Following the recent introduction of GPT‑6 Astra, OpenAI has released two additional GPT‑6 models, GPT‑6 Sol and GPT‑6 Luna, intended to deliver many of Astra’s improvements at lower cost and higher throughput. The company says these models were trained with methods similar to Astra’s to bring advances in professional tasks, factuality, coding, computer use, and alignment to faster, more affordable options.
Cost, pricing, and caching improvements
OpenAI positions the GPT‑6 models across the cost–intelligence curve: infrastructure and caching improvements allow the company to serve these models more cheaply, and OpenAI is passing savings to customers by reducing API prices for Sol and Luna by 50% relative to their GPT‑5.6 promotional pricing. Improved prompt caching increases default cache hit rates and can yield discounts of 90% on cached input‑token reads.
Developers get new tools to monitor and optimize caching: a Prompt Caching Dashboard, diagnostics that explain missed caching opportunities, controls to adjust reasoning effort and tool availability without invalidating cached context, and explicit cache breakpoints for cached prefixes. GitHub reported that these changes reduced the share of prompt tokens requiring fresh processing by more than 50% across billions of OpenAI requests.
Business workflow benchmarks (AutomationBench)
OpenAI published benchmark comparisons showing cost and performance gains:
- On AutomationBench, which tests business workflows across 47 tools and functions, GPT‑6 Sol at xhigh effort outperforms Claude Opus 5 at max effort while costing only 9% of Opus 5’s cost per task. GPT‑6 Luna at higher effort improves over its predecessor by 5.4 percentage points while costing 58% less per task.
- OpenAI notes that the reported cost for Claude Fable 5.1 understates its true cost because it omits Opus 5 fallbacks, which occurred on about 40% of tasks.
Long‑horizon professional work (Agents’ Last Exam)
On Agents’ Last Exam — an evaluation of long‑horizon, economically valuable tasks across 55 sub‑industries — GPT‑6 Sol at max effort scored 56.4%. OpenAI reports this as higher than Claude Opus 5’s top score in the same evaluation, with Sol’s cost per task about 60% lower.
Factuality
In internal factuality evaluations based on de‑identified real‑world conversations where users had flagged prior model mistakes, GPT‑6 Sol makes roughly half as many factual errors as its predecessor, approaching Astra‑level reliability at much lower cost. GPT‑6 Luna also improves substantially; at higher effort levels it matches GPT‑5.6 Sol’s factuality while costing about one‑hundredth as much per task.
OpenAI stresses these tests focus on error‑inducing conversations and therefore do not represent typical usage, where factual errors are less common.
Coding and software engineering benchmarks
OpenAI says internal use of coding agents has risen sharply; valued at API prices, daily token usage exceeded $600 for the median researcher and $7,000 for researchers at the 90th percentile. As agents take on longer, more demanding tasks, sustained cost matters more—OpenAI positions Sol and Luna as options that combine strong coding performance with lower API prices.
Key results cited:
- FrontierCode: GPT‑6 Sol shows substantial improvement over GPT‑5.6 Sol and can match Claude Fable 5.1 xhigh at much lower cost. FrontierCode grades code not just for correctness but for "mergeability," including test quality, scope discipline, style, and codebase standards.
- DeepSWE v1.1: GPT‑6 Sol at max effort scored 68.8%, within 1.1 percentage points of Claude Fable 5’s top score of 69.9% at xhigh effort, while costing roughly 80% less per task. GPT‑6 Luna at max effort scored 66.6%, comparable to Claude Opus 5 and Fable 5 at medium effort, and costs 93% less per task than Opus 5 and 96% less than Fable 5.
Computer‑use tasks and communication style
On OSWorld 2.0 offline, GPT‑6 Sol at xhigh effort scored 60.5%, similar to Claude Opus 5 at medium effort (60.3%), at approximately 80% lower cost per task. GPT‑6 Luna (max) outperformed GPT‑5.6 Sol (medium) while costing about one‑tenth as much.
OpenAI also reports that Sol and Luna inherit Astra’s improved communication style: clearer, less jargon‑heavy answers with fewer low‑value details and slightly shorter responses in technical and coding conversations.
Alignment and safety evaluations
According to OpenAI’s internal alignment evaluations, Sol and Luna show improvements over GPT‑5.6 models, including lower rates of misleading claims about coding work. The company notes these tests are deliberately challenging and do not measure typical failure rates; full results are documented in the model system card.
Availability
GPT‑6 Sol and GPT‑6 Luna are available starting today in ChatGPT Work and Codex for all Plus, Pro, Business, Enterprise, and Edu users. Free and Go users can access GPT‑6 Luna in the desktop app. The models are not yet available in Chat. In the OpenAI API they are accessible as gpt-6-sol and gpt-6-luna.
OpenAI plans a gradual roll‑out within ChatGPT to preserve service stability, so users who do not immediately see the new models are advised to try again later. The company also notes that benchmark evaluations were run in research environments or via the API and may differ slightly from production ChatGPT outputs due to different system prompts and available tools.
