Quick Summary
- OpenAI has released GPT-6 Sol and GPT-6 Luna, two new models built to bring frontier-level AI to everyday tasks at a lower cost.
- Both models use training methods similar to GPT-6 Astra, OpenAI’s top-tier model released earlier this month.
- API prices for Sol and Luna are 50% lower than their GPT-5.6 predecessors.
- GPT-6 Sol outperforms competing models on business workflow and coding benchmarks while costing significantly less per task.
- Prompt caching improvements now offer up to 90% discounts on reused context for agents and long conversations.
- Both models show measurable gains in factual accuracy and alignment compared to their predecessors.
- GPT-6 Sol and Luna are rolling out now in ChatGPT Work, Codex, and the OpenAI API.
OpenAI has introduced GPT-6 Sol and Luna, two new additions to its GPT-6 model family. These models are designed to bring the same underlying advances found in GPT-6 Astra to more affordable and efficient versions. The goal is simple. Not every task needs the most powerful model available, and many teams need strong performance at a lower cost.
GPT-6 Astra, released earlier this month, was positioned as OpenAI’s most capable and aligned model to date. It remains the top choice for the most demanding projects. Everyday work happens at different speeds and budgets. That is where Sol and Luna come in.
What Are GPT-6 Sol and Luna
GPT-6 Sol and GPT-6 Luna are two new models trained using methods similar to those behind GPT-6 Astra. They carry forward many of Astra’s strengths in professional tasks, factual accuracy, coding, and computer use, but at a fraction of the cost. Sol sits between Astra and Luna in terms of raw capability, while Luna is the most budget-friendly option in the lineup.
According to OpenAI, these models are meant to distribute the benefits of frontier intelligence more broadly. Rather than reserving high-level performance for the most expensive tier, the company says it is passing efficiency gains from better caching and inference infrastructure directly to users.
GPT-6 Sol and Luna Pricing
One of the most notable changes is cost. OpenAI has cut API prices for both models by 50% compared to their GPT-5.6 promotional pricing.
GPT-6 Sol now costs $2 per million input tokens and $10 per million output tokens, down from $4 and $20 under GPT-5.6 Sol. GPT-6 Luna costs even less, at $0.10 per million input tokens and $0.50 per million output tokens, compared to $0.20 and $1.20 previously. These prices apply to the OpenAI API and are aimed at developers who need to run large volumes of requests without high recurring costs.
GPT-6 Astra remains the most expensive and most capable model in the family. OpenAI positions it as the choice for teams that want the best possible results regardless of cost.
Performance in Professional Work
OpenAI shared benchmark results showing GPT-6 Sol performing well against competing models on real-world business tasks. On AutomationBench, a test that measures how AI agents handle workflows across common business apps, GPT-6 Sol at its highest effort setting outperformed Claude Opus 5 at its own highest setting. Notably, Sol did this at roughly 9% of Opus 5’s cost per task.
GPT-6 Luna also showed gains. At high effort, it improved on its predecessor by 5.4 percentage points while costing 58% less per task. GPT-6 Sol also surpassed Claude Fable 5.1 on the same benchmark at a much lower price, and even outperformed a lower-effort version of GPT-6 Astra itself.
On a separate evaluation called Agents’ Last Exam, which tests AI agents on complex professional workflows across more than 50 industries, GPT-6 Sol scored 56.4% at maximum effort. That score topped Claude Opus 5’s best result on the same test, while costing 60% less per task.
Factual Accuracy Improvements
Getting facts right remains one of the biggest challenges for AI models, and OpenAI says both Sol and Luna show real progress here. Using an internal evaluation based on real conversations where users had previously flagged factual errors, GPT-6 Sol made roughly half as many mistakes as its predecessor. That puts it close to Astra-level reliability, but at a much lower cost.
GPT-6 Luna also improved substantially. At higher effort levels, it now matches the factual accuracy of GPT-5.6 Sol while costing about a hundredth as much to run. OpenAI notes that the conversations used for this evaluation were selected specifically because they contained past errors, so they are not representative of typical everyday use, where mistakes are far less common.
Coding Performance
AI coding assistants have taken on increasingly complex and long-running tasks over the past year, and OpenAI reports that internal usage of these tools has grown sharply within the company. Because of this, the cost of running coding agents over extended periods matters more than ever.
