OpenAI has launched GPT-6 Sol and GPT-6 Luna, extending its GPT-6 range at lower API prices than their GPT-5.6 predecessors.
The models sit below GPT-6 Astra, OpenAI's top model for the most demanding work. Sol and Luna are aimed at faster, cheaper use across routine professional tasks.
OpenAI cut API prices for both models by 50% from the promotional pricing of GPT-5.6 Sol and GPT-5.6 Luna. GPT-6 Sol costs USD $2 per 1 million input tokens and USD $10 per 1 million output tokens, down from USD $4 and USD $20. GPT-6 Luna costs USD $0.10 per 1 million input tokens and USD $0.50 per 1 million output tokens, down from USD $0.20 and USD $1.20.
It attributed the lower prices to changes in caching and inference that reduce serving costs, with those savings passed on to users.
Benchmark results
OpenAI used internal and external evaluations to position the new models against earlier versions and rival systems. In business workflow testing on AutomationBench, GPT-6 Sol at xhigh effort scored 33.2% at a reported cost of USD $0.27 per task. That was above Claude Opus 5 at max effort on 26.9%, which OpenAI said cost 11.1 times more per task.
On the same benchmark, GPT-6 Sol also outperformed Claude Fable 5.1 at a lower reported cost and beat low-effort GPT-6 Astra. GPT-6 Luna at high effort improved on its predecessor by 5.4 percentage points at 58% lower cost per task, according to OpenAI.
In Agents' Last Exam, which measures long-horizon professional workflows across 55 sub-industries, GPT-6 Sol at max effort scored 56.4%. OpenAI said that was above Claude Opus 5's highest score in the evaluation while costing 60% less per task.
Factuality and coding
OpenAI highlighted factual reliability as a key area of improvement. In an internal evaluation built from de-identified real-world conversations in which users had flagged errors, GPT-6 Sol made about half as many mistakes as its predecessor and approached Astra-level reliability at lower cost, according to the company.
GPT-6 Luna also improved on that measure, OpenAI said. At higher effort levels, Luna matched GPT-5.6 Sol at about one-hundredth of the cost, it said.
Coding was another focus. OpenAI said internal use of coding agents had grown sharply, with daily token usage valued at API prices above USD $600 for the median researcher and USD $7,000 for researchers at the 90th percentile.
On FrontierCode, which assesses whether coding agents can produce changes ready to merge into real codebases, GPT-6 Sol improved substantially over GPT-5.6 Sol and matched Claude Fable 5.1 xhigh at lower cost, according to OpenAI. On DeepSWE v1.1, GPT-6 Sol at max effort scored 68.8%, within 1.1 percentage points of Claude Fable 5's highest score of 69.9% at xhigh effort, at what OpenAI said was about 80% lower cost per task.
GPT-6 Luna at max effort scored 66.6% on DeepSWE, which OpenAI said was comparable to Claude Opus 5 and Claude Fable 5 at medium effort. In those comparisons, Luna cost 93% less per task than Opus 5 and 96% less than Fable 5, according to the company.
Computer use
OpenAI said Astra remained its strongest model for computer use, but presented Sol and Luna as cheaper options with improved performance over earlier versions. On OSWorld 2.0 offline, GPT-6 Sol at xhigh effort achieved 60.5%, compared with Claude Opus 5 at medium effort on 60.3%, at what OpenAI said was about 80% lower cost per task.
GPT-6 Luna at max effort outperformed GPT-5.6 Sol at medium effort at one-tenth of the cost, according to OpenAI.
OpenAI also said it had adjusted the communication style of Sol and Luna to reduce jargon, cut low-value detail and produce shorter answers without losing substance, particularly in technical and coding exchanges.
Caching changes
Alongside the new models, OpenAI said it had improved prompt caching to increase cache hit rates by default. That should help developers building agents and long conversations reuse more context and qualify for 90% discounts on cached input-token reads, it said.
Developers can now monitor caching performance through a Prompt Caching Dashboard and a diagnostics tool, OpenAI said. It also introduced controls that let developers adjust reasoning effort and tool availability without breaking cached context, along with explicit breakpoints to choose where cached prompt prefixes end.
GitHub has seen material effects from those changes, according to OpenAI. "Over the past several months, these improvements have reduced the share of prompt tokens requiring fresh processing by more than 50% across billions of requests to OpenAI models, helping Copilot respond faster," OpenAI said.
Access tiers
GPT-6 Sol and GPT-6 Luna are available in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise and Edu users. Free and Go users can access GPT-6 Luna in the desktop app. Both models are also available through the OpenAI API as gpt-6-sol and gpt-6-luna.
OpenAI said Sol and Luna build on alignment work introduced with Astra and show lower rates of misleading claims about coding work than their GPT-5.6 counterparts in internal testing.