GPT-6 Sol and Luna: Features, Price, Performance and What’s New

 

On September 23, 2026, OpenAI officially introduced GPT-6 Sol and GPT-6 Luna as the newest additions to the GPT-6 model family. Rather than relying on a single monolithic system, OpenAI has split the capabilities associated with its flagship GPT-6 Astra into two specialized tiers designed to solve the two biggest operational hurdles for developers: deep reasoning capability and deployment cost.

The primary headline for developers is pricing. OpenAI says API prices for both Sol and Luna represent a 50% reduction compared with GPT-5.6 promotional pricing. Paired with a redesigned prompt caching architecture offering discounts of up to 90%, this release targets the financial friction of running persistent, multi-step AI agents and high-volume data pipelines.

Key Takeaways

• Two specialized models: GPT-6 Sol is built for demanding coding, complex reasoning, and agentic workflows; GPT-6 Luna is optimized for high-volume, cost-sensitive execution.
• 50% API price reduction: OpenAI says standard API token rates reflect a 50% cut compared with GPT-5.6 promotional pricing.
• Prompt caching discounts: Eligible cached input-token reads receive discounts of up to 90% within a 30-minute reuse window.
• Improved factuality on edge cases: OpenAI reports GPT-6 Sol reduced errors by approximately half on its internal evaluation of historically error-prone conversations.
• Rollout: Available via the API and in ChatGPT Work and Codex plans, with Luna accessible in the desktop app for Free and Go tiers.

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

The introduction of Sol and Luna reflects how enterprise AI architecture has matured. In real-world software, not every task demands the highest tier of reasoning. Running simple data extraction through a frontier model wastes budget, while routing complex full-stack coding tasks through a lightweight model leads to logic breakdown.

GPT-6 Sol is engineered for high capability. OpenAI positions it as the model built for demanding professional tasks, full-stack software development, and long-running agentic workflows that require sustained context and logical coherence.

GPT-6 Luna is engineered for efficiency and scale. It is designed to handle high-throughput, latency-sensitive tasks where fast execution and minimal cost are paramount, offering a lightweight entry point into the GPT-6 ecosystem.

What Is Different Between Sol and Luna?

The core difference lies in how compute is balanced between depth and throughput. While Sol is given the architectural headroom to evaluate difficult edge cases and navigate large codebases, Luna is streamlined to deliver rapid, low-cost responses across repetitive tasks.

Feature	GPT-6 Sol	GPT-6 Luna
Primary focus	Complex reasoning, software engineering, persistent agents	High-volume processing, focused tasks, cost efficiency
Standard input price	$2.00 / 1M tokens	$0.10 / 1M tokens
Cached input price	$0.20 / 1M tokens	$0.01 / 1M tokens
Standard output price	$10.00 / 1M tokens	$0.50 / 1M tokens
Coding suitability	Repository-level debugging, architectural refactoring	Unit test generation, boilerplate scripts, code review
Agentic workflow fit	Autonomous multi-step execution	Fast intermediate task execution in agent pipelines

New Features and Technical Advancements

Coding and Agentic Workflows

OpenAI describes GPT-6 Sol as a model tailored for demanding coding environments, capable of maintaining context across extensive repositories and handling iterative debugging without losing track of project requirements. For agentic workflows, Sol is tuned to operate in closed execution loops: generating code, testing for errors, reading console outputs, and self-correcting before returning a final deliverable.

Computer Use

Moving beyond static text generation, OpenAI reports improvements in computer-use tasks. These allow models to parse visual interfaces, interact with desktop operating systems, and manipulate desktop tools to perform practical workflows autonomously.

Factuality and Collaboration Style

OpenAI states that GPT-6 Sol and Luna feature an updated collaboration style intended to reduce conversational filler and work more effectively as direct technical partners. In terms of factual accuracy, OpenAI reports that GPT-6 Sol reduces mistakes by approximately half compared with its predecessor on its internal factuality evaluation.

Performance and Benchmark Results

To quantify progress, OpenAI released initial benchmark evaluations. Note: All figures below are reported directly by OpenAI and reflect specific testing environments.

• AutomationBench: OpenAI reports that GPT-6 Sol scored 33.2% at "xhigh effort," with an average cost of $0.27 per task. In the same cited comparison, OpenAI reports GPT-6 Astra scored 30.3% at "low effort." OpenAI also notes that Claude Opus 5 scored 26.9% at "max effort," though it explicitly highlights that evaluation conditions and comparator datasets differ between labs.
• Agents’ Last Exam: On this multi-discipline technical evaluation, OpenAI reports GPT-6 Sol achieved 56.4% at "max effort."
• DeepSWE v1.1: In software engineering benchmarks, OpenAI reports GPT-6 Sol scored 68.8% at "max effort." GPT-6 Luna also performed strongly for a compact model, scoring 66.6% under the same evaluation conditions.
• OSWorld 2.0 (Offline): For computer-use capabilities, OpenAI reports GPT-6 Sol scored 60.5% at "xhigh effort," reflecting an improved ability to complete multi-step operating system actions.