On FrontierCode, a benchmark that evaluates whether AI-generated code is ready to merge into real codebases, GPT-6 Sol improved significantly over its predecessor and matched the performance of Claude Fable 5.1 at a much lower cost. On DeepSWE, another benchmark focused on complex software engineering tasks, GPT-6 Sol scored 68.8% at maximum effort. That is within about one percentage point of Claude Fable 5’s best score, achieved at roughly 80% lower cost per task.
GPT-6 Luna also performed well here, scoring 66.6% on the same benchmark. That result is comparable to Claude Opus 5 and Fable 5 at medium effort, but Luna achieved it while costing 93% less than Opus 5 and 96% less than Fable 5 per task.
Computer Use Capabilities
Computer use refers to an AI model’s ability to operate a computer directly, navigating software and completing tasks the way a person would. GPT-6 Astra remains OpenAI’s top performer in this category, but Sol and Luna offer stronger cost efficiency than their predecessors.
On OSWorld 2.0, a benchmark that tests long, multi-step computer tasks, GPT-6 Sol achieved a score close to Claude Opus 5, 60.5% compared to 60.3%, while costing about 80% less per task. GPT-6 Luna also outperformed GPT-5.6 Sol on the same test while costing roughly one tenth as much.
Communication Style
Alongside performance gains, OpenAI says it has carried over improvements in communication style from GPT-6 Astra to both Sol and Luna. Users should notice clearer answers with less jargon and fewer unnecessary details, particularly in technical and coding conversations.
Responses are also expected to be somewhat shorter overall, without losing important substance. OpenAI shared an example comparing GPT-5.6 Sol and GPT-6 Sol responding to the same coding request, noting that the newer model gave a more direct and less repetitive answer.
Prompt Caching for Agents
Beyond lower token prices, OpenAI has made changes to how prompt caching works for GPT-6 models. Improved caching allows developers to reuse more context across requests, which speeds up response times and reduces costs. Cached input tokens now come with a 90% discount compared to fresh processing.
OpenAI also introduced new tools to help developers manage this more effectively. A prompt caching dashboard shows how much input is being cached over time. A diagnostics tool helps identify missed caching opportunities.
Developers can also adjust reasoning effort mid-conversation without losing cached context. They can set explicit breakpoints to control which parts of a prompt get cached.
According to OpenAI, GitHub reported that these caching improvements reduced the share of prompt tokens requiring fresh processing by more than 50% across billions of requests. That change helped tools like Copilot respond faster to everyday coding requests.
Alignment and Safety
GPT-6 Sol and Luna also build on the alignment work introduced with GPT-6 Astra, which OpenAI describes as its most aligned model so far. Internal evaluations show both new models improving over their GPT-5.6 counterparts, including lower rates of misleading claims about coding tasks.
OpenAI notes that these alignment evaluations are designed to test especially difficult situations and do not reflect failure rates under typical use. Full results are available in the model’s system card for readers who want more detail.
Availability
GPT-6 Sol and Luna are rolling out now in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise, and Edu users. Free and Go tier users can access GPT-6 Luna through the desktop app. Neither model is yet available in standard Chat.
In the OpenAI API, developers can access the models using the identifiers gpt-6-sol and gpt-6-luna. OpenAI says it plans to roll out these models gradually throughout the day to keep service stable, so some users may need to check back later if the models are not immediately visible.
Final Thoughts
GPT-6 Sol and Luna represent OpenAI’s push to make frontier-level AI more accessible for everyday work. By pairing strong performance in coding, professional tasks, and factual accuracy with significantly lower prices, the company is aiming to close the gap between top-tier capability and practical, everyday affordability.
For readers following the pace of AI development, this release is a reminder that progress is not only about building the most powerful model. It is also about making strong models cheaper and more usable for the tasks people handle every day. As GPT-6 Sol and Luna roll out across ChatGPT and the API, more users will get direct access to that shift.
Discover how AI is reshaping technology, business, and healthcare—without the hype.
Visit InfluenceOfAI.com for easy-to-understand insights, expert analysis, and real-world applications of artificial intelligence. From the latest tools to emerging trends, we help you navigate the AI landscape with clarity and confidence
