OpenAI has introduced GPT-6 Sol and Luna as lower-cost entrants in the GPT-6 family, extending many of the Astra-era improvements in coding, factuality and agent performance to a broader set of use cases. GPT-6 Sol and Luna are presented as cost-efficient alternatives to GPT-6 Astra, trained with methods similar to Astra and supported by infrastructure changes that reduce serving costs.
The most immediate change for developers is pricing: OpenAI has cut API prices for these models by 50% relative to the earlier GPT-5.6 promotional rates. Pricing is quoted per 1 million tokens. GPT-6 Sol’s input and output prices move from $4/$20 to $2/$10 per million tokens, while GPT-6 Luna’s pricing shifts from $0.20/$1.20 to $0.10/$0.50 per million tokens. OpenAI continues to recommend GPT-6 Astra for projects that require the highest capability and alignment, positioning Sol and Luna as the practical choice where cost matters.
OpenAI published benchmark results showing that the new models deliver notable capability gains while lowering cost per task across a range of professional evaluations. On AutomationBench, which measures end-to-end business workflows across 47 tools, GPT-6 Sol at xhigh effort recorded a 33.2% score and an estimated cost per task of $0.27. OpenAI says this both outperforms Claude Opus 5 at its max effort and does so at roughly one-tenth of Opus 5’s per-task cost, while also surpassing Claude Fable 5.1 at far lower cost and outperforming a low-effort GPT-6 Astra setting.
For agent-style professional workflows, GPT-6 Sol at max effort scored 56.4% on Agents’ Last Exam, a result OpenAI reports as higher than the best score achieved by Claude Opus 5 in that evaluation and accomplished at about a 60% lower cost per task. These comparisons aim to show cost-efficiency as well as capability improvements for common multi-step workflows.
Factual reliability and alignment are highlighted as important gains. In internal factuality evaluations based on de-identified real-world conversations flagged for mistakes, GPT-6 Sol made about half as many mistakes as its predecessor, approaching Astra-level reliability at lower running cost. GPT-6 Luna also showed substantial factuality gains and, at higher effort settings, matched GPT-5.6 Sol’s factuality at roughly a hundredth of that previous cost. OpenAI adds that both models build on Astra’s alignment work and demonstrate lower rates of misleading claims about coding tasks compared with GPT-5.6 models.
Developer-focused benchmarks show meaningful improvements in coding and long-horizon computer use. On FrontierCode, GPT-6 Sol improves over GPT-5.6 Sol and can match Claude Fable 5.1 at lower cost. In DeepSWE v1.1, which tests complex software-engineering tasks in real codebases, GPT-6 Sol at max effort scored 68.8%, within 1.1 percentage points of Claude Fable 5’s best result (69.9%) but at roughly 80% lower cost per task. GPT-6 Luna at max effort scored 66.6% on DeepSWE v1.1, a level OpenAI says is comparable to Claude Opus 5 and Fable 5 at medium effort, while costing 93% less than Opus 5 and 96% less than Fable 5 per task.
On OSWorld 2.0 offline, which evaluates long-horizon computer-use workflows, GPT-6 Sol at xhigh effort achieved 60.5%, roughly similar to Claude Opus 5 at medium effort but at about an 80% lower cost per task. GPT-6 Luna (max) exceeded GPT-5.6 Sol (medium) at approximately one-tenth of the cost. OpenAI also notes that Sol and Luna inherit Astra’s improved collaboration and communication style, producing clearer, more concise responses with fewer low-value details in technical and coding conversations.
To further cut costs for agents and long conversations, OpenAI has improved prompt caching for GPT-6, raising cache hit rates by default and offering discounts of 90% on cached input-token reads. The company offers a Prompt Caching Dashboard and a diagnostics tool that explains missed caching opportunities and recommends fixes. GitHub reported that prompt-caching improvements reduced the share of prompt tokens requiring fresh processing by more than 50% across billions of requests to OpenAI models.
Availability for developers and organizations is immediate: GPT-6 Sol and GPT-6 Luna are available in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise and Edu customers. Free and Go users can access GPT-6 Luna in the desktop app. OpenAI says the models are not yet available in Chat and will be rolled out gradually to maintain stability. In the OpenAI API the models are exposed as gpt-6-sol and gpt-6-luna.
By making Astra-era advances more affordable, OpenAI aims to make advanced AI capabilities practical for everyday tasks and large-scale deployments. GPT-6 Sol and Luna lower the cost barrier for professional coding, agent workflows and factuality improvements, combining model refinements with infrastructure and caching optimizations to reduce per-task costs while preserving many of the performance and alignment gains introduced with GPT-6 Astra.
In short, GPT-6 Sol and Luna offer a marked shift toward cost-effective access to higher-performing models, giving developers more options to balance capability and price for production-scale AI applications.
Source: Read the original source
Leave a Reply