Pricing Breakdown: What the 50% Reduction Actually Means

OpenAI's standard API pricing for the new models is structured as follows:

Model	Input / 1M Tokens	Cached Input / 1M Tokens	Output / 1M Tokens
GPT-6 Sol	$2.00	$0.20	$10.00
GPT-6 Luna	$0.10	$0.01	$0.50

While OpenAI states that these rates represent a 50% reduction against GPT-5.6 promotional pricing, developers should remember that token rate does not equal total project cost. Real-world bills depend on:

• Reasoning effort settings: Running models at "xhigh effort" consumes substantially more internal tokens before producing a response.
• Output length: Output tokens remain significantly more expensive than input tokens ($10/1M on Sol).
• Prompt architecture: Applications that do not structure calls to reuse cached prefixes will pay standard input rates every cycle.

GPT-6 Prompt Caching Explained

Prompt caching allows the API to store pre-processed input tokens on OpenAI's servers so subsequent requests sharing the same prefix do not need to be recomputed from scratch.

OpenAI reports that GPT-6 prompt caching provides discounts of up to 90% on eligible cached input-token reads. Under this system, eligible shared prefixes can be reused within a 30-minute window.

This provides immediate savings for three common scenarios:

1. Coding assistants: Where the entire codebase schema or multiple project files are sent along with every follow-up question.
2. Long conversational agents: Where multi-turn histories grow progressively larger with every exchange.
3. Enterprise search and document analysis: Where thousands of user queries query the same static reference documentation.

The 90% discount applies strictly to cached input reads; standard rates still apply to novel input and all generated outputs.

Availability

Availability at launch depends on user tier and platform:

• ChatGPT Work and Codex: Available for Plus, Pro, Business, Enterprise, and Edu subscribers.
• Desktop Application: Free and Go plan users receive access to GPT-6 Luna inside the official desktop app.
• Developer API: Both models are accessible through the standard API under the model IDs gpt-6-sol and gpt-6-luna.
• Consumer Web Interface: OpenAI notes that deployment to the standard consumer web interface follows a gradual rollout schedule.

Limitations and Context

While these announcements demonstrate strong technological progress, several important limitations should be kept in mind:

• Internal benchmarks are not universal rankings: Benchmark results reported by model creators reflect specific test configurations and prompts. They do not guarantee identical performance on proprietary enterprise codebases.
• Factuality testing was conducted on edge cases: OpenAI explicitly clarifies that its internal factuality test was based on de-identified conversations in which users had previously reported factual errors. OpenAI notes that these error-inducing conversations are not representative of all typical usage, so the "50% error reduction" cannot be assumed across every general topic.
• No full autonomy: Despite improved performance on agentic benchmarks, these models remain predictive systems requiring structured constraints, human monitoring, and fallback protocols.

Practical Use Cases: Sol vs. Luna

When to Use GPT-6 Sol

• Architecting complex software systems and refactoring large codebases.
• Running persistent agents tasked with multi-step research and terminal execution.
• Analyzing lengthy legal, regulatory, or technical documentation requiring nuanced reasoning.

When to Use GPT-6 Luna

• High-throughput data extraction, classification, and metadata tagging.
• Real-time customer support routing and basic query answering.
• Acting as the fast orchestrator in a multi-model agent system, delegating difficult sub-tasks to Sol.

Frequently Asked Questions (FAQ)

What is GPT-6 Sol?
GPT-6 Sol is a high-capability foundational model within the GPT-6 family, designed by OpenAI for demanding coding, professional knowledge work, and complex agentic tasks.

What is GPT-6 Luna?
GPT-6 Luna is an efficient, low-cost model optimized for high-volume execution, fast response times, and scalable everyday tasks.

How much does GPT-6 Sol cost via API?
GPT-6 Sol costs $2.00 per 1 million input tokens ($0.20 for cached input reads) and $10.00 per 1 million output tokens.

How much does GPT-6 Luna cost via API?
GPT-6 Luna costs $0.10 per 1 million input tokens ($0.01 for cached input reads) and $0.50 per 1 million output tokens.

How does prompt caching save money on GPT-6?
Prompt caching provides a discount of up to 90% on eligible cached input tokens when an identical prefix is reused within a 30-minute window, drastically reducing costs for repeated context.

Are GPT-6 Sol and Luna available in the free version of ChatGPT?
Free and Go tier users can access GPT-6 Luna within the official desktop application. Standard consumer web rollout occurs gradually.

Can developers use GPT-6 Sol and Luna in production today?
Yes, both models are available via the OpenAI API under gpt-6-sol and gpt-6-luna.

Conclusion

With GPT-6 Sol and Luna, OpenAI is adjusting its strategy to match how production AI systems are actually built. By dividing frontier capabilities into a heavy-reasoning model (Sol) and a fast, cost-effective model (Luna), OpenAI gives developers more granular control over both intelligence and infrastructure budgets. Combined with lower base token pricing and aggressive prompt caching discounts, these models make building autonomous, persistent software agents far more economically viable than previous generations.

---

About the Author

Vishal Tiwari is a researcher, developer, and writer covering advancements in artificial intelligence, software tools, and digital infrastructure for VISRAONJ.

Last Updated: September 23, 2026