OpenAI Releases GPT 6 Sol And GPT 6 Luna With 50% Lower Prices Than GPT 5.6

OpenAI released GPT-6 Sol and GPT-6 Luna, two models positioned below its flagship GPT-6 Astra, with API pricing 50% lower than GPT-5.6 promotional rates — Sol at $2 per 1M input tokens and $10 per 1M output tokens, and Luna at $0.10 per 1M input and $0.50 per 1M output. OpenAI attributes the cuts to improvements in caching and inference efficiency, and says Sol outperforms Anthropic's Claude Opus 5 on AutomationBench at its xhigh effort setting at roughly a tenth the cost per task. GPT-6 Astra remains OpenAI's top-performing model overall.

OpenAI has widened its GPT-6 lineup with two new additions, GPT-6 Sol and GPT-6 Luna, slotting in below the flagship GPT-6 Astra https://officechai.com/ai/openai-says-its-models-have-resolved-more-than-100-long-standing-mathematical-problems-in-addition-to-navier-stokes/ model the company introduced earlier this month. While Astra remains OpenAI’s top pick for the toughest and most demanding projects, the company says everyday work happens at different scales and budgets, and Sol and Luna are meant to bring much of Astra’s underlying capability to faster, cheaper models built for that kind of everyday use. Both new models were trained using methods similar to Astra, and OpenAI says they carry forward its gains in professional work, factual accuracy, coding, computer use, and alignment. The headline change, though, is price: API costs for Sol and Luna are down 50% compared to their GPT-5.6 promotional pricing, a cut OpenAI attributes to improvements in caching and inference efficiency that it says it is passing straight on to developers and customers. GPT-6 Pricing | Model | Input per 1M tokens | Output per 1M tokens | Price Cut | |---|---|---|---| | GPT-6 Sol vs GPT-5.6 Sol | $4 → $2 | $20 → $10 | 50% cheaper | | GPT-6 Luna vs GPT-5.6 Luna | $0.20 → $0.10 | $1.20 → $0.50 | 50% cheaper | GPT-6 Astra continues to sit at the top of the lineup as OpenAI’s best-performing model overall, for users who want the strongest possible results regardless of cost. A Step Up Across The Model Family OpenAI is positioning Sol and Luna as meaningful intelligence upgrades over their predecessors, not just cheaper versions of the same models, and has shared benchmark comparisons across several categories to back that up. Professional Work On AutomationBench, a test that measures how well AI agents handle real business workflows across dozens of apps, OpenAI says GPT-6 Sol running at its highest “xhigh” effort setting outperforms Anthropic’s Claude Opus 5 at its own top effort setting, while costing roughly a tenth as much per task. GPT-6 Luna, meanwhile, is said to have improved on GPT-5.6 Luna by over 5 percentage points while costing more than half as much per task. The comparison extends to Agents’ Last Exam, a benchmark covering long, complex professional workflows across dozens of industries. Here, OpenAI reports that GPT-6 Sol at maximum effort scores above Claude Opus 5’s best result in the test, at roughly 60% lower cost per task. Factuality OpenAI also highlights factuality gains, based on an internal evaluation built from real conversations where users had previously flagged mistakes. By this measure, GPT-6 Sol is said to make about half as many factual errors as GPT-5.6 Sol, bringing it closer to Astra-level reliability at a fraction of the cost. GPT-6 Luna also shows a sizable improvement, reportedly matching GPT-5.6 Sol’s factuality at a small fraction of what that model costs to run. Coding Coding continues to be one of the most closely watched categories in AI releases, especially as agents take on longer and more complex tasks. On FrontierCode, a benchmark that evaluates whether an AI agent’s code changes are actually ready to merge into a real codebase, OpenAI says GPT-6 Sol improves substantially on its predecessor and is able to match rival Claude Fable 5.1 at its highest effort setting, at considerably lower cost. On DeepSWE, a benchmark for complex, long-horizon software engineering tasks, GPT-6 Sol at maximum effort is reported to score within roughly a percentage point of Fable 5’s best result, at around 80% lower cost per task, while GPT-6 Luna is positioned as comparable to Claude Opus 5 and Fable 5 at medium effort settings, at a fraction of their per-task cost. Computer Use GPT-6 Astra remains OpenAI’s strongest model for computer-use tasks, but the company says Sol and Luna are more cost-efficient than their predecessors here too. On OSWorld 2.0, an offline benchmark for long-horizon computer-use workflows, GPT-6 Sol at its highest effort setting is reported to score close to Claude Opus 5 at medium effort, at around 80% lower cost per task. A Different Communication Style OpenAI says it has also carried over Astra’s revised communication style to Sol and Luna, something it expects users to notice most in technical and coding conversations — less jargon, fewer unnecessary details, and answers that are slightly shorter without losing substance. The company shared a side-by-side example of GPT-5.6 Sol and GPT-6 Sol responding to the same website design request, arguing that the newer model’s reply was more direct and less prone to restating things the user already knew, while also being more transparent about what it had and hadn’t actually checked. Cheaper Caching For Agents Alongside the token price cuts, OpenAI says it has improved prompt caching for GPT-6, aiming for higher cache hit rates by default so that agents and long conversations can reuse more context. Cached input-token reads now come with a 90% discount, and OpenAI has added tools including a Prompt Caching Dashboard and a diagnostics tool to help developers see how much of their input is being cached and where they’re leaving savings on the table. Developers can now also adjust reasoning effort or toggle tools on and off without breaking their cache, and can set explicit breakpoints to control which parts of a prompt get cached. Alignment Improvements OpenAI says both Sol and Luna build on the alignment work introduced with Astra, showing improvements over their GPT-5.6 counterparts on internal evaluations, including lower rates of misleading claims about their own coding work. The company notes that these evaluations are deliberately built around challenging, edge-case scenarios designed to surface dishonest behaviour, and don’t reflect failure rates in typical day-to-day use — a distinction that matters given how closely rivals’ safety and reliability numbers are being scrutinised as the release cadence across the industry keeps accelerating https://officechai.com/ai/xiaomi-mimo-v-2-6-pro-benchmarks/ . Availability GPT-6 Sol and GPT-6 Luna are rolling out today in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise, and Edu users, while Free and Go users can access GPT-6 Luna through the desktop app. Neither model is yet available in standard ChatGPT Chat. On the API side, developers can access them as gpt-6-sol and gpt-6-luna . OpenAI says the ChatGPT rollout will happen gradually over the course of the day to keep things stable, so users who don’t immediately see the new models in ChatGPT Work or Codex are advised to check back later. The release comes at a moment when the pace of frontier model launches shows no signs of slowing, with rivals such as SpaceXAI’s Grok 4.7 https://officechai.com/ai/grok-4-7-benchmarks/ also pushing out new benchmark claims in the same window, keeping the pressure on every major lab to ship faster and cheaper models without giving up ground on capability.