.

GPT-6 Sol Brings Major Coding Improvements

Software development is one of the main areas OpenAI emphasizes with the new models. Coding agents increasingly handle more than code completion: they inspect repositories, identify bugs, modify multiple files, run tests, and attempt to deliver changes that are ready for review.

GPT-6 Sol is designed for these longer engineering workflows, where both performance and operating costs matter.

OpenAI reports that GPT-6 Sol achieved a score of 68.8% on DeepSWE 1.1 at maximum reasoning effort. The benchmark evaluates agents on complex software-engineering tasks involving real codebases.

GPT-6 Luna achieved 66.6% on the same evaluation at maximum effort, according to OpenAI. The company says Luna’s performance was comparable to some competing models running at medium effort, at substantially lower cost per task.

These figures are based on OpenAI’s published testing and should not be treated as guarantees for every development environment. Results can vary depending on the benchmark, reasoning settings, tools, and task requirements.

For developers, the potential benefits include:

• Debugging complex errors across multiple files.
• Understanding unfamiliar repositories.
• Refactoring existing applications.
• Generating and improving automated tests.
• Handling multi-step software-engineering tasks.
• Working with Codex and other supported coding-agent workflows.

The combination of improved coding capabilities and lower token prices could make it more practical to run AI agents for longer sessions or use them across a larger number of projects.

Better Factual Accuracy and Professional Workflows

Coding is not the only area receiving attention. OpenAI says GPT-6 Sol makes approximately half as many mistakes as its predecessor on the company’s internal factuality evaluation.

The evaluation uses de-identified ChatGPT conversations in which users had previously flagged factual errors. OpenAI notes that these conversations were selected because they contained mistakes and are not representative of ordinary usage.

GPT-6 Luna also improves on its predecessor, with OpenAI reporting that it can match GPT-5.6 Sol on this evaluation at higher reasoning-effort settings.

These results suggest progress in factual reliability, but they do not mean either model is error-free. Users should continue to verify important claims, especially when the output will be used in technical documentation, research, financial decisions, or other consequential work.

OpenAI also highlights improvements in professional workflows, including tasks involving business applications, research, reporting, and multi-step automation.

In its AutomationBench evaluation, GPT-6 Sol achieved a score of 33.2% at extra-high reasoning effort, with an estimated cost of $0.27 per task. OpenAI says it outperformed several competing configurations on that evaluation at a lower cost.

The results illustrate the company’s broader strategy: deliver useful performance for real-world tasks without requiring customers to use the most expensive model for every interaction.

Improved Computer Use for AI Agents

AI systems are increasingly expected to interact with software rather than simply respond to prompts. This includes navigating interfaces, interpreting screenshots, working with browser-based applications, and completing workflows across multiple tools.

OpenAI says GPT-6 Sol and Luna improve computer-use performance compared with their predecessors.

On the OSWorld 2.0 offline evaluation, GPT-6 Sol at extra-high reasoning effort achieved a score of 60.5%, compared with 60.3% for Claude Opus 5 at medium effort. OpenAI estimates that Sol completed the evaluated tasks at approximately 80% lower cost per task.

GPT-6 Astra remains OpenAI’s recommended model for users seeking its strongest computer-use capabilities. Sol and Luna instead offer more cost-efficient options for workflows where affordability is an important consideration.

For businesses, these capabilities could support automated data entry, software testing, repetitive administrative tasks, and interactions with web applications. Actual results will depend on the tools available to the agent and the reliability of the surrounding automation system.

Prompt Caching Makes Long AI Sessions Cheaper

OpenAI is also improving prompt caching for developers building applications with GPT-6.

Prompt caching allows a system to reuse previously processed portions of a prompt instead of repeatedly processing the same context from scratch. This is especially useful for coding agents and applications that send long instructions, project documentation, or conversation histories with each request.

OpenAI says its caching improvements help deliver higher cache-hit rates by default, with discounts of up to 90% on cached input-token reads.

Developers also receive more control over caching behavior. They can monitor cache performance, diagnose missed caching opportunities, and adjust reasoning effort or tool availability without necessarily losing the benefits of previously cached context.

For AI agents that operate over long sessions, these improvements can reduce costs and help responses arrive faster. The savings depend on how much context can be reused and how the application structures its requests.

GPT-6 Sol vs. GPT-6 Luna: Which Model Should You Choose?

The two models are designed for different priorities.

Choose GPT-6 Sol when a task requires more sophisticated reasoning, complex code changes, or demanding professional workflows. GPT-6 Luna is worth considering for routine automation, large-scale processing, and applications where keeping per-request costs low is essential.

For especially difficult tasks where maximum performance matters more than price, OpenAI continues to position GPT-6 Astra as its flagship model.

Availability: How to Access GPT-6 Sol and Luna

OpenAI initially announced GPT-6 Sol and GPT-6 Luna for ChatGPT Work and Codex, with access for eligible Plus, Pro, Business, Enterprise, and Edu users. At launch, Free and Go users could access GPT-6 Luna in the desktop app, while the models were not yet available in the standard Chat experience.

Developers can also access the models through the OpenAI API using these model identifiers:

• gpt-6-sol
• gpt-6-luna

Availability may differ by product, subscription, region, and rollout status. OpenAI has since announced GPT-6.1 Sol, so developers should check the current model documentation before choosing a model for a new integration.

Official resources:

• OpenAI’s GPT-6 Sol and Luna announcement
• OpenAI API documentation
• OpenAI platform

Final Thoughts

GPT-6 Sol and GPT-6 Luna reflect OpenAI’s effort to make advanced AI more practical for everyday professional work and large-scale applications. Both models offer improvements in coding, factual reliability, computer use, and AI-agent workflows, while lower API prices make sustained use more affordable.

GPT-6 Sol is the more compelling option for demanding engineering and professional tasks, while GPT-6 Luna targets users who prioritize low costs and efficient performance. Improved prompt caching adds another potential source of savings for applications that reuse large amounts of context.

Neither model eliminates the need to test outputs or compare performance against alternatives. However, for developers building AI-powered products and businesses automating repetitive work, the combination of capability and lower costs makes the GPT-6 Sol and Luna launch worth watching.

Learn more: Introducing GPT-6 Sol and Luna