OpenAI's cheaper, faster GPT-6 models arrive with halved API rates and strong agent-benchmark scores, pitching everyday AI work against Anthropic's newest flagship in an intensifying pricing and performance race.
OpenAI unveiled two new artificial intelligence models, GPT-6 Sol and GPT-6 Luna, on Tuesday (September 22), minutes after rival Anthropic introduced Claude Opus 5.5, its latest flagship.
The releases slot beneath GPT-6 Astra, which OpenAI described as the world's most intelligent and aligned model when it debuted on September 3. Astra remains the company's choice for the most demanding tasks, while Sol and Luna target routine, high-volume work where speed and cost matter more.
The structure closely resembles Anthropic's approach. Astra corresponds to Anthropic's Fable, Sol to Opus, and Luna to Haiku once that model is upgraded. A mid-range counterpart to Sonnet is also expected to follow.
Prices cut in half
OpenAI also lowered prices to stay competitive. API rates for both models are 50 percent below GPT-5.6's promotional levels. AI companies bill by the token, a unit of text usually slightly shorter than a word, and customers typically pay per million of them.
Sol now costs $2 per million input tokens and $10 per million output tokens, down from $4 and $20. Luna falls to $0.10 for input and $0.50 for output, compared with $0.20 and $1.20 previously. By tier, OpenAI is now the cheaper option against Anthropic.
Benchmark claims
OpenAI's performance argument rests largely on its own reported figures. On AutomationBench, a Zapier-built test in which agents complete business workflows across 47 tools in sales, marketing, operations, support, finance and HR, Sol at its highest reasoning setting passed 33.2 percent of tasks at $0.27 each. Claude Opus 5 scored 26.9 percent while costing more than 11 times as much per task, OpenAI said. Sol also narrowly beat Anthropic's costlier Claude Fable 5.1.
On Agents' Last Exam, which measures long, economically valuable work across 55 sub-industries, Sol scored 56.4 percent. OpenAI says that surpasses Opus 5's best result at 60 percent lower cost per task.
Sol also nearly matched Opus 5 on OSWorld 2.0, a test of an AI operating software like a human user, scoring 60.5 percent against 60.3 percent, at roughly 80 percent lower cost. The reasoning-effort setting, which OpenAI added, makes a model spend more time and computing power checking its work, typically improving accuracy while raising cost.
Availability
Sol and Luna are live in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise and Edu subscribers. Luna also reaches free and Go users through the desktop app. Neither model is yet in the standard ChatGPT app, and OpenAI is staging the rollout through the day to keep the service stable.
