OpenAI has announced the release of GPT-6 Sol and GPT-6 Luna, two significant additions to its formidable GPT-6 model series. Positioned as more economical and swifter alternatives to the flagship GPT-6 Astra, these models aim to democratize advanced AI capabilities for a broader range of applications.
Both Sol and Luna are immediately accessible via the OpenAI API, designated as gpt-6-sol and gpt-6-luna, respectively. It is important to note that these models are exclusively API-based, with no provisions for self-hosting.
A Three-Tiered AI Strategy
The GPT-6 ecosystem now comprises three distinct tiers, each optimized for specific computational demands. GPT-6 Astra remains the premier choice for highly complex and demanding tasks. Following this, GPT-6 Sol is engineered for intricate coding projects and professional-grade applications, offering a balance of power and cost-efficiency. Meanwhile, GPT-6 Luna is tailored for high-volume, rapid processing of everyday tasks, emphasizing speed and affordability.
Revolutionary Pricing Structure Announced
A core highlight of this release is a dramatic reduction in API pricing for the new models. OpenAI attributes these cost savings to enhanced caching and more efficient inference mechanisms. API access for Sol and Luna is now approximately 50% cheaper compared to their previous GPT-5.6 promotional rates. For instance, the input cost for GPT-6 Sol drops from $4.00 to $2.00 per million tokens, while its output cost decreases from $20.00 to $10.00. GPT-6 Luna sees its input price fall from $0.20 to $0.10 and output price from $1.20 to $0.50 per million tokens, representing a remarkable 58% reduction for output.
Unpacking Performance Benchmarks
Independent evaluations and internal metrics released by OpenAI showcase the impressive capabilities and cost-effectiveness of Sol and Luna across various domains. On the demanding AutomationBench 1.0.6, GPT-6 Sol achieved a 33.2% score at a mere $0.27 per task, significantly outperforming a top competitor at over eleven times the cost. Similarly, Sol's performance on Agents' Last Exam surpassed rival models while maintaining a 60% lower cost per task.
For coding tasks, DeepSWE v1.1 results indicate GPT-6 Sol scored 68.8%, nearing a competitor's performance but at roughly 80% less expense. GPT-6 Luna, scoring 66.6% on the same benchmark, provides comparable results to more expensive alternatives for a fraction of the price, sometimes costing 93% to 96% less per task. On FrontierCode 1.1 Main, Sol demonstrates parity with leading models for code readiness, again at a substantially lower operational expense.
In computer usage scenarios, OSWorld 2.0 offline tests showed Sol achieving 60.5%, matching or slightly exceeding a competitor's score at approximately 80% lower cost. Luna also demonstrated superior performance over its predecessor with a tenfold cost reduction. Furthermore, internal factuality tests reveal Sol makes half as many errors as its predecessor, and Luna, with greater effort, matches prior Sol models at about one-hundredth of the cost.
These new models also inherit the enhanced communication style first seen in Astra, characterized by clearer, more concise responses and reduced technical jargon, particularly beneficial in coding dialogues.
Advanced Prompt Caching for Developers
A significant architectural enhancement accompanying this release is an upgraded prompt caching system, critical for optimizing long-running agent workflows. Agents frequently resubmit identical instructions, tool definitions, and historical context with each interaction. The improved system boasts higher cache hit rates by default, offering discounts of up to 90% on cached input reads. This applies to shared prefixes reused within a thirty-minute window.
Developers now have granular control over caching behavior through new features:
- A Prompt Caching Dashboard for monitoring hit rates over time.
- A diagnostics tool that explains cache misses, providing specific reasons.
- Explicit breakpoints allowing precise definition of where a cached prefix ends.
- The ability to modify reasoning effort mid-conversation via
configuration_updatewithout invalidating the cache. allowed_toolsfor restricting callable tools while keeping definitions stable.- Prewarming capabilities for preparing known context before the first user request.
Reports from GitHub suggest these caching innovations have already reduced the proportion of prompt tokens requiring fresh processing by over 50%, contributing to faster response times for tools like Copilot.
Availability Details
The GPT-6 Sol and Luna models are now available through the OpenAI API. They are also integrated into ChatGPT Work and Codex for Plus, Pro, Business, Enterprise, and Education users. Additionally, GPT-6 Luna is accessible within the ChatGPT desktop application for Free and Go tier users. The full rollout across various ChatGPT platforms will proceed gradually throughout the launch period.
This article is a rewritten summary based on publicly available reporting. For the original story, visit the source.
Source: MarkTechPost
